Search

Purpose: Medical images acquired using different scanners and protocols can differ substantially in their appearance. This phenomenon, scanner domain shift, can result in a drop in the performance of deep neural networks which are trained on data acquired by one scanner and tested on another. This significant practical issue is well-acknowledged, however, no systematic study of the issue is available across different modalities and diagnostic tasks. Materials and Methods: In this paper, we present a broad experimental study evaluating the impact of scanner domain shift on convolutional neural network performance for different automated diagnostic tasks. We evaluate this phenomenon in common radiological modalities, including X-ray, CT, and MRI. Results: We find that network performance on data from a different scanner is almost always worse than on same-scanner data, and we quantify the degree of performance drop across different datasets. Notably, we find that this drop is most severe for MRI, moderate for X-ray, and quite small for CT, on average, which we attribute to the standardized nature of CT acquisition systems which is not present in MRI or X-ray. We also study how injecting varying amounts of target domain data into the training set, as well as adding noise to the training data, helps with generalization. Conclusion: Our results provide extensive experimental evidence and quantification of the extent of performance drop caused by scanner domain shift in deep learning across different modalities, with the goal of guiding the future development of robust deep learning models for medical image analysis.

Creating annotations for 3D medical data is time-consuming and often requires highly specialized expertise. Various tools have been implemented to aid this process. Segment Anything Model 2 (SAM 2) offers a general-purpose prompt-based segmentation algorithm designed to annotate videos. In this paper, we adapt this model to the annotation of 3D medical images and offer our implementation in the form of an extension to the popular annotation software: 3D Slicer. Our extension allows users to place point prompts on 2D slices to generate annotation masks and propagate these annotations across entire volumes in either single-directional or bi-directional manners. Our code is publicly available on https://github.com/mazurowski-lab/SlicerSegmentWithSAM and can be easily installed directly from the Extension Manager of 3D Slicer as well., Comment: Future work: support for box and mask inputs for the video predictor of SAM 2

In recent years, there has been interest in how geometric properties such as intrinsic dimension (ID) of a neural network's hidden representations change through its layers, and how such properties are predictive of important model behavior such as generalization ability. However, evidence has begun to emerge that such behavior can change significantly depending on the domain of the network's training data, such as natural versus medical images. Here, we further this inquiry by exploring how the ID of a network's learned representations changes through its layers, in essence, characterizing how the network successively refines the information content of input data to be used for predictions. Analyzing eleven natural and medical image datasets across six network architectures, we find that how ID changes through the network differs noticeably between natural and medical image models. Specifically, medical image models peak in representation ID earlier in the network, implying a difference in the image features and their abstractness that are typically used for downstream tasks in these domains. Additionally, we discover a strong correlation of this peak representation ID with the ID of the data in its input space, implying that the intrinsic information content of a model's learned representations is guided by that of the data it was trained on. Overall, our findings emphasize notable discrepancies in network behavior between natural and non-natural imaging domains regarding hidden representation information content, and provide further insights into how a network's learned features are shaped by its training data.

Segment Anything Model (SAM) has gained significant attention because of its ability to segment various objects in images given a prompt. The recently developed SAM 2 has extended this ability to video inputs. This opens an opportunity to apply SAM to 3D images, one of the fundamental tasks in the medical imaging field. In this paper, we extensively evaluate SAM 2's ability to segment both 2D and 3D medical images by first collecting 21 medical imaging datasets, including surgical videos, common 3D modalities such as computed tomography (CT), magnetic resonance imaging (MRI), and positron emission tomography (PET) as well as 2D modalities such as X-ray and ultrasound. Two evaluation settings of SAM 2 are considered: (1) multi-frame 3D segmentation, where prompts are provided to one or multiple slice(s) selected from the volume, and (2) single-frame 2D segmentation, where prompts are provided to each slice. The former only applies to videos and 3D modalities, while the latter applies to all datasets. Our results show that SAM 2 exhibits similar performance as SAM under single-frame 2D segmentation, and has variable performance under multi-frame 3D segmentation depending on the choices of slices to annotate, the direction of the propagation, the predictions utilized during the propagation, etc. We believe our work enhances the understanding of SAM 2's behavior in the medical field and provides directions for future work in adapting SAM 2 to this domain. Our code is available at: https://github.com/mazurowski-lab/segment-anything2-medical-evaluation., Comment: 20 pages, 13 figures. Codes are available at https://github.com/mazurowski-lab/segment-anything2-medical-evaluation

Automated segmentation is a fundamental medical image analysis task, which enjoys significant advances due to the advent of deep learning. While foundation models have been useful in natural language processing and some vision tasks for some time, the foundation model developed with image segmentation in mind - Segment Anything Model (SAM) - has been developed only recently and has shown similar promise. However, there are still no systematic analyses or "best-practice" guidelines for optimal fine-tuning of SAM for medical image segmentation. This work summarizes existing fine-tuning strategies with various backbone architectures, model components, and fine-tuning algorithms across 18 combinations, and evaluates them on 17 datasets covering all common radiology modalities. Our study reveals that (1) fine-tuning SAM leads to slightly better performance than previous segmentation methods, (2) fine-tuning strategies that use parameter-efficient learning in both the encoder and decoder are superior to other strategies, (3) network architecture has a small impact on final performance, (4) further training SAM with self-supervised learning can improve final model performance. We also demonstrate the ineffectiveness of some methods popular in the literature and further expand our experiments into few-shot and prompt-based settings. Lastly, we released our code and MRI-specific fine-tuned weights, which consistently obtained superior performance over the original SAM, at https://github.com/mazurowski-lab/finetune-SAM., Comment: Code available at https://github.com/mazurowski-lab/finetune-SAM

Modern medical image translation methods use generative models for tasks such as the conversion of CT images to MRI. Evaluating these methods typically relies on some chosen downstream task in the target domain, such as segmentation. On the other hand, task-agnostic metrics are attractive, such as the network feature-based perceptual metrics (e.g., FID) that are common to image translation in general computer vision. In this paper, we investigate evaluation metrics for medical image translation on two medical image translation tasks (GE breast MRI to Siemens breast MRI and lumbar spine MRI to CT), tested on various state-of-the-art translation methods. We show that perceptual metrics do not generally correlate with segmentation metrics due to them extending poorly to the anatomical constraints of this sub-field, with FID being especially inconsistent. However, we find that the lesser-used pixel-level SWD metric may be useful for subtle intra-modality translation. Our results demonstrate the need for further research into helpful metrics for medical image translation.

