Peritumoral MRI Radiomics Features Increase the Evaluation Efficiency for Response to Chemotherapy in Patients With Epithelial Ovarian Cancer.

BACKGROUND: It remains unclear whether extracting peritumoral volume (PTV) radiomics features are useful tools for evaluating response to chemotherapy of epithelial ovarian cancer (EOC). PURPOSE: To evaluate MRI radiomics signatures (RS) capturing subtle changes of PTV and their added evaluation performance to whole tumor volume (WTV) for response to chemotherapy in patients with EOC. STUDY TYPE: Retrospective. POPULATION: 219 patients aged from 15 to 79 years were enrolled. FIELD STRENGTH/SEQUENCE: 3.0 or 1.5T, axial fat-suppressed T2-weighted imaging (FS-T2WI), diffusion-weighted imaging (DWI), and contrast enhanced T1-weighted imaging (CE-T1WI). ASSESSMENT: MRI features were extracted from the four axial sequences and six different volumes of interest (VOIs) (WTV and WTV + PTV (WPTV)) with different peritumor sizes (PS) ranging from 1 to 5 mm. Those features underwent preprocessing, and the most informative features were selected using minimum redundancy maximum relevance and least absolute shrinkage and selection operator to construct the RS. The optimal RS, with the highest area under the curve (AUC) of receiver operating characteristic was then integrated with independent clinical characteristics through multivariable logistic regression to construct the radiomics-clinical model (RCM). STATISTICAL TESTS: Mann-Whitney U test, chi-squared test, DeLong test, log-rank test. P < 0.05 indicated a significant difference. RESULTS: All the RSs constructed on WPTV exhibited higher AUCs (0.720-0.756) than WTV (0.671). Of which, RS with PS = 2 mm displayed a significantly better performance (AUC = 0.756). International Federation of Gynecology and Obstetrics (FIGO) stage was identified as the exclusive independent clinical evaluation characteristic, and the RCM demonstrated higher AUC (0.790) than the RS, but without statistical significance (P = 0.261). DATA CONCLUSION: The radiomics features extracted from PTV could increase the efficiency of WTV radiomics for evaluating the chemotherapy response of EOC. The cut-off of 2 mm PTV was a reasonable value to obtain effective evaluation efficiency. LEVEL OF EVIDENCE: 4 TECHNICAL EFFICACY: Stage 2.

Multicenter Evaluation of a Weakly Supervised Deep Learning Model for Lymph Node Diagnosis in Rectal Cancer at MRI.

Purpose To develop a Weakly supervISed model DevelOpment fraMework (WISDOM) model to construct a lymph node (LN) diagnosis model for patients with rectal cancer (RC) that uses preoperative MRI data coupled with postoperative patient-level pathologic information. Materials and Methods In this retrospective study, the WISDOM model was built using MRI (T2-weighted and diffusion-weighted imaging) and patient-level pathologic information (the number of postoperatively confirmed metastatic LNs and resected LNs) based on the data of patients with RC between January 2016 and November 2017. The incremental value of the model in assisting radiologists was investigated. The performances in binary and ternary N staging were evaluated using area under the receiver operating characteristic curve (AUC) and the concordance index (C index), respectively. Results A total of 1014 patients (median age, 62 years; IQR, 54-68 years; 590 male) were analyzed, including the training cohort (n = 589) and internal test cohort (n = 146) from center 1 and two external test cohorts (cohort 1: 117; cohort 2: 162) from centers 2 and 3. The WISDOM model yielded an overall AUC of 0.81 and C index of 0.765, significantly outperforming junior radiologists (AUC = 0.69, P < .001; C index = 0.689, P < .001) and performing comparably with senior radiologists (AUC = 0.79, P = .21; C index = 0.788, P = .22). Moreover, the model significantly improved the performance of junior radiologists (AUC = 0.80, P < .001; C index = 0.798, P < .001) and senior radiologists (AUC = 0.88, P < .001; C index = 0.869, P < .001). Conclusion This study demonstrates the potential of WISDOM as a useful LN diagnosis method using routine rectal MRI data. The improved radiologist performance observed with model assistance highlights the potential clinical utility of WISDOM in practice. Keywords: MR Imaging, Abdomen/GI, Rectum, Computer Applications-Detection/Diagnosis Supplemental material is available for this article. Published under a CC BY 4.0 license.

Deep Learning-Based Fully Automated Segmentation of Regional Muscle Volume and Spatial Intermuscular Fat Using CT.

