Search

To develop a U-net deep learning method for breast tissue segmentation on fat-sat T1-weighted (T1W) MRI using transfer learning (TL) from a model developed for non-fat-sat images. The training dataset (N = 126) was imaged on a 1.5 T MR scanner, and the independent testing dataset (N = 40) was imaged on a 3 T scanner, both using fat-sat T1W pulse sequence. Pre-contrast images acquired in the dynamic-contrast-enhanced (DCE) MRI sequence were used for analysis. All patients had unilateral cancer, and the segmentation was performed using the contralateral normal breast. The ground truth of breast and fibroglandular tissue (FGT) segmentation was generated using a template-based segmentation method with a clustering algorithm. The deep learning segmentation was performed using U-net models trained with and without TL, by using initial values of trainable parameters taken from the previous model for non-fat-sat images. The ground truth of each case was used to evaluate the segmentation performance of the U-net models by calculating the dice similarity coefficient (DSC) and the overall accuracy based on all pixels. Pearson’s correlation was used to evaluate the correlation of breast volume and FGT volume between the U-net prediction output and the ground truth. In the training dataset, the evaluation was performed using tenfold cross-validation, and the mean DSC with and without TL was 0.97 vs. 0.95 for breast and 0.86 vs. 0.80 for FGT. When the final model developed with and without TL from the training dataset was applied to the testing dataset, the mean DSC was 0.89 vs. 0.83 for breast and 0.81 vs. 0.81 for FGT, respectively. Application of TL not only improved the DSC, but also decreased the required training case number. Lastly, there was a high correlation (R2 > 0.90) for both the training and testing datasets between the U-net prediction output and ground truth for breast volume and FGT volume. U-net can be applied to perform breast tissue segmentation on fat-sat images, and TL is an efficient strategy to develop a specific model for each different dataset.

PURPOSE: To improve image quality and CT number accuracy of daily cone-beam computed tomography (CBCT) through a deep-learning methodology with Generative Adversarial Network. METHODS: 150 paired pelvic CT and CBCT scans were used for model training and validation. An unsupervised deep-learning method, 2.5D pixel-to-pixel generative adversarial network (GAN) model with feature mapping was proposed. A total of 12000 slice pairs of CT and CBCT were used for model training, while 10-cross validation was applied to verify model robustness. Paired CT-CBCT scans from an additional 15 pelvic patients and 10 head-and-neck (HN) patients with CBCT images collected at a different machine were used for independent testing purpose. Besides the proposed method above, other network architectures were also tested as: 2D vs. 2.5D; GAN model with vs. without feature mapping; GAN model with vs. without additional perceptual loss; and previously reported models as U-net and cycleGAN with or without identity loss. Image quality of deep-learning generated synthetic CT (sCT) images were quantitatively compared against the reference CT (rCT) image using mean absolute error (MAE) of Hounsfield units (HU) and peak signal-to-noise ratio (PSNR). The dosimetric calculation accuracy was further evaluated with both photon and proton beams. RESULTS: The deep-learning generated synthetic CTs (sCT) showed improved image quality with reduced artifact distortion and improved soft tissue contrast. The proposed algorithm of 2.5 Pix2pix GAN with feature matching (FM) was shown to be the best model among all tested methods producing the highest PSNR and the lowest MAE to reference CT (rCT). The dose distribution demonstrated a high accuracy in the scope of photon based planning, yet more work is needed for proton based treatment. Once the model was trained, it took 11–12 ms to process one slice, and could generate a 3D-volume of dCBCT (80 slices) in less than a second using a NVIDIA GeForce GTX Titan X GPU (12GB, Maxwell architecture). CONCLUSION: The proposed deep-learning algorithm is promising to improve CBCT image quality in an efficient way, thus has a potential to support online CBCT-based adaptive radiotherapy.