Scoliosis, a prevalent condition characterized by abnormal spinal curvature leading to deformity, requires precise assessment methods for effective diagnosis and management. The Cobb angle is a widely used scoliosis quantification method that measures the degree of curvature between the tilted vertebrae. Yet, manual measuring of Cobb angles is time-consuming and labor-intensive, fraught with significant interobserver and intraobserver variability. To address these challenges and the lack of interpretability found in certain existing automated methods, we have created fully automated software that not only precisely measures the Cobb angle but also provides clear visualizations of these measurements. This software integrates deep neural network-based spine region detection and segmentation, spine centerline identification, pinpointing the most significantly tilted vertebrae, and direct visualization of Cobb angles on the original images. Upon comparison with the assessments of 7 expert readers, our algorithm exhibited a mean deviation in Cobb angle measurements of 4.17 degrees, notably surpassing the manual approach's average intra-reader discrepancy of 5.16 degrees. The algorithm also achieved intra-class correlation coefficients (ICC) exceeding 0.96 and Pearson correlation coefficients above 0.944, reflecting robust agreement with expert assessments and superior measurement reliability. Through the comprehensive reader study and statistical analysis, we believe this algorithm not only ensures a higher consensus with expert readers but also enhances interpretability and reproducibility during assessments. It holds significant promise for clinical application, potentially aiding physicians in more accurate scoliosis assessment and diagnosis, thereby improving patient care., Comment: 17 pages, 5 figures

Accurately translating medical images across different modalities (e.g., CT to MRI) has numerous downstream clinical and machine learning applications. While several methods have been proposed to achieve this, they often prioritize perceptual quality with respect to output domain features over preserving anatomical fidelity. However, maintaining anatomy during translation is essential for many tasks, e.g., when leveraging masks from the input domain to develop a segmentation model with images translated to the output domain. To address these challenges, we propose ContourDiff, a novel framework that leverages domain-invariant anatomical contour representations of images. These representations are simple to extract from images, yet form precise spatial constraints on their anatomical content. We introduce a diffusion model that converts contour representations of images from arbitrary input domains into images in the output domain of interest. By applying the contour as a constraint at every diffusion sampling step, we ensure the preservation of anatomical content. We evaluate our method by training a segmentation model on images translated from CT to MRI with their original CT masks and testing its performance on real MRIs. Our method outperforms other unpaired image translation methods by a significant margin, furthermore without the need to access any input domain information during training., Comment: Code will be released on GitHub

Test-time adaptation (TTA) refers to adapting a trained model to a new domain during testing. Existing TTA techniques rely on having multiple test images from the same domain, yet this may be impractical in real-world applications such as medical imaging, where data acquisition is expensive and imaging conditions vary frequently. Here, we approach such a task, of adapting a medical image segmentation model with only a single unlabeled test image. Most TTA approaches, which directly minimize the entropy of predictions, fail to improve performance significantly in this setting, in which we also observe the choice of batch normalization (BN) layer statistics to be a highly important yet unstable factor due to only having a single test domain example. To overcome this, we propose to instead integrate over predictions made with various estimates of target domain statistics between the training and test statistics, weighted based on their entropy statistics. Our method, validated on 24 source/target domain splits across 3 medical image datasets surpasses the leading method by 2.9% Dice coefficient on average., Comment: Code and pre-trained weights: https://github.com/mazurowski-lab/single-image-test-time-adaptation

Diffusion models have enabled remarkably high-quality medical image generation, yet it is challenging to enforce anatomical constraints in generated images. To this end, we propose a diffusion model-based method that supports anatomically-controllable medical image generation, by following a multi-class anatomical segmentation mask at each sampling step. We additionally introduce a random mask ablation training algorithm to enable conditioning on a selected combination of anatomical constraints while allowing flexibility in other anatomical areas. We compare our method ("SegGuidedDiff") to existing methods on breast MRI and abdominal/neck-to-pelvis CT datasets with a wide range of anatomical objects. Results show that our method reaches a new state-of-the-art in the faithfulness of generated images to input anatomical masks on both datasets, and is on par for general anatomical realism. Finally, our model also enjoys the extra benefit of being able to adjust the anatomical similarity of generated images to real images of choice through interpolation in its latent space. SegGuidedDiff has many applications, including cross-modality translation, and the generation of paired or counterfactual data. Our code is available at https://github.com/mazurowski-lab/segmentation-guided-diffusion., Comment: Accepted at MICCAI 2024. Code and synthetic dataset: https://github.com/mazurowski-lab/segmentation-guided-diffusion

Magnetic Resonance Imaging (MRI) is pivotal in radiology, offering non-invasive and high-quality insights into the human body. Precise segmentation of MRIs into different organs and tissues would be highly beneficial since it would allow for a higher level of understanding of the image content and enable important measurements, which are essential for accurate diagnosis and effective treatment planning. Specifically, segmenting bones in MRI would allow for more quantitative assessments of musculoskeletal conditions, while such assessments are largely absent in current radiological practice. The difficulty of bone MRI segmentation is illustrated by the fact that limited algorithms are publicly available for use, and those contained in the literature typically address a specific anatomic area. In our study, we propose a versatile, publicly available deep-learning model for bone segmentation in MRI across multiple standard MRI locations. The proposed model can operate in two modes: fully automated segmentation and prompt-based segmentation. Our contributions include (1) collecting and annotating a new MRI dataset across various MRI protocols, encompassing over 300 annotated volumes and 8485 annotated slices across diverse anatomic regions; (2) investigating several standard network architectures and strategies for automated segmentation; (3) introducing SegmentAnyBone, an innovative foundational model-based approach that extends Segment Anything Model (SAM); (4) comparative analysis of our algorithm and previous approaches; and (5) generalization analysis of our algorithm across different anatomical locations and MRI sequences, as well as an external dataset. We publicly release our model at https://github.com/mazurowski-lab/SegmentAnyBone., Comment: 15 pages, 15 figures

This paper investigates discrepancies in how neural networks learn from different imaging domains, which are commonly overlooked when adopting computer vision techniques from the domain of natural images to other specialized domains such as medical images. Recent works have found that the generalization error of a trained network typically increases with the intrinsic dimension ($d_{data}$) of its training set. Yet, the steepness of this relationship varies significantly between medical (radiological) and natural imaging domains, with no existing theoretical explanation. We address this gap in knowledge by establishing and empirically validating a generalization scaling law with respect to $d_{data}$, and propose that the substantial scaling discrepancy between the two considered domains may be at least partially attributed to the higher intrinsic ``label sharpness'' ($K_\mathcal{F}$) of medical imaging datasets, a metric which we propose. Next, we demonstrate an additional benefit of measuring the label sharpness of a training set: it is negatively correlated with the trained model's adversarial robustness, which notably leads to models for medical images having a substantially higher vulnerability to adversarial attack. Finally, we extend our $d_{data}$ formalism to the related metric of learned representation intrinsic dimension ($d_{repr}$), derive a generalization scaling law with respect to $d_{repr}$, and show that $d_{data}$ serves as an upper bound for $d_{repr}$. Our theoretical results are supported by thorough experiments with six models and eleven natural and medical imaging datasets over a range of training set sizes. Our findings offer insights into the influence of intrinsic dataset properties on generalization, representation learning, and robustness in deep neural networks. Code link: https://github.com/mazurowski-lab/intrinsic-properties, Comment: ICLR 2024. Code: https://github.com/mazurowski-lab/intrinsic-properties