RATIONALE AND OBJECTIVES: We aim to develop a CT-based deep learning (DL) system for fully automatic segmentation of regional muscle volume and measurement of the spatial intermuscular fat distribution of the gluteus maximus muscle. MATERIALS AND METHODS: A total of 472 subjects were enrolled and randomly assigned to one of three groups: a training set, test set 1, and test set 2. For each subject in the training set and test set 1, we selected six slices of the CT images as the region of interest for manual segmentation by a radiologist. For each subject in test set 2, we selected all slices of the gluteus maximus muscle on the CT images for manual segmentation. The DL system was constructed using Attention U-Net and the Otsu binary thresholding method to segment the muscle and measure the fat fraction of the gluteus maximus muscle. The segmentation results of the DL system were evaluated using the Dice similarity coefficient (DSC), Hausdorff distance (HD), and the average surface distance (ASD) as metrics. Intraclass correlation coefficients (ICCs) and Bland-Altman plots were used to assess agreement in the measurements of fat fraction between the radiologist and the DL system. RESULTS: The DL system showed good segmentation performance on the two test sets, with DSCs of 0.930 and 0.873, respectively. The fat fraction of the gluteus maximus muscle measured by the DL system was in agreement with the radiologist (ICC=0.748). CONCLUSION: The proposed DL system showed accurate, fully automated segmentation performance and good agreement with the radiologist at fat fraction evaluation, and can be further used for muscle evaluation.

Contribution of whole slide imaging-based deep learning in the assessment of intraoperative and postoperative sections in neuropathology.

The pathological diagnosis of intracranial germinoma (IG), oligodendroglioma, and low-grade astrocytoma on intraoperative frozen section (IFS) and hematoxylin-eosin (HE)-staining section directly determines patients' treatment options, but it is a difficult task for pathologists. We aimed to investigate whether whole-slide imaging (WSI)-based deep learning can contribute new precision to the diagnosis of IG, oligodendroglioma, and low-grade astrocytoma. Two types of WSIs (500 IFSs and 832 HE-staining sections) were collected from 379 patients at multiple medical centers. Patients at Center 1 were split into the training, testing, and internal validation sets (3:1:1), while the other centers were the external validation sets. First, we subdivided WSIs into small tiles and selected tissue tiles using a tissue tile selection model. Then a tile-level classification model was established, and the majority voting method was used to determine the final diagnoses. Color jitter was applied to the tiles so that the deep learning (DL) models could adapt to the variations in the staining. Last, we investigated the effectiveness of model assistance. The internal validation accuracies of the IFS and HE models were 93.9% and 95.3%, respectively. The external validation accuracies of the IFS and HE models were 82.0% and 76.9%, respectively. Furthermore, the IFS and HE models can predict Ki-67 positive cell areas with R2 of 0.81 and 0.86, respectively. With model assistance, the IFS and HE diagnosis accuracy of pathologists improved from 54.6%-69.7% and 53.5%-83.7% to 87.9%-93.9% and 86.0%-90.7%, respectively. Both the IFS model and the HE model can differentiate the three tumors, predict the expression of Ki-67, and improve the diagnostic accuracy of pathologists. The use of our model can assist clinicians in providing patients with optimal and timely treatment options.

Deep learning-based segmentation of epithelial ovarian cancer on T2-weighted magnetic resonance images.

Background: Epithelial ovarian cancer (EOC) segmentation is an indispensable step in assessing the extent of disease and guiding the treatment plan that follows. Currently, manual segmentation is the most commonly used method, despite it being tedious, time-consuming and subject to inter- and intra-observer variability. This study aims to assess the feasibility of deep learning methods in the automatic segmentation of EOC on T2-weighted magnetic resonance images. Methods: A total of 339 EOC patients from eight different clinical centers were enrolled and divided into 4 groups: training set (n=154), validation set (n=25), internal test set (n=50) and external test set (n=110). Six common-used algorithms, including convolutional neural networks (CNNs) (U-Net, DeepLabv3, U-Net++ and PSPNet) and transformers (TransUnet and Swin-Unet), were used to conduct automatic segmentations. The performances of these automatic segmentation methods were evaluated by means of dice similarity coefficient (DSC), Hausdorff distance (HD), average symmetric surface distance (ASSD), precision and recall. Results: All the results look promising, which demonstrates the feasibility of using deep learning for EOC segmentation. Overall, CNNs and transformers showed similar performances in both internal and external test sets. Among all the models, U-Net++ performed best with a DSC, HD, ASSD, precision and recall of 0.851, 25.3, 1.75, 0.838, 0.882 and 0.740, 42.5, 4.21, 0.825, 0.725 in internal and external test sets, respectively. Conclusions: Fully automated segmentation of EOC is possible with deep learning. The segmentation performance is related to the International Federation of Gynecology and Obstetrics (FIGO) stages and histological types of EOC.