Background: The combination of anastrozole and fulvestrant (A+F) is associated with improvement in overall survival (OS) as compared to anastrozole (A) in patients (pts) with hormone receptor-positive (HR+) breast cancer (BC), specifically in pts with endocrine naïve disease or in pts with long disease-free interval (Mehta et al. NEJM 2019). We performed an IRB-approved retrospective assessment of the time-to-treatment failure (TTF) and overall survival (OS) outcomes in pts treated at our institution with A+F, and assessed the impact of A+F in various molecularly defined subsets of pts. Methods: We reviewed charts of 118 pts with advanced HR+ BC who received A+F. Pts with brain metastases were not excluded. Performance status (PS) ranged from 0-3. We compared TTF and OS outcomes in pts with and without molecular aberrations including HER2 overexpression/amplification, or mutations in PIK3CA, ESR1 or BRCA1/2. HER2 status was tested through tumor. PIK3CA and ESR1 were tested on tumor or ctDNA by next generation sequencing. Germline/somatic BRCA1/2 status was assessed either through ctDNA or germline testing. Kaplan-Meir survival curves were constructed. Primary statistical analysis was log-rank test, followed by Cox regression to estimate the hazard ratio and 95% confidence intervals. All p values are 2-sided. Results: Overall 68 patients had endocrine-sensitive (ES) disease and 50 patients had acquired endocrine-resistant (ER) disease. Median TTF was 26 months in the ES group vs. 10 months in the ER group (HR: 0.47;95% Cl: 0.30 - 0.75, p= 0.001). Median OS was 57 months in the ES group vs. 38 months in the ER group (HR: 0.51; 95% Cl: 0.30 - 0.86, p= 0.01). Among the 31 pts tested for PIK3CA mutation, 21 pts were negative (control) and 10 pts were positive (PIK3CA mutant). Median TTF was 23 months for control vs.14 months for PIK3CA mutant (HR: 0.89; 95% CI: 0.40 -1.95, p = 0.76). Median OS was not reached for control vs. 44 months for PIK3CA mutant (HR: 0.50; 95% Cl: 0.16 - 1.56, p = 0.23). Among the 35 pts tested for ESR1 mutation, 27 pts were negative (control) and 8 patients were positive (ESR1 mutant). Median TTF was 26 months for control vs. 10 months for ESR1 mutant (HR: 0.59; 95% CI: 0.23 - 1.52, p = 0.20). Median OS was 66 months for control vs. 65 months for ESR1 mutant (HR: 1.63; 95% Cl: 0.44 - 5.95, p = 0.46). In the 17 pts tested for BRCA mutation (germline or somatic), 6 of them were positive (BRCA mutant) and 11 of them were negative (control). Median TTF was 36 months for control vs. 10 months for BRCA mutant (HR: 0.40; 95% CI: 0.12 - 1.40, p = 0.09). Median OS was 35 months for control but has not been reached in the BRCA mutant group (HR: 1.16; 95% Cl: 0.23 - 5.88, p = 0.93). Of 118 patients, 92 pts were HER2 negative (control) and 26 patients were HER2 positive. Median TTF was 23 months for control vs. 14 months for HER2+ group (HR: 0.80; 95% CI: 0.46 - 1.39, p = 0.42). Median OS was 52 months for control vs. 40 months for HER2+ group (HR: 0.87; 95% Cl: 0.47 - 1.64, p = 0.67). Conclusions: The time-to-treatment failure and overall survival in the endocrine-sensitive disease are 2-fold higher compared to the endocrine-resistant disease. Patients with PIK3CA, ESR1 or BRCA1/2 mutation or with HER2 overexpression/amplification historically have poor survival outcomes, but in hypothesis-generating analysis, we showed statistically similar overall survival outcomes in these patients treated with anastrozole plus fulvestrant, compared to controls without these poor prognostic features. We suggest that anastrozole plus fulvestrant should be the preferred partner with molecularly targeted agents. Citation Format: Tianyi Tang, Yu Cheng, Ritesh Parajuli, Neha Patel, Min-Ying Su, Chaitali Nangia, Rita S Mehta. Overall survival outcomes of combination anastrozole and fulvestrant in molecular subsets of hormone receptor-positive breast cancer [abstract]. In: Proceedings of the 2020 San Antonio Breast Cancer Virtual Symposium; 2020 Dec 8-11; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2021;81(4 Suppl):Abstract nr PS5-32.

PURPOSE: To improve the performance of less experienced clinicians in the diagnosis of benign and malignant spinal fracture on MRI, we applied the ResNet50 algorithm to develop a decision support system. METHODS: A total of 190 patients, 50 with malignant and 140 with benign fractures, were studied. The visual diagnosis was made by one senior MSK radiologist, one fourth-year resident, and one first-year resident. The MSK radiologist also gave the binary score for 15 qualitative imaging features. Deep learning was implemented using ResNet50, using one abnormal spinal segment selected from each patient as input. The T1W and T2W images of the lesion slice and its two neighboring slices were considered. The diagnostic performance was evaluated using tenfold cross-validation. RESULTS: The overall reading accuracy was 98, 96, and 66% for the senior MSK radiologist, fourth-year resident, and first-year resident, respectively. Of the 15 imaging features, 10 showed a significant difference between benign and malignant groups with p < = 0.001. The accuracy achieved by using the ResNet50 deep learning model for the identified abnormal vertebral segment was 92%. Compared to the first-year resident’s reading, the model improved the sensitivity from 78 to 94% (p < 0.001) and the specificity from 61 to 91% (p < 0.001). CONCLUSION: Our deep learning-based model may provide information to assist less experienced clinicians in the diagnosis of spinal fractures on MRI. Other findings away from the vertebral body need to be considered to improve the model, and further investigation is required to generalize our findings to real-world settings.

Football exposes its players to traumatic brain, neck, and spinal injury. It is unknown whether the adolescent football player develops imaging abnormalities of the brain and spine that are detectable on magnetic resonance imaging (MRI). The objective of this observational study was to identify potential MRI signatures of early brain and cervical spine (c-spine) injury in high school football players. Eighteen football players (mean age, 17.0 ± 1.5 years; mean career length, 6.3 ± 4.0 years) had a baseline brain MRI, and 7 had a follow-up scan 9-42 months later. C-spine MRIs were performed on 11 of the 18 subjects, and 5 had a follow-up scan. C-spine MRIs from 12 age-matched hospital controls were also retrospectively retrieved. Brain MRIs were reviewed by a neuroradiologist, and no cerebral microbleeds were detected. Three readers (a neuroradiologist, a neurosurgeon, and an orthopedic spine surgeon) studied the cervical intervertebral discs at six different cervical levels and graded degeneration using an established five-grade scoring system. We observed no statistically significant difference in disc degeneration or any trend toward increased disc degeneration in the c-spine of football players as compared with age-matched controls. Further research is needed to validate our findings and better understand the true impact of contact sports on young athletes.