Deep neural networks (DNNs) have been deployed for many image segmentation tasks and achieved outstanding performance. However, preparing a dataset for training segmentation DNNs is laborious and costly since typically pixel-level annotations are provided for each object of interest. To alleviate this issue, one can provide only weak labels such as bounding boxes or scribbles, or less accurate (noisy) annotations of the objects. These are significantly faster to generate and thus result in more annotated images given the same time budget. However, the reduction in quality might negatively affect the segmentation performance of the resulting model. In this study, we perform a thorough cost-effectiveness evaluation of several weak and noisy labels. We considered 11 variants of annotation strategies and 4 datasets. We conclude that the common practice of accurately outlining the objects of interest is virtually never the optimal approach when the annotation time is limited, even if notable annotation time is available (10s of hours). Annotation approaches that stood out in such scenarios were (1) contour-based annotation with rough continuous traces, (2) polygon-based annotation with few vertices, and (3) box annotations combined with the Segment Anything Model (SAM). In situations where unlimited annotation time was available, precise annotations still lead to the highest segmentation model performance.

Medical image analysis (MedIA) has become an essential tool in medicine and healthcare, aiding in disease diagnosis, prognosis, and treatment planning, and recent successes in deep learning (DL) have made significant contributions to its advances. However, deploying DL models for MedIA in real-world situations remains challenging due to their failure to generalize across the distributional gap between training and testing samples - a problem known as domain shift. Researchers have dedicated their efforts to developing various DL methods to adapt and perform robustly on unknown and out-of-distribution data distributions. This paper comprehensively reviews domain generalization studies specifically tailored for MedIA. We provide a holistic view of how domain generalization techniques interact within the broader MedIA system, going beyond methodologies to consider the operational implications on the entire MedIA workflow. Specifically, we categorize domain generalization methods into data-level, feature-level, model-level, and analysis-level methods. We show how those methods can be used in various stages of the MedIA workflow with DL equipped from data acquisition to model prediction and analysis. Furthermore, we critically analyze the strengths and weaknesses of various methods, unveiling future research opportunities.

Large numbers of radiographic images are available in knee radiology practices which could be used for training of deep learning models for diagnosis of knee abnormalities. However, those images do not typically contain readily available labels due to limitations of human annotations. The purpose of our study was to develop an automated labeling approach that improves the image classification model to distinguish normal knee images from those with abnormalities or prior arthroplasty. The automated labeler was trained on a small set of labeled data to automatically label a much larger set of unlabeled data, further improving the image classification performance for knee radiographic diagnosis. We developed our approach using 7,382 patients and validated it on a separate set of 637 patients. The final image classification model, trained using both manually labeled and pseudo-labeled data, had the higher weighted average AUC (WAUC: 0.903) value and higher AUC-ROC values among all classes (normal AUC-ROC: 0.894; abnormal AUC-ROC: 0.896, arthroplasty AUC-ROC: 0.990) compared to the baseline model (WAUC=0.857; normal AUC-ROC: 0.842; abnormal AUC-ROC: 0.848, arthroplasty AUC-ROC: 0.987), trained using only manually labeled data. DeLong tests show that the improvement is significant on normal (p-value<0.002) and abnormal (p-value<0.001) images. Our findings demonstrated that the proposed automated labeling approach significantly improves the performance of image classification for radiographic knee diagnosis, allowing for facilitating patient care and curation of large knee datasets., Comment: This is the preprint version

The class imbalance problem in deep learning has been explored in several studies, but there has yet to be a systematic analysis of this phenomenon in object detection. Here, we present comprehensive analyses and experiments of the foreground-background (F-B) imbalance problem in object detection, which is very common and caused by small, infrequent objects of interest. We experimentally study the effects of different aspects of F-B imbalance (object size, number of objects, dataset size, object type) on detection performance. In addition, we also compare 9 leading methods for addressing this problem, including Faster-RCNN, SSD, OHEM, Libra-RCNN, Focal-Loss, GHM, PISA, YOLO-v3, and GFL with a range of datasets from different imaging domains. We conclude that (1) the F-B imbalance can indeed cause a significant drop in detection performance, (2) The detection performance is more affected by F-B imbalance when fewer training data are available, (3) in most cases, decreasing object size leads to larger performance drop than decreasing number of objects, given the same change in the ratio of object pixels to non-object pixels, (6) among all selected methods, Libra-RCNN and PISA demonstrate the best performance in addressing the issue of F-B imbalance. (7) When the training dataset size is large, the choice of method is not impactful (8) Soft-sampling methods, including focal-loss, GHM, and GFL, perform fairly well on average but are relatively unstable.

Shape learning, or the ability to leverage shape information, could be a desirable property of convolutional neural networks (CNNs) when target objects have specific shapes. While some research on the topic is emerging, there is no systematic study to conclusively determine whether and under what circumstances CNNs learn shape. Here, we present such a study in the context of segmentation networks where shapes are particularly important. We define shape and propose a new behavioral metric to measure the extent to which a CNN utilizes shape information. We then execute a set of experiments with synthetic and real-world data to progressively uncover under which circumstances CNNs learn shape and what can be done to encourage such behavior. We conclude that (i) CNNs do not learn shape in typical settings but rather rely on other features available to identify the objects of interest, (ii) CNNs can learn shape, but only if the shape is the only feature available to identify the object, (iii) sufficiently large receptive field size relative to the size of target objects is necessary for shape learning; (iv) a limited set of augmentations can encourage shape learning; (v) learning shape is indeed useful in the presence of out-of-distribution data., Comment: Accepted by Pattern Recognition, reformatted for Arxiv for readability