Preoperative CECT-based Radiomic Signature for Predicting the Response of Transarterial Chemoembolization (TACE) Therapy in Hepatocellular Carcinoma.

PURPOSE: To evaluate the efficiency of radiomics signatures in predicting the response of transarterial chemoembolization (TACE) therapy based on preoperative contrast-enhanced computed tomography (CECT). MATERIALS: This study consisted of 111 patients with intermediate-stage hepatocellular carcinoma who underwent CECT at both the arterial phase (AP) and venous phase (VP) before and after TACE. According to mRECIST 1.1, patients were divided into an objective-response group (n = 38) and a non-response group (n = 73). Among them, 79 patients were assigned as the training dataset, and the remaining 32 cases were assigned as the test dataset. METHODS: Radiomics features were extracted from CECT images. Two feature ranking methods and three classifiers were used to find the best single-phase radiomics signatures for both AP and VP on the training set. Meanwhile, multi-phase radiomics signatures were built upon integration of images from two CECT phases by decision-level fusion and feature-level fusion. Finally, multivariable logistic regression was used to develop a nomogram by combining radiomics signatures and clinic-radiologic characteristics. The prediction performance was evaluated by AUC on the test dataset. RESULTS: The multi-phase radiomics signature (AUC = 0.883) performed better in predicting TACE therapy response compared to the best single-phase radiomics signature (AUC = 0.861). The nomogram (AUC = 0.913) showed better performance than any radiomics signatures. CONCLUSION: The radiomics signatures and nomogram were developed and validated for predicting responses to TACE therapy, and the radiomics model may play a positive role in identifying patients who may benefit from TACE therapy in clinical practice.

Deep learning-based fully automated diagnosis of melanocytic lesions by using whole slide images.

BACKGROUND: Erroneous diagnoses of melanocytic lesions (benign, atypical, and malignant types) result in inappropriate surgical treatment plans. OBJECTIVE: To propose a deep learning (DL)-based fully automated diagnostic method using whole slide images (WSIs) for melanocytic lesions. METHODS: The method consisted of patch prediction using a DL model and patient diagnosis using an aggregation module. The method was developed with 745 WSIs and evaluated using internal and external testing sets comprising 182 WSIs and 54 WSIs, respectively. The results were compared with those of the classification by one junior and two senior pathologists. Furthermore, we compared the performance of the three pathologists in the classification of melanocytic lesions with and without the assistance of our method. RESULTS: The method achieved an accuracy of 0.963 and 0.930 on the internal and external testing sets, respectively, which was significantly higher than that of the junior pathologist (0.419 and 0.535). With assistance from the method, all three pathologists achieved higher accuracy on the internal and external testing sets; the accuracy of the junior pathologist increased by 39.0% and 30.2%, respectively (p < .05). CONCLUSION: This generalizable method can accurately classify melanocytic lesions and effectively improve the diagnostic accuracy of pathologists.

CT-based radiomics signatures can predict the tumor response of non-small cell lung cancer patients treated with first-line chemotherapy and targeted therapy.

OBJECTIVES: The goal of this study was to evaluate the effectiveness of radiomics signatures on pre-treatment computed tomography (CT) images of lungs to predict the tumor responses of non-small cell lung cancer (NSCLC) patients treated with first-line chemotherapy, targeted therapy, or a combination of both. MATERIALS AND METHODS: This retrospective study included 322 NSCLC patients who were treated with first-line chemotherapy, targeted therapy, or a combination of both. Of these patients, 224 were randomly assigned to a cohort to help develop the radiomics signature. A total of 1946 radiomics features were obtained from each patient's CT scan. The top-ranked features were selected by the Minimum Redundancy Maximum Relevance (MRMR) feature-ranking method and used to build a lightweight radiomics signature with the Random Forest (RF) classifier. The independent predictive (IP) features (AUC > 0.6, p value < 0.05) were further identified from the top-ranked features and used to build a refined radiomics signature by the RF classifier. Its prediction performance was tested on the validation cohort, which consisted of the remaining 98 patients. RESULTS: The initial lightweight radiomics signature constructed from 15 top-ranked features had an AUC of 0.721 (95% CI, 0.619-0.823). After six IP features were further identified and a refined radiomics signature was built, it had an AUC of 0.746 (95% CI, 0.646-0.846). CONCLUSIONS: Radiomics signatures based on pre-treatment CT scans can accurately predict tumor response in NSCLC patients after first-line chemotherapy or targeted therapy treatments. Radiomics features could be used as promising prognostic imaging biomarkers in the future. KEY POINTS: The radiomics signature extracted from baseline CT images in patients with NSCLC can predict response to first-line chemotherapy, targeted therapy, or both treatments with an AUC = 0.746 (95% CI, 0.646-0.846). The radiomics signature could be used as a new biomarker for quantitative analysis in radiology, which might provide value in decision-making and to define personalized treatments for cancer patients.