Rationale and Objectives Breast segmentation using the U-net architecture was implemented and tested in independent validation datasets to quantify fibroglandular tissue volume in breast MRI. Materials and Methods Two datasets were used. The training set was MRI of 286 patients with unilateral breast cancer. The segmentation was done on the contralateral normal breasts. The ground truth for the breast and fibroglandular tissue (FGT) was obtained by using a template-based segmentation method. The U-net deep learning algorithm was implemented to analyze the training set, and the final model was obtained using 10-fold cross-validation. The independent validation set was MRI of 28 normal volunteers acquired using four different MR scanners. Dice Similarity Coefficient (DSC), voxel-based accuracy, and Pearson's correlation were used to evaluate the performance. Results For the 10-fold cross-validation in the initial training set of 286 patients, the DSC range was 0.83–0.98 (mean 0.95 ± 0.02) for breast and 0.73–0.97 (mean 0.91 ± 0.03) for FGT; and the accuracy range was 0.92–0.99 (mean 0.98 ± 0.01) for breast and 0.87–0.99 (mean 0.97 ± 0.01) for FGT. For the entire 224 testing breasts of the 28 normal volunteers in the validation datasets, the mean DSC was 0.86 ± 0.05 for breast, 0.83 ± 0.06 for FGT; and the mean accuracy was 0.94 ± 0.03 for breast and 0.93 ± 0.04 for FGT. The testing results for MRI acquired using four different scanners were comparable. Conclusion Deep learning based on the U-net algorithm can achieve accurate segmentation results for the breast and FGT on MRI. It may provide a reliable and efficient method to process large number of MR images for quantitative analysis of breast density.

ObjectivesTo evaluate the diagnostic performance of Kaiser score (KS) adjusted with the apparent diffusion coefficient (ADC) (KS+) and machine learning (ML) modeling.MethodsA dataset of 402 malignant and 257 benign lesions was identified. Two radiologists assigned the KS. If a lesion with KS > 4 had ADC > 1.4 × 10-3 mm2/s, the KS was reduced by 4 to become KS+. In order to consider the full spectrum of ADC as a continuous variable, the KS and ADC values were used to train diagnostic models using 5 ML algorithms. The performance was evaluated using the ROC analysis, compared by the DeLong test. The sensitivity, specificity, and accuracy achieved using the threshold of KS > 4, KS+ > 4, and ADC ≤ 1.4 × 10-3 mm2/s were obtained and compared by the McNemar test.ResultsThe ROC curves of KS, KS+, and all ML models had comparable AUC in the range of 0.883-0.921, significantly higher than that of ADC (0.837, p < 0.0001). The KS had sensitivity = 97.3% and specificity = 59.1%; and the KS+ had sensitivity = 95.5% with significantly improved specificity to 68.5% (p < 0.0001). However, when setting at the same sensitivity of 97.3%, KS+ could not improve specificity. In ML analysis, the logistic regression model had the best performance. At sensitivity = 97.3% and specificity = 65.3%, i.e., compared to KS, 16 false-positives may be avoided without affecting true cancer diagnosis (p = 0.0015).ConclusionUsing dichotomized ADC to modify KS to KS+ can improve specificity, but at the price of lowered sensitivity. Machine learning algorithms may be applied to consider the ADC as a continuous variable to build more accurate diagnostic models.Key points• When using ADC to modify the Kaiser score to KS+, the diagnostic specificity according to the results of two independent readers was improved by 9.4-9.7%, at the price of slightly degraded sensitivity by 1.5-1.8%, and overall had improved accuracy by 2.6-2.9%. • When the KS and the continuous ADC values were combined to train models by machine learning algorithms, the diagnostic specificity achieved by the logistic regression model could be significantly improved from 59.1 to 65.3% (p = 0.0015), while maintaining at the high sensitivity of KS = 97.3%, and thus, the results demonstrated the potential of ML modeling to further evaluate the contribution of ADC. • When setting the sensitivity at the same levels, the modified KS+ and the original KS have comparable specificity; therefore, KS+ with consideration of ADC may not offer much practical help, and the original KS without ADC remains as an excellent robust diagnostic method.

ObjectiveTo build radiomics models using features extracted from DCE-MRI and mammography for diagnosis of breast cancer.Materials and Methods266 patients receiving MRI and mammography, who had well-enhanced lesions on MRI and histologically confirmed diagnosis were analyzed. Training dataset had 146 malignant and 56 benign, and testing dataset had 48 malignant and 18 benign lesions. Fuzzy-C-means clustering algorithm was used to segment the enhanced lesion on subtraction MRI maps. Two radiologists manually outlined the corresponding lesion on mammography by consensus, with the guidance of MRI maximum intensity projection. Features were extracted using PyRadiomics from three DCE-MRI parametric maps, and from the lesion and a 2-cm bandshell margin on mammography. The support vector machine (SVM) was applied for feature selection and model building, using 5 datasets: DCE-MRI, mammography lesion-ROI, mammography margin-ROI, mammography lesion+margin, and all combined.ResultsIn the training dataset evaluated using 10-fold cross-validation, the diagnostic accuracy of the individual model was 83.2% for DCE-MRI, 75.7% for mammography lesion, 64.4% for mammography margin, and 77.2% for lesion+margin. When all features were combined, the accuracy was improved to 89.6%. By adding mammography features to MRI, the specificity was significantly improved from 69.6% (39/56) to 82.1% (46/56), pConclusionThe radiomics model built from the combined MRI and mammography has the potential to provide a machine learning-based diagnostic tool and decrease the false positive diagnosis of contrast-enhanced benign lesions on MRI.