Automated tumor detection in Digital Breast Tomosynthesis (DBT) is a difficult task due to natural tumor rarity, breast tissue variability, and high resolution. Given the scarcity of abnormal images and the abundance of normal images for this problem, an anomaly detection/localization approach could be well-suited. However, most anomaly localization research in machine learning focuses on non-medical datasets, and we find that these methods fall short when adapted to medical imaging datasets. The problem is alleviated when we solve the task from the image completion perspective, in which the presence of anomalies can be indicated by a discrepancy between the original appearance and its auto-completion conditioned on the surroundings. However, there are often many valid normal completions given the same surroundings, especially in the DBT dataset, making this evaluation criterion less precise. To address such an issue, we consider pluralistic image completion by exploring the distribution of possible completions instead of generating fixed predictions. This is achieved through our novel application of spatial dropout on the completion network during inference time only, which requires no additional training cost and is effective at generating diverse completions. We further propose minimum completion distance (MCD), a new metric for detecting anomalies, thanks to these stochastic completions. We provide theoretical as well as empirical support for the superiority over existing methods of using the proposed method for anomaly localization. On the DBT dataset, our model outperforms other state-of-the-art methods by at least 10\% AUROC for pixel-level detection., Comment: Accepted in Medical Image Analysis (2023). Our code is at https://github.com/mazurowski-lab/picard

Training segmentation models for medical images continues to be challenging due to the limited availability of data annotations. Segment Anything Model (SAM) is a foundation model that is intended to segment user-defined objects of interest in an interactive manner. While the performance on natural images is impressive, medical image domains pose their own set of challenges. Here, we perform an extensive evaluation of SAM's ability to segment medical images on a collection of 19 medical imaging datasets from various modalities and anatomies. We report the following findings: (1) SAM's performance based on single prompts highly varies depending on the dataset and the task, from IoU=0.1135 for spine MRI to IoU=0.8650 for hip X-ray. (2) Segmentation performance appears to be better for well-circumscribed objects with prompts with less ambiguity and poorer in various other scenarios such as the segmentation of brain tumors. (3) SAM performs notably better with box prompts than with point prompts. (4) SAM outperforms similar methods RITM, SimpleClick, and FocalClick in almost all single-point prompt settings. (5) When multiple-point prompts are provided iteratively, SAM's performance generally improves only slightly while other methods' performance improves to the level that surpasses SAM's point-based performance. We also provide several illustrations for SAM's performance on all tested datasets, iterative segmentation, and SAM's behavior given prompt ambiguity. We conclude that SAM shows impressive zero-shot segmentation performance for certain medical imaging datasets, but moderate to poor performance for others. SAM has the potential to make a significant impact in automated medical image segmentation in medical imaging, but appropriate care needs to be applied when using it., Comment: Link to our code: https://github.com/mazurowski-lab/segment-anything-medical-evaluation

Three-dimensional segmentation in magnetic resonance images (MRI), which reflects the true shape of the objects, is challenging since high-resolution isotropic MRIs are rare and typical MRIs are anisotropic, with the out-of-plane dimension having a much lower resolution. A potential remedy to this issue lies in the fact that often multiple sequences are acquired on different planes. However, in practice, these sequences are not orthogonal to each other, limiting the applicability of many previous solutions to reconstruct higher-resolution images from multiple lower-resolution ones. We propose a weakly-supervised deep learning-based solution to generating high-resolution masks from multiple low-resolution images. Our method combines segmentation and unsupervised registration networks by introducing two new regularizations to make registration and segmentation reinforce each other. Finally, we introduce a multi-view fusion method to generate high-resolution target object masks. The experimental results on two datasets show the superiority of our methods. Importantly, the advantage of not using high-resolution images in the training process makes our method applicable to a wide variety of MRI segmentation tasks., Comment: 15 pages, 5 figures

The image acquisition parameters (IAPs) used to create MRI scans are central to defining the appearance of the images. Deep learning models trained on data acquired using certain parameters might not generalize well to images acquired with different parameters. Being able to recover such parameters directly from an image could help determine whether a deep learning model is applicable, and could assist with data harmonization and/or domain adaptation. Here, we introduce a neural network model that can predict many complex IAPs used to generate an MR image with high accuracy solely using the image, with a single forward pass. These predicted parameters include field strength, echo and repetition times, acquisition matrix, scanner model, scan options, and others. Even challenging parameters such as contrast agent type can be predicted with good accuracy. We perform a variety of experiments and analyses of our model's ability to predict IAPs on many MRI scans of new patients, and demonstrate its usage in a realistic application. Predicting IAPs from the images is an important step toward better understanding the relationship between image appearance and IAPs. This in turn will advance the understanding of many concepts related to the generalizability of neural network models on medical images, including domain shift, domain adaptation, and data harmonization., Comment: Paper accepted at MIDL 2023. Code available at https://github.com/mazurowski-lab/MRI-IAP-prediction

In this work we introduce a novel medical image style transfer method, StyleMapper, that can transfer medical scans to an unseen style with access to limited training data. This is made possible by training our model on unlimited possibilities of simulated random medical imaging styles on the training set, making our work more computationally efficient when compared with other style transfer methods. Moreover, our method enables arbitrary style transfer: transferring images to styles unseen in training. This is useful for medical imaging, where images are acquired using different protocols and different scanner models, resulting in a variety of styles that data may need to be transferred between. Methods: Our model disentangles image content from style and can modify an image's style by simply replacing the style encoding with one extracted from a single image of the target style, with no additional optimization required. This also allows the model to distinguish between different styles of images, including among those that were unseen in training. We propose a formal description of the proposed model. Results: Experimental results on breast magnetic resonance images indicate the effectiveness of our method for style transfer. Conclusion: Our style transfer method allows for the alignment of medical images taken with different scanners into a single unified style dataset, allowing for the training of other downstream tasks on such a dataset for tasks such as classification, object detection and others., Comment: preprint version, accepted in the Journal of Digital Imaging (JDIM). 16 pages (+ author names + references + supplementary), 6 figures

Objectives: The purpose is to apply a previously validated deep learning algorithm to a new thyroid nodule ultrasound image dataset and compare its performances with radiologists. Methods: Prior study presented an algorithm which is able to detect thyroid nodules and then make malignancy classifications with two ultrasound images. A multi-task deep convolutional neural network was trained from 1278 nodules and originally tested with 99 separate nodules. The results were comparable with that of radiologists. The algorithm was further tested with 378 nodules imaged with ultrasound machines from different manufacturers and product types than the training cases. Four experienced radiologists were requested to evaluate the nodules for comparison with deep learning. Results: The Area Under Curve (AUC) of the deep learning algorithm and four radiologists were calculated with parametric, binormal estimation. For the deep learning algorithm, the AUC was 0.69 (95% CI: 0.64 - 0.75). The AUC of radiologists were 0.63 (95% CI: 0.59 - 0.67), 0.66 (95% CI:0.61 - 0.71), 0.65 (95% CI: 0.60 - 0.70), and 0.63 (95%CI: 0.58 - 0.67). Conclusion: In the new testing dataset, the deep learning algorithm achieved similar performances with all four radiologists. The relative performance difference between the algorithm and the radiologists is not significantly affected by the difference of ultrasound scanner., Comment: Clinical Imaging (2023)