MRI-Based Multiple Instance Convolutional Neural Network for Increased Accuracy in the Differentiation of Borderline and Malignant Epithelial Ovarian Tumors.

BACKGROUND: Preoperative differentiation of borderline from malignant epithelial ovarian tumors (BEOT vs. MEOT) is challenging and can significantly impact surgical management. PURPOSE: To develop a multiple instance convolutional neural network (MICNN) that can differentiate BEOT from MEOT, and to compare its diagnostic performance with that of radiologists. STUDY TYPE: Retrospective study of eight clinical centers. SUBJECTS: Between January 2010 and June 2018, a total of 501 women (mean age, 48.93 ± 14.05 years) with histopathologically confirmed BEOT (N = 165) or MEOT (N = 336) were divided into the training (N = 342) and validation cohorts (N = 159). FIELD STRENGTH/SEQUENCE: Three axial sequences from 1.5 or 3 T scanner were used: fast spin echo T2-weighted imaging with fat saturation (T2WI FS), echo planar diffusion-weighted imaging, and 2D volumetric interpolated breath-hold examination of contrast-enhanced T1-weighted imaging (CE-T1WI) with FS. ASSESSMENT: Three monoparametric MICNN models were built based on T2WI FS, apparent diffusion coefficient map, and CE-T1WI. Based on these monoparametric models, we constructed an early multiparametric (EMP) model and a late multiparametric (LMP) model using early and late information fusion methods, respectively. The diagnostic performance of the models was evaluated using the receiver operating characteristic (ROC) curve and compared to the performance of six radiologists with varying levels of experience. STATISTICAL TESTS: We used DeLong test, chi-square test, Mann-Whitney U-test, and t-test, with significance level of 0.05. RESULTS: Both EMP and LMP models differentiated BEOT from MEOT, with an area under the ROC curve (AUC) of 0.855 (95% CI, 0.795-0.915) and 0.884 (95% CI, 0.831-0.938), respectively. The AUC of the LMP model was significantly higher than the radiologists' pooled AUC (0.884 vs. 0.797). DATA CONCLUSION: The developed MICNN models can effectively differentiate BEOT from MEOT and the diagnostic performances (AUCs) were more superior than that of the radiologists' assessments. LEVEL OF EVIDENCE: 3 TECHNICAL EFFICACY STAGE: 2.

Multiple instance convolutional neural network with modality-based attention and contextual multi-instance learning pooling layer for effective differentiation between borderline and malignant epithelial ovarian tumors.

Malignant epithelial ovarian tumors (MEOTs) are the most lethal gynecologic malignancies, accounting for 90% of ovarian cancer cases. By contrast, borderline epithelial ovarian tumors (BEOTs) have low malignant potential and are generally associated with a good prognosis. Accurate preoperative differentiation between BEOTs and MEOTs is crucial for determining the appropriate surgical strategies and improving the postoperative quality of life. Multimodal magnetic resonance imaging (MRI) is an essential diagnostic tool. Although state-of-the-art artificial intelligence technologies such as convolutional neural networks can be used for automated diagnoses, their application have been limited owing to their high demand for graphics processing unit memory and hardware resources when dealing with large 3D volumetric data. In this study, we used multimodal MRI with a multiple instance learning (MIL) method to differentiate between BEOT and MEOT. We proposed the use of MAC-Net, a multiple instance convolutional neural network (MICNN) with modality-based attention (MA) and contextual MIL pooling layer (C-MPL). The MA module can learn from the decision-making patterns of clinicians to automatically perceive the importance of different MRI modalities and achieve multimodal MRI feature fusion based on their importance. The C-MPL module uses strong prior knowledge of tumor distribution as an important reference and assesses contextual information between adjacent images, thus achieving a more accurate prediction. The performance of MAC-Net is superior, with an area under the receiver operating characteristic curve of 0.878, surpassing that of several known MICNN approaches. Therefore, it can be used to assist clinical differentiation between BEOTs and MEOTs.