Glenohumeral joint (GHJ) space narrowing has been demonstrated to be an important morphologic parameter of glenohumeral osteoarthritis (GHO). However, the morphology of GHJ space is irregular because of degeneration of subchondral bone and articular cartilage. Thus, we devised GHJ cartilage cross-sectional area (GHJCCSA) as a new diagnostic morphological parameter to assess the irregular morphologic change of GHJ. GHJ samples were acquired from 33 patients with GHO and from 33 normal controls without evidence of GHO based on shoulder magnetic resonance imaging. T2-weighted coronal MRIs were collected at the GHJ level for all individuals. GHJCCSA and GHJ cartilage thickness (GHJCT) at the GHJ were measured on MRIs using a graphic measuring system. The GHJCCSA was measured as the whole cartilage cross-sectional area of the GHJ. The average GHJCCSA was 115.28 ± 17.36 mm2 in normal individuals and 61.77 ± 13.74 mm2 in the GHO group. The mean GHJCT was 2.06 ± 0.35 mm in normal individuals and 1.50 ± 0.28 mm in the GHO group. GHO patients had significantly lower GHJCCSA (P .001) and GHJCT (P .001) than normal individuals. Receiver operator characteristics curve analysis revealed that the optimal cutoff score of the GHJCCSA was 82.21 mm2, with a sensitivity of 97.0%, a specificity of 97.0%, and an area under the curve of 0.99 (95% CI: 0.97-1.00). Although GHJCCSA and GHJCT were both significantly associated with GHO, the GHJCCSA was a more sensitive measurement parameter.

A narrowed sacroiliac joint (SIJ) space has been considered to be a major morphologic parameter of ankylosing spondylitis (AS). Previous studies revealed that the sacroiliac joint thickness (SIJT) correlated with AS in patients. However, irregular narrowing is different from thickness. Thus, we devised a method using the sacroiliac joint cross-sectional area (SIJA) as a new morphological parameter for use in evaluating AS. We hypothesized that the SIJA is a key morphologic parameter in diagnosing AS. SIJ samples were collected from 107 patients with AS, and from 85 control subjects who underwent SIJ-view X-rays that revealed no evidence of AS. We measured the SIJT and SIJA at the SIJ margin on X-rays using our picture archiving and communications system. The SIJT was measured at the narrowest point between the sacrum and the ilium. The SIJA was measured as the entire cross-sectional joint space area of the SIJ in the X-ray images. The average SIJT was 3.09 ± 0.61 mm in the control group, and 1.59 ± 0.52 mm in the AS group. The average SIJA was 166.74 ± 39.98 mm2 in the control group, and 68.65 ± 24.11 mm2 in the AS group. AS patients had significantly lower SIJT (P .001) and SIJA (P .001) than the control subjects. Receiver operating characteristics curve analysis showed that the best cutoff point for the SIJT was 2.33 mm, with 92.5% sensitivity, 94.1% specificity, and an area under the curve of 0.97 (95% confidence interval: 0.95-0.99). The optimal cutoff point for the SIJA was 106.19 mm2, with 93.5% sensitivity, 95.3% specificity, and an area under the curve of 0.98 (95% confidence interval: 0.97-1.00). Although the SIJT and SIJA were both significantly associated with AS, the SIJA parameter was a more sensitive measurement. We concluded that the SIJA is an easy-to-use, fast, cheap, and useful new morphological parameter for predicting AS.

Background: To investigate the impact of intraventricular hemorrhage (IVH) on the classification of hematoma expansion (HE), and the development of radiomics models using features extracted from the baseline hematoma to predict HE.Methods: Eighty-four patients with baseline and follow-up non-contrast CT within 4~24 hours were included. The intraparenchymal hemorrhage (IPH) and IVH were separately outlined by an experienced neuroradiologist. HE was defined as an absolute hematoma growth >6 mL or percentage growth >33%. HE was determined based on two criteria, using IPH alone (HEP) or IPH+IVH (HEP+V). The radiomics analysis was performed by using PyRadiomics to extract features, followed by random forest algorithm to select features, and lastly the decision tree to build classification models. Results: The classification of expansion showed 37 (44%) HEP and 47 (56%) non-HEP based on IPH alone, and similar results of 38 (45%) HEP+V and 46 (55%) non-HEP+V based on IPH+IVH. The majority, >94% of HE patients, had a poor outcome (death or mRS>3 at discharge). Three radiomics analysis (RA) models were built. The first model using baseline IPH to predict HEP (RAP-P) showed an accuracy of 80% but loss of correlation with the clinical outcome; the second model using IPH+IVH to predict HEP (RAPV-V) had a slightly higher accuracy of 81% and resumed the poor outcome association with HE; and the third model using IPH+IVH to predict HEP+V (RAPV-PV) had the highest accuracy of 86% with preserved clinical outcome correlation of HE. The sensitivity, specificity, and accuracy of three decision trees (RAP-P, RAPV-P, RAPV-PV) were 0.8/ 0.68/ 0.89; 0.81/ 0.92/ 0.72 and 0.86/ 0.82/ 0.89, respectively.Conclusions: The proposed radiomics approach with additional IVH information could be used to classify the expansion status highly associated with the clinical outcome and provide a robust tool for the enrollment of high-risk ICH cases in the anti-expansion trials.