A fully-automated deep learning algorithm matched performance of radiologists in assessment of knee osteoarthritis severity in radiographs using the Kellgren-Lawrence grading system. To develop an automated deep learning-based algorithm that jointly uses Posterior-Anterior (PA) and Lateral (LAT) views of knee radiographs to assess knee osteoarthritis severity according to the Kellgren-Lawrence grading system. We used a dataset of 9739 exams from 2802 patients from Multicenter Osteoarthritis Study (MOST). The dataset was divided into a training set of 2040 patients, a validation set of 259 patients and a test set of 503 patients. A novel deep learning-based method was utilized for assessment of knee OA in two steps: (1) localization of knee joints in the images, (2) classification according to the KL grading system. Our method used both PA and LAT views as the input to the model. The scores generated by the algorithm were compared to the grades provided in the MOST dataset for the entire test set as well as grades provided by 5 radiologists at our institution for a subset of the test set. The model obtained a multi-class accuracy of 71.90% on the entire test set when compared to the ratings provided in the MOST dataset. The quadratic weighted Kappa coefficient for this set was 0.9066. The average quadratic weighted Kappa between all pairs of radiologists from our institution who took a part of study was 0.748. The average quadratic-weighted Kappa between the algorithm and the radiologists at our institution was 0.769. The proposed model performed demonstrated equivalency of KL classification to MSK radiologists, but clearly superior reproducibility. Our model also agreed with radiologists at our institution to the same extent as the radiologists with each other. The algorithm could be used to provide reproducible assessment of knee osteoarthritis severity.

The manifold hypothesis is a core mechanism behind the success of deep learning, so understanding the intrinsic manifold structure of image data is central to studying how neural networks learn from the data. Intrinsic dataset manifolds and their relationship to learning difficulty have recently begun to be studied for the common domain of natural images, but little such research has been attempted for radiological images. We address this here. First, we compare the intrinsic manifold dimensionality of radiological and natural images. We also investigate the relationship between intrinsic dimensionality and generalization ability over a wide range of datasets. Our analysis shows that natural image datasets generally have a higher number of intrinsic dimensions than radiological images. However, the relationship between generalization ability and intrinsic dimensionality is much stronger for medical images, which could be explained as radiological images having intrinsic features that are more difficult to learn. These results give a more principled underpinning for the intuition that radiological images can be more challenging to apply deep learning to than natural image datasets common to machine learning research. We believe rather than directly applying models developed for natural images to the radiological imaging domain, more care should be taken to developing architectures and algorithms that are more tailored to the specific characteristics of this domain. The research shown in our paper, demonstrating these characteristics and the differences from natural images, is an important first step in this direction., Comment: preprint version, accepted for MICCAI 2022 (25th International Conference on Medical Image Computing and Computer Assisted Intervention). 8 pages (+ author names + references + supplementary), 4 figures. Code available at https://github.com/mazurowski-lab/radiologyintrinsicmanifolds

Machine learning (ML) for diagnosis of thyroid nodules on ultrasound is an active area of research. However, ML tools require large, well-labelled datasets, the curation of which is time-consuming and labor-intensive. The purpose of our study was to develop and test a deep-learning-based tool to facilitate and automate the data annotation process for thyroid nodules; we named our tool Multistep Automated Data Labelling Procedure (MADLaP). MADLaP was designed to take multiple inputs included pathology reports, ultrasound images, and radiology reports. Using multiple step-wise modules including rule-based natural language processing, deep-learning-based imaging segmentation, and optical character recognition, MADLaP automatically identified images of a specific thyroid nodule and correctly assigned a pathology label. The model was developed using a training set of 378 patients across our health system and tested on a separate set of 93 patients. Ground truths for both sets were selected by an experienced radiologist. Performance metrics including yield (how many labeled images the model produced) and accuracy (percentage correct) were measured using the test set. MADLaP achieved a yield of 63% and an accuracy of 83%. The yield progressively increased as the input data moved through each module, while accuracy peaked part way through. Error analysis showed that inputs from certain examination sites had lower accuracy (40%) than the other sites (90%, 100%). MADLaP successfully created curated datasets of labeled ultrasound images of thyroid nodules. While accurate, the relatively suboptimal yield of MADLaP exposed some challenges when trying to automatically label radiology images from heterogeneous sources. The complex task of image curation and annotation could be automated, allowing for enrichment of larger datasets for use in machine learning development.

The assessment of knee osteoarthritis (KOA) severity on knee X-rays is a central criteria for the use of total knee arthroplasty. However, this assessment suffers from imprecise standards and a remarkably high inter-reader variability. An algorithmic, automated assessment of KOA severity could improve overall outcomes of knee replacement procedures by increasing the appropriateness of its use. We propose a novel deep learning-based five-step algorithm to automatically grade KOA from posterior-anterior (PA) views of radiographs: (1) image preprocessing (2) localization of knees joints in the image using the YOLO v3-Tiny model, (3) initial assessment of the severity of osteoarthritis using a convolutional neural network-based classifier, (4) segmentation of the joints and calculation of the joint space narrowing (JSN), and (5), a combination of the JSN and the initial assessment to determine a final Kellgren-Lawrence (KL) score. Furthermore, by displaying the segmentation masks used to make the assessment, our algorithm demonstrates a higher degree of transparency compared to typical "black box" deep learning classifiers. We perform a comprehensive evaluation using two public datasets and one dataset from our institution, and show that our algorithm reaches state-of-the art performance. Moreover, we also collected ratings from multiple radiologists at our institution and showed that our algorithm performs at the radiologist level. The software has been made publicly available at https://github.com/MaciejMazurowski/osteoarthritis-classification.