3D multi-scale, multi-task, and multi-label deep learning for prediction of lymph node metastasis in T1 lung adenocarcinoma patients' CT images.

The diagnosis of preoperative lymph node (LN) metastasis is crucial to evaluate possible therapy options for T1 lung adenocarcinoma patients. Radiologists preoperatively diagnose LN metastasis by evaluating signs related to LN metastasis, like spiculation or lobulation of pulmonary nodules in CT images. However, this type of evaluation is subjective and time-consuming, which may result in poor consistency and low efficiency of diagnoses. In this study, a 3D Multi-scale, Multi-task, and Multi-label classification network (3M-CN) was proposed to predict LN metastasis, as well as evaluate multiple related signs of pulmonary nodules in order to improve the accuracy of LN metastasis prediction. The following key approaches were adapted for this method. First, a multi-scale feature fusion module was proposed to aggregate the features from different levels for which different labels be best modeled at different levels; second, an auxiliary segmentation task was applied to force the model to focus more on the nodule region and less on surrounding unrelated structures; and third, a cross-modal integration module called the refine layer was designed to integrate the related risk factors into the model to further improve its confidence level. The 3M-CN was trained using data from 401 cases and then validated on both internal and external datasets, which consisted of 100 cases and 53 cases, respectively. The proposed 3M-CN model was then compared with existing state-of-the-art methods for prediction of LN metastasis. The proposed model outperformed other methods, achieving the best performance with AUCs of 0.945 and 0.948 in the internal and external test datasets, respectively. The proposed model not only obtain strong generalization, but greatly enhance the interpretability of the deep learning model, increase doctors' confidence in the model results, conform to doctors' diagnostic process, and may also be transferable to the diagnosis of other diseases.

MR image-based radiomics to differentiate type Ι and type ΙΙ epithelial ovarian cancers.

OBJECTIVES: Epithelial ovarian cancers (EOC) can be divided into type I and type II according to etiology and prognosis. Accurate subtype differentiation can substantially impact patient management. In this study, we aimed to construct an MR image-based radiomics model to differentiate between type I and type II EOC. METHODS: In this multicenter retrospective study, a total of 294 EOC patients from January 2010 to February 2019 were enrolled. Quantitative MR imaging features were extracted from the following axial sequences: T2WI FS, DWI, ADC, and CE-T1WI. A combined model was constructed based on the combination of these four MR sequences. The diagnostic performance was evaluated by ROC-AUC. In addition, an occlusion test was carried out to identify the most critical region for EOC differentiation. RESULTS: The combined radiomics model exhibited superior diagnostic capability over all four single-parametric radiomics models, both in internal and external validation cohorts (AUC of 0.806 and 0.847, respectively). The occlusion test revealed that the most critical region for differential diagnosis was the border zone between the solid and cystic components, or the less compact areas of solid component on direct visual inspection. CONCLUSIONS: MR image-based radiomics modeling can differentiate between type I and type II EOC and identify the most critical region for differential diagnosis. KEY POINTS: • Combined radiomics models exhibited superior diagnostic capability over all four single-parametric radiomics models, both in internal and external validation cohorts (AUC of 0.834 and 0.847, respectively). • The occlusion test revealed that the most crucial region for differentiating type Ι and type ΙΙ EOC was the border zone between the solid and cystic components, or the less compact areas of solid component on direct visual inspection on T2WI FS. • The light-combined model (constructed by T2WI FS, DWI, and ADC sequences) can be used for patients who are not suitable for contrast agent use.

Deep Learning Predicts Overall Survival of Patients With Unresectable Hepatocellular Carcinoma Treated by Transarterial Chemoembolization Plus Sorafenib.

OBJECTIVES: To develop and validate a deep learning-based overall survival (OS) prediction model in patients with hepatocellular carcinoma (HCC) treated with transarterial chemoembolization (TACE) plus sorafenib. METHODS: This retrospective multicenter study consisted of 201 patients with treatment-naïve, unresectable HCC who were treated with TACE plus sorafenib. Data from 120 patients were used as the training set for model development. A deep learning signature was constructed using the deep image features from preoperative contrast-enhanced computed tomography images. An integrated nomogram was built using Cox regression by combining the deep learning signature and clinical features. The deep learning signature and nomograms were also externally validated in an independent validation set of 81 patients. C-index was used to evaluate the performance of OS prediction. RESULTS: The median OS of the entire set was 19.2 months and no significant difference was found between the training and validation cohort (18.6 months vs. 19.5 months, P = 0.45). The deep learning signature achieved good prediction performance with a C-index of 0.717 in the training set and 0.714 in the validation set. The integrated nomogram showed significantly better prediction performance than the clinical nomogram in the training set (0.739 vs. 0.664, P = 0.002) and validation set (0.730 vs. 0.679, P = 0.023). CONCLUSION: The deep learning signature provided significant added value to clinical features in the development of an integrated nomogram which may act as a potential tool for individual prognosis prediction and identifying HCC patients who may benefit from the combination therapy of TACE plus sorafenib.