Objectives To explore the predictive value of preoperative imaging in patients with spinal giant cell tumor of bone (GCTB) for postoperative recurrence and risk stratification. Methods Clinical data for 62 cases of spinal GCTB diagnosed and treated at our hospital from 2008 to 2018 were identified. All patients were followed up for more than 2 years according to the clinical guidelines after surgery. Medical history data including baseline demographic and clinical characteristics, computed tomography (CT) and magnetic resonance imaging (MRI) findings of recurrent and non‐recurrent patients were compared. Two musculoskeletal radiologists read the images and were blinded to the clinical data. The imaging features associated with postoperative recurrence were analyzed by multivariate logistic regression, and receiver operating characteristic (ROC) curve analysis was performed to determine the optimal cutoff value of the largest lesion diameter predicting recurrence after surgery. Results According to whether the disease recurred within the follow‐up period, patients were divided into the recurrence group and the non‐recurrence group. Of 62 patients (29 males and 33 females), 17 had recurrence and 45 did not. The recurrence rate was 27.4%. The mean follow‐up time was 73.66 (± 32.92) months. The three major treatments were total en bloc spondylectomy (n = 26), intralesional spondylectomy (n = 20), and curettage(n = 16). A total of 16 CT and MRI features were analyzed. A univariate analysis showed no significant difference in age, sex, treatment, multi‐vertebral body involvement, location, boundary, expansile mass, residual bone crest, paravertebral soft tissue mass, CT value, and MRI signal on T1‐weighted imaging (WI), T2‐WI, and T2‐WI fat suppression (FS) sequences (P > 0.05). The largest lesion diameter [(4.68 ± 1.79) vs (5.92 ± 2.17) cm, t = 2.287, P = 0.026] and the vertebral compression fracture (51% vs 82%, χ 2 = 5.005, P = 0.025) were significantly different between the non‐recurrence and recurrence groups. Logistic regression analysis showed that both largest lesion diameter (odds ratio [OR], 1.584; 95% confidence interval [CI], 1.108–2.264; P = 0.012) and compression fracture (OR, 8.073; 95%CI, 1.481–11.003; P = 0.016) were independent predictors of postoperative recurrence. When we set the cutoff value for the largest lesion diameter at 4.2 cm, the sensitivity and specificity for distinguishing the recurrence and non‐recurrence of GCTB were 94.1% and 42.2%, respectively, and the area under the curve (AUC) was 0.671. The combined model achieved a sensitivity, specificity and accuracy of 47.1%, 97.8% and 83.9%, respectively. Conclusions In spinal GCTB, maximum lesion diameter and the vertebral compression fracture are associated with tumor recurrence after surgery, which may provide helpful information for planning personalized treatment., Giant cell tumor of bone (GCTB) is currently classified as an intermediate tumor. This current study of 62 patients found that the largest diameter of the tumor and the vertebral compression were independent predictors of tumor recurrence after surgery, which may provide more helpful information for development of a personalized treatment plan.

Carpal tunnel syndrome (CTS) is correlated with increased intracarpal canal pressure (ICP). The effect of palmaris longus tendon (PLT) loading on ICP is documented in previous researches. PLT loading induces the greatest absolute increase in ICP. Therefore, to analyze the connection between the PLT and CTS, we newly made the measurement of the PLT cross-sectional area (PLTCSA). We assumed that PLTCSA is a reliable diagnostic parameter in the CTS. PLTCSA measurement data were acquired from 21 patients with CTS, and from 21 normal subjects who underwent wrist magnetic resonance imaging (W-MRI). We measured the PLTCSA at the level of pisiform on W-MRI. The PLTCSA was measured on the outlining of PLT. The two different cutoff values in the analysis were determined using receiver operating characteristic (ROC) analysis. The mean PLTCSA was 2.34 ± 0.82 mm2 in the normal group and 3.97 ± 1.18 mm2 in the CTS group. ROC curve analysis concluded that the best cutoff point for the PLTCSA was 2.81 mm2, with 76.2% sensitivity, 71.4% specificity, and area under the curve of 0.88 (95% CI, 0.78-0.98). PLTCSA is a sensitive, new, objective morphological parameter for evaluating CTS.