Research studies of artificial intelligence models in medical imaging have been hampered by poor generalization. This problem has been especially concerning over the last year with numerous applications of deep learning for COVID-19 diagnosis. Virtual imaging trials (VITs) could provide a solution for objective evaluation of these models. In this work utilizing the VITs, we created the CVIT-COVID dataset including 180 virtually imaged computed tomography (CT) images from simulated COVID-19 and normal phantom models under different COVID-19 morphology and imaging properties. We evaluated the performance of an open-source, deep-learning model from the University of Waterloo trained with multi-institutional data and an in-house model trained with the open clinical dataset called MosMed. We further validated the model's performance against open clinical data of 305 CT images to understand virtual vs. real clinical data performance. The open-source model was published with nearly perfect performance on the original Waterloo dataset but showed a consistent performance drop in external testing on another clinical dataset (AUC=0.77) and our simulated CVIT-COVID dataset (AUC=0.55). The in-house model achieved an AUC of 0.87 while testing on the internal test set (MosMed test set). However, performance dropped to an AUC of 0.65 and 0.69 when evaluated on clinical and our simulated CVIT-COVID dataset. The VIT framework offered control over imaging conditions, allowing us to show there was no change in performance as CT exposure was changed from 28.5 to 57 mAs. The VIT framework also provided voxel-level ground truth, revealing that performance of in-house model was much higher at AUC=0.87 for diffuse COVID-19 infection size >2.65% lung volume versus AUC=0.52 for focal disease with <2.65% volume. The virtual imaging framework enabled these uniquely rigorous analyses of model performance., Comment: 7 pages, 5 figures, 2 tables, presented at the Medical Imaging 2022: Computer-Aided Diagnosis, 2022

Organ segmentation of medical images is a key step in virtual imaging trials. However, organ segmentation datasets are limited in terms of quality (because labels cover only a few organs) and quantity (since case numbers are limited). In this study, we explored the tradeoffs between quality and quantity. Our goal is to create a unified approach for multi-organ segmentation of body CT, which will facilitate the creation of large numbers of accurate virtual phantoms. Initially, we compared two segmentation architectures, 3D-Unet and DenseVNet, which were trained using XCAT data that is fully labeled with 22 organs, and chose the 3D-Unet as the better performing model. We used the XCAT-trained model to generate pseudo-labels for the CT-ORG dataset that has only 7 organs segmented. We performed two experiments: First, we trained 3D-UNet model on the XCAT dataset, representing quality data, and tested it on both XCAT and CT-ORG datasets. Second, we trained 3D-UNet after including the CT-ORG dataset into the training set to have more quantity. Performance improved for segmentation in the organs where we have true labels in both datasets and degraded when relying on pseudo-labels. When organs were labeled in both datasets, Exp-2 improved Average DSC in XCAT and CT-ORG by 1. This demonstrates that quality data is the key to improving the model's performance., Comment: 6 pages, 3 figures, 2 tables, Accepted and Presented at SPIE Medical Imaging 2022

Lesion detection in digital breast tomosynthesis (DBT) is an important and a challenging problem characterized by a low prevalence of images containing tumors. Due to the label scarcity problem, large deep learning models and computationally intensive algorithms are likely to fail when applied to this task. In this paper, we present a practical yet lightweight backbone to improve the accuracy of tumor detection. Specifically, we propose a novel modification of visual transformer (ViT) on image feature patches to connect the feature patches of a tumor with healthy backgrounds of breast images and form a more robust backbone for tumor detection. To the best of our knowledge, our model is the first work of Transformer backbone object detection for medical imaging. Our experiments show that this model can considerably improve the accuracy of lesion detection and reduce the amount of labeled data required in typical ViT. We further show that with additional augmented tumor data, our model significantly outperforms the Faster R-CNN model and state-of-the-art SWIN transformer model.

What we expect from radiology AI algorithms will shape the selection and implementation of AI in the radiologic practice. In this paper I consider prevailing expectations of AI and compare them to expectations that we have of human readers. I observe that the expectations from AI and radiologists are fundamentally different. The expectations of AI are based on a strong and justified mistrust about the way that AI makes decisions. Because AI decisions are not well understood, it is difficult to know how the algorithms will behave in new, unexpected situations. However, this mistrust is not mirrored in our expectations of human readers. Despite well-proven idiosyncrasies and biases in human decision making, we take comfort from the assumption that others make decisions in a way as we do, and we trust our own decision making. Despite poor ability to explain decision making processes in humans, we accept explanations of decisions given by other humans. Because the goal of radiology is the most accurate radiologic interpretation, our expectations of radiologists and AI should be similar, and both should reflect a healthy mistrust of complicated and partially opaque decision processes undergoing in computer algorithms and human brains. This is generally not the case now.

Purpose: To develop high throughput multi-label annotators for body (chest, abdomen, and pelvis) Computed Tomography (CT) reports that can be applied across a variety of abnormalities, organs, and disease states. Approach: We used a dictionary approach to develop rule-based algorithms (RBA) for extraction of disease labels from radiology text reports. We targeted three organ systems (lungs/pleura, liver/gallbladder, kidneys/ureters) with four diseases per system based on their prevalence in our dataset. To expand the algorithms beyond pre-defined keywords, attention-guided recurrent neural networks (RNN) were trained using the RBA-extracted labels to classify reports as being positive for one or more diseases or normal for each organ system. Confounding effects on model performance were evaluated using random initialization or pre-trained embedding as well as different sizes of training datasets. Performance was evaluated using the receiver operating characteristic (ROC) area under the curve (AUC) against 2,158 manually obtained labels. Results: Our models extracted disease labels from 261,229 radiology reports of 112,501 unique subjects. Pre-trained models outperformed random initialization across all diseases. As the training dataset size was reduced, performance was robust except for a few diseases with relatively small number of cases. Pre-trained classification AUCs achieved > 0.95 for all five disease outcomes across all three organ systems. Conclusions: Our label-extracting pipeline was able to encompass a variety of cases and diseases by generalizing beyond strict rules with exceptional accuracy. This method can be easily adapted to enable automated labeling of hospital-scale medical data sets for training image-based disease classifiers.

Breast cancer screening is one of the most common radiological tasks with over 39 million exams performed each year. While breast cancer screening has been one of the most studied medical imaging applications of artificial intelligence, the development and evaluation of the algorithms are hindered due to the lack of well-annotated large-scale publicly available datasets. This is particularly an issue for digital breast tomosynthesis (DBT) which is a relatively new breast cancer screening modality. We have curated and made publicly available a large-scale dataset of digital breast tomosynthesis images. It contains 22,032 reconstructed DBT volumes belonging to 5,610 studies from 5,060 patients. This included four groups: (1) 5,129 normal studies, (2) 280 studies where additional imaging was needed but no biopsy was performed, (3) 112 benign biopsied studies, and (4) 89 studies with cancer. Our dataset included masses and architectural distortions which were annotated by two experienced radiologists. Additionally, we developed a single-phase deep learning detection model and tested it using our dataset to serve as a baseline for future research. Our model reached a sensitivity of 65% at 2 false positives per breast. Our large, diverse, and highly-curated dataset will facilitate development and evaluation of AI algorithms for breast cancer screening through providing data for training as well as common set of cases for model validation. The performance of the model developed in our study shows that the task remains challenging and will serve as a baseline for future model development.