Deep learning-based fully automated detection and segmentation of lymph nodes on multiparametric-mri for rectal cancer: A multicentre study.

BACKGROUND: Accurate lymph nodes (LNs) assessment is important for rectal cancer (RC) staging in multiparametric magnetic resonance imaging (mpMRI). However, it is incredibly time-consumming to identify all the LNs in scan region. This study aims to develop and validate a deep-learning-based, fully-automated lymph node detection and segmentation (auto-LNDS) model based on mpMRI. METHODS: In total, 5789 annotated LNs (diameter ≥ 3 mm) in mpMRI from 293 patients with RC in a single center were enrolled. Fused T2-weighted images (T2WI) and diffusion-weighted images (DWI) provided input for the deep learning framework Mask R-CNN through transfer learning to generate the auto-LNDS model. The model was then validated both on the internal and external datasets consisting of 935 LNs and 1198 LNs, respectively. The performance for LNs detection was evaluated using sensitivity, positive predictive value (PPV), and false positive rate per case (FP/vol), and segmentation performance was evaluated using the Dice similarity coefficient (DSC). FINDINGS: For LNs detection, auto-LNDS achieved sensitivity, PPV, and FP/vol of 80.0%, 73.5% and 8.6 in internal testing, and 62.6%, 64.5%, and 8.2 in external testing, respectively, significantly better than the performance of junior radiologists. The time taken for model detection and segmentation was 1.3 s/case, compared with 200 s/case for the radiologists. For LNs segmentation, the DSC of the model was in the range of 0.81-0.82. INTERPRETATION: This deep learning-based auto-LNDS model can achieve pelvic LNseffectively based on mpMRI for RC, and holds great potential for facilitating N-staging in clinical practice.

A cross-modal 3D deep learning for accurate lymph node metastasis prediction in clinical stage T1 lung adenocarcinoma.

OBJECTIVES: The evaluation of lymph node (LN) status by radiologists based on preoperative computed tomography (CT) lacks high precision for early lung cancer patients; erroneous evaluations result in inappropriate therapeutic plans and increase the risk of complications. This study aims to develop a cross-modal 3D neural network based on CT images and prior clinical knowledge for accurate prediction of LN metastasis in clinical stage T1 lung adenocarcinoma. PATIENTS AND METHODS: Five hundred one lung adenocarcinoma patients with clinical stage T1 were enrolled. Data including: corresponding 3D nodule-centered patches of CT; prior clinical features; and pathological labels of LN status were obtained. We proposed a cross-modal deep learning system, which can successfully incorporate prior clinical knowledge and CT images into a 3D neural network to predict LN metastasis. We trained and validated our system with 401 cases and tested its performance with 100 cases. The result was compared with that of the logistic regression integration model, the single deep learning model without prior clinical knowledge integration, radiomics method, and manual evaluation by radiologists. RESULTS: The model proposed DensePriNet achieved an AUC of 0.926, which is significantly higher than the logistic regression integration model (0.904) single deep learning model (0.880), and radiomics method (0.891). The Matthews Correlation Coefficient (MCC) of DensePriNet (0.705) was significantly higher than manual classification by one senior radiologist (0.534) and one junior radiologist (0.416), respectively. CONCLUSION: The performance of the single deep learning method is significantly higher than the radiomics method and the radiologists, and integration of prior clinical knowledge into the deep learning model enhance the diagnostic precision of LN status and facilitate the application of precision medicine.

Preoperative Prediction of Microvascular Invasion in Hepatocellular Carcinoma: Initial Application of a Radiomic Algorithm Based on Grayscale Ultrasound Images.

