Search

Purpose: Deep learning (DL) algorithms have shown promising results for brain tumor segmentation in MRI. However, validation is required prior to routine clinical use. We report the first randomized and blinded comparison of DL and trained technician segmentations. Approach: We compiled a multi-institutional database of 741 pretreatment MRI exams. Each contained a postcontrast T1-weighted exam, a T2-weighted fluid-attenuated inversion recovery exam, and at least one technician-derived tumor segmentation. The database included 729 unique patients (470 males and 259 females). Of these exams, 641 were used for training the DL system, and 100 were reserved for testing. We developed a platform to enable qualitative, blinded, controlled assessment of lesion segmentations made by technicians and the DL method. On this platform, 20 neuroradiologists performed 400 side-by-side comparisons of segmentations on 100 test cases. They scored each segmentation between 0 (poor) and 10 (perfect). Agreement between segmentations from technicians and the DL method was also evaluated quantitatively using the Dice coefficient, which produces values between 0 (no overlap) and 1 (perfect overlap). Results: The neuroradiologists gave technician and DL segmentations mean scores of 6.97 and 7.31, respectively (p

We develop neural-network-based methods for classifying plaque types in clinical intravascular optical coherence tomography (IVOCT) images of coronary arteries. A single IVOCT pullback can consist of >500 microscopic-resolution images, creating both a challenge for physician interpretation during an interventional procedure and an opportunity for automated analysis. In the proposed method, we classify each A-line, a datum element that better captures physics and pathophysiology than a voxel, as a fibrous layer followed by calcification (fibrocalcific), a fibrous layer followed by a lipidous deposit (fibrolipidic), or other. For A-line classification, the usefulness of a convolutional neural network (CNN) is compared with that of a fully connected artificial neural network (ANN). A total of 4469 image frames across 48 pullbacks that are manually labeled using consensus labeling from two experts are used for training, evaluation, and testing. A 10-fold cross-validation using held-out pullbacks is applied to assess classifier performance. Noisy A-line classifications are cleaned by applying a conditional random field (CRF) and morphological processing to pullbacks in the en-face view. With CNN (ANN) approaches, we achieve an accuracy of 77.7 % ± 4.1 % (79.4 % ± 2.9 % ) for fibrocalcific, 86.5 % ± 2.3 % (83.4 % ± 2.6 % ) for fibrolipidic, and 85.3 % ± 2.5 % (82.4 % ± 2.2 % ) for other, across all folds following CRF noise cleaning. The results without CRF cleaning are typically reduced by 10% to 15%. The enhanced performance of the CNN was likely due to spatial invariance of the convolution operation over the input A-line. The predicted en-face classification maps of entire pullbacks agree favorably to the annotated counterparts. In some instances, small error regions are actually hard to call when re-examined by human experts. Even in worst-case pullbacks, it can be argued that the results will not negatively impact usage by physicians, as there is a preponderance of correct calls.

The segmentation of organs at risk is a crucial and time-consuming step in radiotherapy planning. Good automatic methods can significantly reduce the time clinicians have to spend on this task. Due to its variability in shape and low contrast to surrounding structures, segmenting the parotid gland is challenging. Motivated by the recent success of deep learning, we study the use of two-dimensional (2-D), 2-D ensemble, and three-dimensional (3-D) U-Nets for segmentation. The mean Dice similarity to ground truth is ∼0.83 for all three models. A patch-based approach for class balancing seems promising for false-positive reduction. The 2-D ensemble and 3-D U-Net are applied to the test data of the 2015 MICCAI challenge on head and neck autosegmentation. Both deep learning methods generalize well onto independent data (Dice 0.865 and 0.88) and are superior to a selection of model- and atlas-based methods with respect to the Dice coefficient. Since appropriate reference annotations are essential for training but often difficult and expensive to obtain, it is important to know how many samples are needed for training. We evaluate the performance after training with different-sized training sets and observe no significant increase in the Dice coefficient for more than 250 training cases.