Magnetic resonance imaging (MRI) plays an important role in the modern radiation therapy (RT) workflow. In comparison with computed tomography (CT) imaging, which is the dominant imaging modality in RT, MRI possesses excellent soft-tissue contrast for radiographic evaluation. Based on quantitative models, MRI can be used to assess tissue functional and physiological information. With the developments of scanner design, acquisition strategy, advanced data analysis, and modeling, multiparametric MRI (mpMRI), a combination of morphologic and functional imaging modalities, has been increasingly adopted for disease detection, localization, and characterization. Integration of mpMRI techniques into RT enriches the opportunities to individualize RT. In particular, RT response assessment using mpMRI allows for accurate characterization of both tissue anatomical and biochemical changes to support decision-making in monotherapy of radiation treatment and/or systematic cancer management. In recent years, accumulating evidence have, indeed, demonstrated the potentials of mpMRI in RT response assessment regarding patient stratification, trial benchmarking, early treatment intervention, and outcome modeling. Clinical application of mpMRI for treatment response assessment in routine radiation oncology workflow, however, is more complex than implementing an additional imaging protocol; mpMRI requires additional focus on optimal study design, practice standardization, and unified statistical reporting strategy to realize its full potential in the context of RT. In this article, the mpMRI theories, including image mechanism, protocol design, and data analysis, will be reviewed with a focus on the radiation oncology field. Representative works will be discussed to demonstrate how mpMRI can be used for RT response assessment. Additionally, issues and limits of current works, as well as challenges and potential future research directions, will also be discussed.

Purpose To predict the neoadjuvant chemoradiation therapy (CRT) response in patients with locally advanced rectal cancer (LARC) using radiomics and deep learning based on pre-treatment MRI and a mid-radiation follow-up MRI taken 3–4 weeks after the start of CRT. Methods A total of 51 patients were included, 45 with pre-treatment, 41 with mid-radiation therapy (RT), and 35 with both MRI sets. The multi-parametric MRI protocol included T2, diffusion weighted imaging (DWI) with b-values of 0 and 800 s/mm2, and dynamic-contrast-enhanced (DCE) MRI. After completing CRT and surgery, the specimen was examined to determine the pathological response based on the tumor regression grade. The tumor ROI was manually drawn on the post-contrast image and mapped to other sequences. The total tumor volume and mean apparent diffusion coefficient (ADC) were measured. Radiomics using GLCM texture and histogram parameters, and deep learning using a convolutional neural network (CNN), were performed to differentiate pathologic complete response (pCR) vs. non-pCR, and good response (GR) vs. non-GR. Results Tumor volume decreased and ADC increased significantly in the mid-RT MRI compared to the pre-treatment MRI. For predicting pCR vs. non-pCR, combining ROI and radiomics features achieved an AUC of 0.80 for pre-treatment, 0.82 for mid-RT, and 0.86 for both MRI together. For predicting GR vs. non-GR, the AUC was 0.91 for pre-treatment, 0.92 for mid-RT, and 0.93 for both MRI together. In deep learning using CNN, combining pre-treatment and mid-RT MRI achieved a higher accuracy compared to using either dataset alone, with AUC of 0.83 for predicting pCR vs. non-pCR. Conclusion Radiomics based on pre-treatment and early follow-up multi-parametric MRI in LARC patients receiving CRT could extract comprehensive quantitative information to predict final pathologic response.

Accurate segmentation of the breast tissue is a significant challenge in the analysis of breast MR images, especially analysis of breast images with low contrast. Most of the existing methods for breast segmentation are semi-automatic and limited in their ability to achieve accurate results. This is because of difficulties in removing landmarks from noisy magnetic resonance images (MRI). Especially, when tumour is imaged for scanning, how to isolate the tumour region from chest will directly affect the accuracy for tumour to be detected. Due to low intensity levels and the close connection between breast and chest portion in MRIs, this study proposes an innovative, fully automatic and fast segmentation approach which combines histogram with inverse Gaussian gradient for morphology snakes, along with extended Stein’s unbiased risk estimator (eSURE) applied for unsupervised learning of deep neural network Gaussian denoisers, aimed at accurate identification of landmarks such as chest and breast.

Rationale and Objectives Computer-aided methods have been widely applied to diagnose lesions on breast magnetic resonance imaging (MRI). The first step was to identify abnormal areas. A deep learning Mask Regional Convolutional Neural Network (R-CNN) was implemented to search the entire set of images and detect suspicious lesions. Materials and Methods Two DCE-MRI datasets were used, 241 patients acquired using non–fat-sat sequence for training, and 98 patients acquired using fat-sat sequence for testing. All patients have confirmed unilateral mass cancers. The tumor was segmented using fuzzy c-means clustering algorithm to serve as the ground truth. Mask R-CNN was implemented with ResNet-101 as the backbone. The neural network output the bounding boxes and the segmented tumor for evaluation using the Dice Similarity Coefficient (DSC). The detection performance, and the trade-off between sensitivity and specificity, was analyzed using free response receiver operating characteristic. Results When the precontrast and subtraction image of both breasts were used as input, the false positive from the heart and normal parenchymal enhancements could be minimized. The training set had 1469 positive slices (containing lesion) and 9135 negative slices. In 10-fold cross-validation, the mean accuracy = 0.86 and DSC = 0.82. The testing dataset had 1568 positive and 7264 negative slices, with accuracy = 0.75 and DSC = 0.79. When the obtained per-slice results were combined, 240 of 241 (99.5%) lesions in the training and 98 of 98 (100%) lesions in the testing datasets were identified. Conclusion Deep learning using Mask R-CNN provided a feasible method to search breast MRI, localize, and segment lesions. This may be integrated with other artificial intelligence algorithms to develop a fully automatic breast MRI diagnostic system.