Objectives: To establish a radiomic algorithm based on grayscale ultrasound images and to make preoperative predictions of microvascular invasion (MVI) in hepatocellular carcinoma (HCC) patients. Methods: In this retrospective study, 322 cases of histopathologically confirmed HCC lesions were included. The classifications based on preoperative grayscale ultrasound images were performed in two stages: (1) classifier #1, MVI-negative and MVI-positive cases; (2) classifier #2, MVI-positive cases were further classified as M1 or M2 cases. The gross-tumoral region (GTR) and peri-tumoral region (PTR) signatures were combined to generate gross- and peri-tumoral region (GPTR) radiomic signatures. The optimal radiomic signatures were further incorporated with vital clinical information. Multivariable logistic regression was used to build radiomic models. Results: Finally, 1,595 radiomic features were extracted from each HCC lesion. At the classifier #1 stage, the radiomic signatures based on features of GTR, PTR, and GPTR showed area under the curve (AUC) values of 0.708 (95% CI, 0.603-0.812), 0.710 (95% CI, 0.609-0.811), and 0.726 (95% CI, 0.625-0.827), respectively. Upon incorporation of vital clinical information, the AUC of the GPTR radiomic algorithm was 0.744 (95% CI, 0.646-0.841). At the classifier #2 stage, the AUC of the GTR radiomic signature was 0.806 (95% CI, 0.667-0.944). Conclusions: Our radiomic algorithm based on grayscale ultrasound images has potential value to facilitate preoperative prediction of MVI in HCC patients. The GTR radiomic signature may be helpful for further discriminating between M1 and M2 levels among MVI-positive patients.

MRI-Based Machine Learning for Differentiating Borderline From Malignant Epithelial Ovarian Tumors: A Multicenter Study.

BACKGROUND: Preoperative differentiation of borderline from malignant epithelial ovarian tumors (BEOT from MEOT) can impact surgical management. MRI has improved this assessment but subjective interpretation by radiologists may lead to inconsistent results. PURPOSE: To develop and validate an objective MRI-based machine-learning (ML) assessment model for differentiating BEOT from MEOT, and compare the performance against radiologists' interpretation. STUDY TYPE: Retrospective study of eight clinical centers. POPULATION: In all, 501 women with histopathologically-confirmed BEOT (n = 165) or MEOT (n = 336) from 2010 to 2018 were enrolled. Three cohorts were constructed: a training cohort (n = 250), an internal validation cohort (n = 92), and an external validation cohort (n = 159). FIELD STRENGTH/SEQUENCE: Preoperative MRI within 2 weeks of surgery. Single- and multiparameter (MP) machine-learning assessment models were built utilizing the following four MRI sequences: T2 -weighted imaging (T2 WI), fat saturation (FS), diffusion-weighted imaging (DWI), apparent diffusion coefficient (ADC), and contrast-enhanced (CE)-T1 WI. ASSESSMENT: Diagnostic performance of the models was assessed for both whole tumor (WT) and solid tumor (ST) components. Assessment of the performance of the model in discriminating BEOT vs. early-stage MEOT was made. Six radiologists of varying experience also interpreted the MR images. STATISTICAL TESTS: Mann-Whitney U-test: significance of the clinical characteristics; chi-square test: difference of label; DeLong test: difference of receiver operating characteristic (ROC). RESULTS: The MP-ST model performed better than the MP-WT model for both the internal validation cohort (area under the curve [AUC] = 0.932 vs. 0.917) and external validation cohort (AUC = 0.902 vs. 0.767). The model showed capability in discriminating BEOT vs. early-stage MEOT, with AUCs of 0.909 and 0.920, respectively. Radiologist performance was considerably poorer than both the internal (mean AUC = 0.792; range, 0.679-0.924) and external (mean AUC = 0.797; range, 0.744-0.867) validation cohorts. DATA CONCLUSION: Performance of the MRI-based ML model was robust and superior to subjective assessment of radiologists. If our approach can be implemented in clinical practice, improved preoperative prediction could potentially lead to preserved ovarian function and fertility for some women. LEVEL OF EVIDENCE: Level 4. TECHNICAL EFFICACY: Stage 2. J. Magn. Reson. Imaging 2020;52:897-904.

Full convolutional network based multiple side-output fusion architecture for the segmentation of rectal tumors in magnetic resonance images: A multi-vendor study.