Weakly supervised disease classification of CT imaging suffers from poor localization owing to case-level annotations, where even a positive scan can hold hundreds to thousands of negative slices along multiple planes. Furthermore, although deep learning segmentation and classification models extract distinctly unique combinations of anatomical features from the same target class(es), they are typically seen as two independent processes in a computer-aided diagnosis (CAD) pipeline, with little to no feature reuse. In this research, we propose a medical classifier that leverages the semantic structural concepts learned via multi-resolution segmentation feature maps, to guide weakly supervised 3D classification of chest CT volumes. Additionally, a comparative analysis is drawn across two different types of feature aggregation to explore the vast possibilities surrounding feature fusion. Using a dataset of 1593 scans labeled on a case-level basis via rule-based model, we train a dual-stage convolutional neural network (CNN) to perform organ segmentation and binary classification of four representative diseases (emphysema, pneumonia/atelectasis, mass and nodules) in lungs. The baseline model, with separate stages for segmentation and classification, results in AUC of 0.791. Using identical hyperparameters, the connected architecture using static and dynamic feature aggregation improves performance to AUC of 0.832 and 0.851, respectively. This study advances the field in two key ways. First, case-level report data is used to weakly supervise a 3D CT classifier of multiple, simultaneous diseases for an organ. Second, segmentation and classification models are connected with two different feature aggregation strategies to enhance the classification performance., Comment: Accepted to 2020 SPIE Medical Imaging: Computer-Aided Diagnosis [oral presentation]

Purpose: To design multi-disease classifiers for body CT scans for three different organ systems using automatically extracted labels from radiology text reports.Materials & Methods: This retrospective study included a total of 12,092 patients (mean age 57 +- 18; 6,172 women) for model development and testing (from 2012-2017). Rule-based algorithms were used to extract 19,225 disease labels from 13,667 body CT scans from 12,092 patients. Using a three-dimensional DenseVNet, three organ systems were segmented: lungs and pleura; liver and gallbladder; and kidneys and ureters. For each organ, a three-dimensional convolutional neural network classified no apparent disease versus four common diseases for a total of 15 different labels across all three models. Testing was performed on a subset of 2,158 CT volumes relative to 2,875 manually derived reference labels from 2133 patients (mean age 58 +- 18;1079 women). Performance was reported as receiver operating characteristic area under the curve (AUC) with 95% confidence intervals by the DeLong method. Results: Manual validation of the extracted labels confirmed 91% to 99% accuracy across the 15 different labels. AUCs for lungs and pleura labels were: atelectasis 0.77 (95% CI: 0.74, 0.81), nodule 0.65 (0.61, 0.69), emphysema 0.89 (0.86, 0.92), effusion 0.97 (0.96, 0.98), and no apparent disease 0.89 (0.87, 0.91). AUCs for liver and gallbladder were: hepatobiliary calcification 0.62 (95% CI: 0.56, 0.67), lesion 0.73 (0.69, 0.77), dilation 0.87 (0.84, 0.90), fatty 0.89 (0.86, 0.92), and no apparent disease 0.82 (0.78, 0.85). AUCs for kidneys and ureters were: stone 0.83 (95% CI: 0.79, 0.87), atrophy 0.92 (0.89, 0.94), lesion 0.68 (0.64, 0.72), cyst 0.70 (0.66, 0.73), and no apparent disease 0.79 (0.75, 0.83). Conclusion: Weakly-supervised deep learning models were able to classify diverse diseases in multiple organ systems., Comment: 22 pages, 6 figures, 2 tables; Accepted for publication at Radiology: Artificial Intelligence

Machine learning models for radiology benefit from large-scale data sets with high quality labels for abnormalities. We curated and analyzed a chest computed tomography (CT) data set of 36,316 volumes from 19,993 unique patients. This is the largest multiply-annotated volumetric medical imaging data set reported. To annotate this data set, we developed a rule-based method for automatically extracting abnormality labels from free-text radiology reports with an average F-score of 0.976 (min 0.941, max 1.0). We also developed a model for multi-organ, multi-disease classification of chest CT volumes that uses a deep convolutional neural network (CNN). This model reached a classification performance of AUROC greater than 0.90 for 18 abnormalities, with an average AUROC of 0.773 for all 83 abnormalities, demonstrating the feasibility of learning from unfiltered whole volume CT data. We show that training on more labels improves performance significantly: for a subset of 9 labels - nodule, opacity, atelectasis, pleural effusion, consolidation, mass, pericardial effusion, cardiomegaly, and pneumothorax - the model's average AUROC increased by 10% when the number of training labels was increased from 9 to all 83. All code for volume preprocessing, automated label extraction, and the volume abnormality prediction model will be made publicly available. The 36,316 CT volumes and labels will also be made publicly available pending institutional approval., Comment: 20 pages, 3 figures, 5 tables (appendices additional). Published in Medical Image Analysis (October 2020)

Dynamic Contrast Enhanced-Magnetic Resonance Imaging (DCE-MRI) is widely used to complement ultrasound examinations and x-ray mammography during the early detection and diagnosis of breast cancer. However, images generated by various MRI scanners (e.g. GE Healthcare vs Siemens) differ both in intensity and noise distribution, preventing algorithms trained on MRIs from one scanner to generalize to data from other scanners successfully. We propose a method for image normalization to solve this problem. MRI normalization is challenging because it requires both normalizing intensity values and mapping between the noise distributions of different scanners. We utilize a cycle-consistent generative adversarial network to learn a bidirectional mapping between MRIs produced by GE Healthcare and Siemens scanners. This allows us learning the mapping between two different scanner types without matched data, which is not commonly available. To ensure the preservation of breast shape and structures within the breast, we propose two technical innovations. First, we incorporate a mutual information loss with the CycleGAN architecture to ensure that the structure of the breast is maintained. Second, we propose a modified discriminator architecture which utilizes a smaller field-of-view to ensure the preservation of finer details in the breast tissue. Quantitative and qualitative evaluations show that the second proposed method was able to consistently preserve a high level of detail in the breast structure while also performing the proper intensity normalization and noise mapping. Our results demonstrate that the proposed model can successfully learn a bidirectional mapping between MRIs produced by different vendors, potentially enabling improved accuracy of downstream computational algorithms for diagnosis and detection of breast cancer. All the data used in this study are publicly available., Comment: Final accepted draft in Computer Methods and Programs in Biomedicine