Highlights • Characteristics of 62 patients with spinal GCTB who underwent surgery. • A prognostic classification model was built based on features selected by SVM. • The combined histogram and texture features could predict recurrence of GCTB., Objectives To determine if radiomics analysis based on preoperative computed tomography (CT) can predict early postoperative recurrence of giant cell tumor of bone (GCTB) in the spine. Methods In a retrospective review, 62 patients with pathologically confirmed spinal GCTB from March 2008 to February 2018, with a minimum follow-up of 24 months, were identified. The mean follow-up was 73.7 months (range, 28.7–152.1 months). The clinical information including age, gender, lesion location, multi-vertebral involvement, and surgical methods, were obtained. CT images acquired before the operation were retrieved for radiomics analysis. For each case, the tumor regions of interest (ROI) was manually outlined, and a total of 107 radiomics features were extracted. The features were selected via the sequential selection process by using the support vector machine (SVM), then used to construct classification models with Gaussian kernels. The differentiation between recurrence and non-recurrence groups was evaluated by ROC analysis, using 10-fold cross-validation. Results Of the 62 patients, 17 had recurrence with a recurrence rate of 27.4%. None of the clinical information was significantly different between the two groups. Patients receiving curettage had a higher recurrence rate (6/16 = 37.5%) compared to patients receiving TES (6/26 = 23.1%) or intralesional spondylectomy (5/20 = 25%). The final radiomics model was built using 10 selected features, which achieved an accuracy of 89% with AUC of 0.78. Conclusions The radiomics model developed based on pre-operative CT can achieve a high accuracy to predict the recurrence of spinal GCTB. Patients who have a high risk of early recurrence should be treated more aggressively to minimize recurrence.

To apply deep learning algorithms using a conventional convolutional neural network (CNN) and a recurrent CNN to differentiate three breast cancer molecular subtypes on MRI. A total of 244 patients were analyzed, 99 in training dataset scanned at 1.5 T and 83 in testing-1 and 62 in testing-2 scanned at 3 T. Patients were classified into 3 subtypes based on hormonal receptor (HR) and HER2 receptor: (HR+/HER2−), HER2+, and triple negative (TN). Only images acquired in the DCE sequence were used in the analysis. The smallest bounding box covering tumor ROI was used as the input for deep learning to develop the model in the training dataset, by using a conventional CNN and the convolutional long short-term memory (CLSTM). Then, transfer learning was applied to re-tune the model using testing-1(2) and evaluated in testing-2(1). In the training dataset, the mean accuracy evaluated using tenfold cross-validation was higher by using CLSTM (0.91) than by using CNN (0.79). When the developed model was applied to the independent testing datasets, the accuracy was 0.4–0.5. With transfer learning by re-tuning parameters in testing-1, the mean accuracy reached 0.91 by CNN and 0.83 by CLSTM, and improved accuracy in testing-2 from 0.47 to 0.78 by CNN and from 0.39 to 0.74 by CLSTM. Overall, transfer learning could improve the classification accuracy by greater than 30%. The recurrent network using CLSTM could track changes in signal intensity during DCE acquisition, and achieved a higher accuracy compared with conventional CNN during training. For datasets acquired using different settings, transfer learning can be applied to re-tune the model and improve accuracy. • Deep learning can be applied to differentiate breast cancer molecular subtypes. • The recurrent neural network using CLSTM could track the change of signal intensity in DCE images, and achieved a higher accuracy compared with conventional CNN during training. • For datasets acquired using different scanners with different imaging protocols, transfer learning provided an efficient method to re-tune the classification model and improve accuracy.

Background Computer-aided methods have been widely applied to diagnose lesions detected on breast MRI, but fully-automatic diagnosis using deep learning is rarely reported. Purpose To evaluate the diagnostic accuracy of mass lesions using region of interest (ROI)-based, radiomics and deep-learning methods, by taking peritumor tissues into consideration. Study type Retrospective. Population In all, 133 patients with histologically confirmed 91 malignant and 62 benign mass lesions for training (74 patients with 48 malignant and 26 benign lesions for testing). Field strength/sequence 3T, using the volume imaging for breast assessment (VIBRANT) dynamic contrast-enhanced (DCE) sequence. Assessment 3D tumor segmentation was done automatically by using fuzzy-C-means algorithm with connected-component labeling. A total of 99 texture and histogram parameters were calculated for each case, and 15 were selected using random forest to build a radiomics model. Deep learning was implemented using ResNet50, evaluated with 10-fold crossvalidation. The tumor alone, smallest bounding box, and 1.2, 1.5, 2.0 times enlarged boxes were used as inputs. Statistical tests The malignancy probability was calculated using each model, and the threshold of 0.5 was used to make a diagnosis. Results In the training dataset, the diagnostic accuracy was 76% using three ROI-based parameters, 84% using the radiomics model, and 86% using ROI + radiomics model. In deep learning using the per-slice basis, the area under the receiver operating characteristic (ROC) was comparable for tumor alone, smallest and 1.2 times box (AUC = 0.97-0.99), which were significantly higher than 1.5 and 2.0 times box (AUC = 0.86 and 0.71, respectively). For per-lesion diagnosis, the highest accuracy of 91% was achieved when using the smallest bounding box, and that decreased to 84% for tumor alone and 1.2 times box, and further to 73% for 1.5 times box and 69% for 2.0 times box. In the independent testing dataset, the per-lesion diagnostic accuracy was also the highest when using the smallest bounding box, 89%. Data conclusion Deep learning using ResNet50 achieved a high diagnostic accuracy. Using the smallest bounding box containing proximal peritumor tissue as input had higher accuracy compared to using tumor alone or larger boxes. Level of evidence 3 Technical Efficacy: Stage 2.