PURPOSE: Accurate segmentation of rectal tumors is a basic and crucial task for diagnosis and treatment of rectal cancer. To avoid tedious manual delineation, an automatic rectal tumor segmentation model is proposed. METHODS: A pretrained Resnet50 model was introduced for feature extraction. To reduce the complexity of the model, all layers after the 13th residual block of ResNet50 were removed, and three side-output modules were added to the hidden layer of ResNet50 to guide multiscale feature learning. The final boundaries of tumors were determined by fusion of the predictions from side-output modules. The proposed model was compared with two other models, and the effects of different region of interest (ROI) sizes, loss functions, and side-output fusion strategy were also evaluated. RESULTS: The models were trained and evaluated on data from four clinical centers; T2-weighted magnetic resonance images (T2W-MRIs) from 461 patients in the first clinical center were used for training, while T2W-MRIs from 51 patients in the same clinical center and 56 patients in three other clinical centers were used for performance evaluation. The proposed model was superior to the two other models and achieved an average Dice similarity coefficient of 82.39%, sensitivity of 86.32%, specificity of 92.25%, and Hausdorff distance of 12.10 px. In addition, when the ROI contained rectal tumors, the smaller the ROI size, the higher the segmentation accuracy. For a certain ROI size, there were no considerable differences in segmentation results among the loss functions. Compared to the models with single side-output module, the proposed model performed better. CONCLUSIONS: The results show that the proposed model has potential clinical applications in rectal cancer treatment, particularly with regard to therapeutic response evaluation and preoperative planning.

Can peritumoral radiomics increase the efficiency of the prediction for lymph node metastasis in clinical stage T1 lung adenocarcinoma on CT?

OBJECTIVES: To evaluate the efficiency of radiomics model on CT images of intratumoral and peritumoral lung parenchyma for preoperative prediction of lymph node (LN) metastasis in clinical stage T1 peripheral lung adenocarcinoma patients. METHODS: Three hundred sixty-six peripheral lung adenocarcinoma patients with clinical stage T1 were evaluated using five CT scanners. For each patient, two volumes of interest (VOIs) on CT were defined as the gross tumor volume (GTV) and the peritumoral volume (PTV, 1.5 cm around the tumor). One thousand nine hundred forty-six radiomic features were obtained from each VOI, and then refined for reproducibility and redundancy. The refined features were investigated for usefulness in building radiomic signatures by mRMR feature ranking method and LASSO classifier. Multivariable logistic regression analysis was used to develop a radiomic nomogram incorporating the radiomic signature and clinical parameters. The prediction performance was evaluated on the validation cohort. RESULTS: The radiomic signatures using the features of GTV and PTV showed a good ability in predicting LN metastasis with an AUC of 0.829 (95% CI, 0.745-0.913) and 0.825 (95% CI, 0.733-0.918), respectively. By incorporating the features of GTV and PTV, the AUC of radiomic signature increased to 0.843 (95% CI, 0.770-0.916). The AUC of radiomic nomogram was 0.869 (95% CI, 0.800-0.938). CONCLUSIONS: Radiomic signatures of GTV and PTV both had a good prediction ability in the prediction of LN metastasis, and there is no significant difference of AUC between the two groups. The proposed nomogram can be conveniently used to facilitate the preoperative prediction of LN metastasis in T1 peripheral lung adenocarcinomas. KEY POINTS: • Radiomics from peritumoral lung parenchyma increase the efficiency of the prediction for lymph node metastasis in clinical stage T1 lung adenocarcinoma on CT. • A radiomic nomogram was developed and validated to predict LN metastasis. • Different scan parameters on CT showed that radiomics signature had good predictive performance.

Fully convolutional networks (FCNs)-based segmentation method for colorectal tumors on T2-weighted magnetic resonance images.

Segmentation of colorectal tumors is the basis of preoperative prediction, staging, and therapeutic response evaluation. Due to the blurred boundary between lesions and normal colorectal tissue, it is hard to realize accurate segmentation. Routinely manual or semi-manual segmentation methods are extremely tedious, time-consuming, and highly operator-dependent. In the framework of FCNs, a segmentation method for colorectal tumor was presented. Normalization was applied to reduce the differences among images. Borrowing from transfer learning, VGG-16 was employed to extract features from normalized images. We conducted five side-output blocks from the last convolutional layer of each block of VGG-16 along the network, these side-output blocks can deep dive multiscale features, and produced corresponding predictions. Finally, all of the predictions from side-output blocks were fused to determine the final boundaries of the tumors. A quantitative comparison of 2772 colorectal tumor manual segmentation results from T2-weighted magnetic resonance images shows that the average Dice similarity coefficient, positive predictive value, specificity, sensitivity, Hammoude distance, and Hausdorff distance were 83.56, 82.67, 96.75, 87.85%, 0.2694, and 8.20, respectively. The proposed method is superior to U-net in colorectal tumor segmentation (P < 0.05). There is no difference between cross-entropy loss and Dice-based loss in colorectal tumor segmentation (P > 0.05). The results indicate that the introduction of FCNs contributed to accurate segmentation of colorectal tumors. This method has the potential to replace the present time-consuming and nonreproducible manual segmentation method.