Recent analysis identified distinct genomic subtypes of lower-grade glioma tumors which are associated with shape features. In this study, we propose a fully automatic way to quantify tumor imaging characteristics using deep learning-based segmentation and test whether these characteristics are predictive of tumor genomic subtypes. We used preoperative imaging and genomic data of 110 patients from 5 institutions with lower-grade gliomas from The Cancer Genome Atlas. Based on automatic deep learning segmentations, we extracted three features which quantify two-dimensional and three-dimensional characteristics of the tumors. Genomic data for the analyzed cohort of patients consisted of previously identified genomic clusters based on IDH mutation and 1p/19q co-deletion, DNA methylation, gene expression, DNA copy number, and microRNA expression. To analyze the relationship between the imaging features and genomic clusters, we conducted the Fisher exact test for 10 hypotheses for each pair of imaging feature and genomic subtype. To account for multiple hypothesis testing, we applied a Bonferroni correction. P-values lower than 0.005 were considered statistically significant. We found the strongest association between RNASeq clusters and the bounding ellipsoid volume ratio ($p<0.0002$) and between RNASeq clusters and margin fluctuation ($p<0.005$). In addition, we identified associations between bounding ellipsoid volume ratio and all tested molecular subtypes ($p<0.02$) as well as between angular standard deviation and RNASeq cluster ($p<0.02$). In terms of automatic tumor segmentation that was used to generate the quantitative image characteristics, our deep learning algorithm achieved a mean Dice coefficient of 82% which is comparable to human performance.

Objective: To develop an automatic image normalization algorithm for intensity correction of images from breast dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) acquired by different MRI scanners with various imaging parameters, using only image information. Methods: DCE-MR images of 460 subjects with breast cancer acquired by different scanners were used in this study. Each subject had one T1-weighted pre-contrast image and three T1-weighted post-contrast images available. Our normalization algorithm operated under the assumption that the same type of tissue in different patients should be represented by the same voxel value. We used four tissue/material types as the anchors for the normalization: 1) air, 2) fat tissue, 3) dense tissue, and 4) heart. The algorithm proceeded in the following two steps: First, a state-of-the-art deep learning-based algorithm was applied to perform tissue segmentation accurately and efficiently. Then, based on the segmentation results, a subject-specific piecewise linear mapping function was applied between the anchor points to normalize the same type of tissue in different patients into the same intensity ranges. We evaluated the algorithm with 300 subjects used for training and the rest used for testing. Results: The application of our algorithm to images with different scanning parameters resulted in highly improved consistency in pixel values and extracted radiomics features. Conclusion: The proposed image normalization strategy based on tissue segmentation can perform intensity correction fully automatically, without the knowledge of the scanner parameters. Significance: We have thoroughly tested our algorithm and showed that it successfully normalizes the intensity of DCE-MR images. We made our software publicly available for others to apply in their analyses.

Deep learning is a branch of artificial intelligence where networks of simple interconnected units are used to extract patterns from data in order to solve complex problems. Deep learning algorithms have shown groundbreaking performance in a variety of sophisticated tasks, especially those related to images. They have often matched or exceeded human performance. Since the medical field of radiology mostly relies on extracting useful information from images, it is a very natural application area for deep learning, and research in this area has rapidly grown in recent years. In this article, we review the clinical reality of radiology and discuss the opportunities for application of deep learning algorithms. We also introduce basic concepts of deep learning including convolutional neural networks. Then, we present a survey of the research in deep learning applied to radiology. We organize the studies by the types of specific tasks that they attempt to solve and review the broad range of utilized deep learning algorithms. Finally, we briefly discuss opportunities and challenges for incorporating deep learning in the radiology practice of the future., Comment: 27 pages, 4 figures

Purpose: To determine whether deep learning models can distinguish between breast cancer molecular subtypes based on dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI). Materials and methods: In this institutional review board-approved single-center study, we analyzed DCE-MR images of 270 patients at our institution. Lesions of interest were identified by radiologists. The task was to automatically determine whether the tumor is of the Luminal A subtype or of another subtype based on the MR image patches representing the tumor. Three different deep learning approaches were used to classify the tumor according to their molecular subtypes: learning from scratch where only tumor patches were used for training, transfer learning where networks pre-trained on natural images were fine-tuned using tumor patches, and off-the-shelf deep features where the features extracted by neural networks trained on natural images were used for classification with a support vector machine. Network architectures utilized in our experiments were GoogleNet, VGG, and CIFAR. We used 10-fold crossvalidation method for validation and area under the receiver operating characteristic (AUC) as the measure of performance. Results: The best AUC performance for distinguishing molecular subtypes was 0.65 (95% CI:[0.57,0.71]) and was achieved by the off-the-shelf deep features approach. The highest AUC performance for training from scratch was 0.58 (95% CI:[0.51,0.64]) and the best AUC performance for transfer learning was 0.60 (95% CI:[0.52,0.65]) respectively. For the off-the-shelf approach, the features extracted from the fully connected layer performed the best. Conclusion: Deep learning may play a role in discovering radiogenomic associations in breast cancer.

Purpose: To determine whether deep learning-based algorithms applied to breast MR images can aid in the prediction of occult invasive disease following the di- agnosis of ductal carcinoma in situ (DCIS) by core needle biopsy. Material and Methods: In this institutional review board-approved study, we analyzed dynamic contrast-enhanced fat-saturated T1-weighted MRI sequences of 131 patients at our institution with a core needle biopsy-confirmed diagnosis of DCIS. The patients had no preoperative therapy before breast MRI and no prior history of breast cancer. We explored two different deep learning approaches to predict whether there was a hidden (occult) invasive component in the analyzed tumors that was ultimately detected at surgical excision. In the first approach, we adopted the transfer learning strategy, in which a network pre-trained on a large dataset of natural images is fine-tuned with our DCIS images. Specifically, we used the GoogleNet model pre-trained on the ImageNet dataset. In the second approach, we used a pre-trained network to extract deep features, and a support vector machine (SVM) that utilizes these features to predict the upstaging of the DCIS. We used 10-fold cross validation and the area under the ROC curve (AUC) to estimate the performance of the predictive models. Results: The best classification performance was obtained using the deep features approach with GoogleNet model pre-trained on ImageNet as the feature extractor and a polynomial kernel SVM used as the classifier (AUC = 0.70, 95% CI: 0.58- 0.79). For the transfer learning based approach, the highest AUC obtained was 0.53 (95% CI: 0.41-0.62). Conclusion: Convolutional neural networks could potentially be used to identify occult invasive disease in patients diagnosed with DCIS at the initial core needle biopsy.