Rationale and Objectives Breast cancer occurs more frequently in the upper outer (UO) quadrant, but whether this higher cancer incidence is related to the greater amount of dense tissue is not known. Magnetic resonance imaging acquires three-dimensional volumetric images and is the most suitable among all breast imaging modalities for regional quantification of density. This study applied a magnetic resonance imaging-based method to measure quadrant percent density (QPD), and evaluated its association with the quadrant location of the developed breast cancer. Materials and Methods A total of 126 cases with pathologically confirmed breast cancer were reviewed. Only women who had unilateral breast cancer located in a clear quadrant were selected for analysis. A total of 84 women, including 47 Asian women and 37 western women, were included. An established computer-aided method was used to segment the diseased breast and the contralateral normal breast, and to separate the dense and fatty tissues. Then, a breast was further separated into four quadrants using the nipple and the centroid as anatomic landmarks. The tumor was segmented using a computer-aided method to determine its quadrant location. The distribution of cancer quadrant location, the quadrant with the highest QPD, and the proportion of cancers occurring in the highest QPD were analyzed. Results The highest incidence of cancer occurred in the UO quadrant (36 out of 84, 42.9%). The highest QPD was also noted most frequently in the UO quadrant (31 out of 84, 36.9%). When correlating the highest QPD with the quadrant location of breast cancer, only 17 women out of 84 (20.2%) had breast cancer occurring in the quadrant with the highest QPD. Conclusions The results showed that the development of breast cancer in a specific quadrant could not be explained by the density in that quadrant, and further studies are needed to find the biological reasons accounting for the higher breast cancer incidence in the UO quadrant.

Rationale and Objectives This study aimed to evaluate the diagnostic performance of dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) in differentiation of four spinal lesions by using heuristic and pharmacokinetic parameters analyzed from DCE signal intensity time course. Materials and Methods DCE-MRI of 62 subjects with confirmed myeloma (n = 9), metastatic cancer (n = 22), lymphoma (n = 7), and inflammatory tuberculosis (TB) (n = 24) in the spine were analyzed retrospectively. The region of interest was placed on strongly enhanced tissues. The DCE time course was categorized as the “wash-out,” “plateau,” or “persistent enhancement” pattern. The maximum enhancement, steepest wash-in enhancement, and wash-out slope using the signal intensity at 67 seconds after contrast injection as reference were measured. The Tofts 2-compartmental pharmacokinetic model was applied to obtain Ktrans and kep. Pearson correlation between heuristic and pharmacokinetic parameters was evaluated, and receiver operating characteristic curve analysis was performed for pairwise group differentiation. Results The mean wash-out slope was −22% ± 10% for myeloma, 1% ± 0.4% for metastatic cancer, 3% ± 3% for lymphoma, and 7% ± 10% for TB, and it could significantly distinguish myeloma from metastasis (area under the curve [AUC] = 0.884), lymphoma (AUC = 1.0), and TB (AUC = 1.0) with P = .001, and distinguish metastasis from TB (AUC = 0.741) with P = .005. The kep and wash-out slope were highly correlated (r = 0.92), and they showed a similar diagnostic performance. The Ktrans was significantly correlated with the maximum enhancement (r = 0.71) and the steepest wash-in enhancement (r = 0.85), but they had inferior diagnostic performance compared to the wash-out slope. Conclusions DCE-MRI may provide additional diagnostic information, and a simple wash-out slope had the best diagnostic performance. The heuristic and pharmacokinetic parameters were highly correlated.

Background We describe a novel strategy for power and sample size determination developed for studies utilizing investigational technologies with limited available preliminary data, specifically of imaging biomarkers. We evaluated diffuse optical spectroscopic imaging (DOSI), an experimental noninvasive imaging technique that may be capable of assessing changes in mammographic density. Because there is significant evidence that tamoxifen treatment is more effective at reducing breast cancer risk when accompanied by a reduction of breast density, we designed a study to assess the changes from baseline in DOSI imaging biomarkers that may reflect fluctuations in breast density in premenopausal women receiving tamoxifen. Method While preliminary data demonstrate that DOSI is sensitive to mammographic density in women about to receive neoadjuvant chemotherapy for breast cancer, there is no information on DOSI in tamoxifen treatment. Since the relationship between magnetic resonance imaging (MRI) and DOSI has been established in previous studies, we developed a statistical simulation approach utilizing information from an investigation of MRI assessment of breast density in 16 women before and after treatment with tamoxifen to estimate the changes in DOSI biomarkers due to tamoxifen. Results Three sets of 10,000 pairs of MRI breast density data with correlation coefficients of 0.5, 0.8 and 0.9 were simulated and generated and were used to simulate and generate a corresponding 5,000,000 pairs of DOSI values representing water, ctHHB, and lipid. Minimum sample sizes needed per group for specified clinically-relevant effect sizes were obtained. Conclusion The simulation techniques we describe can be applied in studies of other experimental technologies to obtain the important preliminary data to inform the power and sample size calculations. Electronic supplementary material The online version of this article (doi:10.1186/s12874-017-0329-1) contains supplementary material, which is available to authorized users.