Deep Learning Prostate MRI Segmentation Accuracy and Robustness: A Systematic Review.

Myopic maculopathy (MM) is a major cause of vision impairment globally. Artificial intelligence (AI) and deep learning (DL) algorithms for detecting MM from fundus images could potentially improve diagnosis and assist screening in a variety of health care settings. To evaluate DL algorithms for MM classification and segmentation and compare their performance with that of ophthalmologists. The Myopic Maculopathy Analysis Challenge (MMAC) was an international competition to develop automated solutions for 3 tasks: (1) MM classification, (2) segmentation of MM plus lesions, and (3) spherical equivalent (SE) prediction. Participants were provided 3 subdatasets containing 2306, 294, and 2003 fundus images, respectively, with which to build algorithms. A group of 5 ophthalmologists evaluated the same test sets for tasks 1 and 2 to ascertain performance. Results from model ensembles, which combined outcomes from multiple algorithms submitted by MMAC participants, were compared with each individual submitted algorithm. This study was conducted from March 1, 2023, to March 30, 2024, and data were analyzed from January 15, 2024, to March 30, 2024. DL algorithms submitted as part of the MMAC competition or ophthalmologist interpretation. MM classification was evaluated by quadratic-weighted κ (QWK), F1 score, sensitivity, and specificity. MM plus lesions segmentation was evaluated by dice similarity coefficient (DSC), and SE prediction was evaluated by R2 and mean absolute error (MAE). The 3 tasks were completed by 7, 4, and 4 teams, respectively. MM classification algorithms achieved a QWK range of 0.866 to 0.901, an F1 score range of 0.675 to 0.781, a sensitivity range of 0.667 to 0.778, and a specificity range of 0.931 to 0.945. MM plus lesions segmentation algorithms achieved a DSC range of 0.664 to 0.687 for lacquer cracks (LC), 0.579 to 0.673 for choroidal neovascularization, and 0.768 to 0.841 for Fuchs spot (FS). SE prediction algorithms achieved an R2 range of 0.791 to 0.874 and an MAE range of 0.708 to 0.943. Model ensemble results achieved the best performance compared to each submitted algorithms, and the model ensemble outperformed ophthalmologists at MM classification in sensitivity (0.801; 95% CI, 0.764-0.840 vs 0.727; 95% CI, 0.684-0.768; P = .006) and specificity (0.946; 95% CI, 0.939-0.954 vs 0.933; 95% CI, 0.925-0.941; P = .009), LC segmentation (DSC, 0.698; 95% CI, 0.649-0.745 vs DSC, 0.570; 95% CI, 0.515-0.625; P < .001), and FS segmentation (DSC, 0.863; 95% CI, 0.831-0.888 vs DSC, 0.790; 95% CI, 0.742-0.830; P < .001). In this diagnostic study, 15 AI models for MM classification and segmentation on a public dataset made available for the MMAC competition were validated and evaluated, with some models achieving better diagnostic performance than ophthalmologists.

Multiparametric ultrasound (mpUS) is being investigated as an alternative to magnetic resonance imaging (MRI) for detection of prostate cancer (PC). Automated prostate segmentation facilitates workflows, and zonal segmentation can aid in PC diagnosis, accounting for differences in imaging characteristics and tumor incidence. Our aim was to develop a deep learning algorithm that can automatically segment the prostate and its zones on three-dimensional (3D) contrast-enhanced ultrasound (CEUS) and conventional brightness-mode (B-mode) images (NCT04605276). A total of 259 3D mpUS images were collected from men with suspicion for PC in a prospective multicenter trial to develop a computer-aided diagnosis system for PC. Manual segmentation was performed using a custom tool, and an algorithm was developed using a convolutional neural network based on the U-Net architecture. Cross-validation of the automated segmentation algorithm revealed Dice similarity coefficients (DSCs) of 0.91 (95% confidence interval [CI] 0.90-0.91) for CEUS and 0.94 (95% CI 0.93-0.94) for B-mode ultrasound for 3D prostate segmentation. Zonal segmentation was less accurate, with DSCs of 0.83 (95% CI 0.82-0.84) for CEUS and 0.86 (95% CI 0.85-0.87) for B-mode ultrasound. There was high agreement for prostate volume between automatic segmentation on CEUS and physician-estimated volumes on MRI (R2=0.96). Qualitative assessment of prostate segmentation using a scale from 1 to 5 revealed a median grade of 5 (interquartile range [IQR] 4-5) for manual segmentation and 4 (IQR 4-5) for automated segmentation (p=0.10). Our deep learning algorithm demonstrated strong performance for automatic prostate and zonal segmentation from 3D CEUS and B-mode ultrasound images. We developed a computer tool to automatically identify the prostate in three-dimensional ultrasound images. The results show high accuracy and closely match manual assessments by urologists. This tool has potential for use in a computer-aided diagnostic system for prostate cancer.

MP43-02 LESSONS LEARNED IN APPLYING DEEP LEARNING TO FACILITATE PROSTATE MR-US FUSION BIOPSY WORKFLOW

Manual segmentation in lumbar paraspinal muscle studies is time-consuming, subject to variability, and limits comprehensive evaluation. To solve the problems, we aim to develop a deep learning algorithm for the automatic segmentation and multi-parameter quantification of eight paraspinal muscles in lumbar CT. We collected CT images covering the range from the superior endplate of the L1 vertebra to the inferior endplate of the S1 vertebra. The dataset was partitioned into training, validation, and test sets in a ratio of 7:1:2. Six convolutional neural networks were employed to automatically segment the bilateral psoas major, quadratus lumborum, erector spinae, and multifidus across all slices. Model performance was evaluated using the Dice Similarity Coefficient (DSC), Hausdorff Distance (HD), and mean Intersection over Union (mIoU). Muscle parameters including cross-sectional area, muscle volume, fat infiltration, CT density, and paraspinal muscle index were calculated. The Intraclass Correlation Coefficient [ICC(3,1)] was used to assess agreement between model's segmentations and manual ground truth. 100 lumbar CT scans (age, 55.02 ± 16.2; 62 female) were used for model development. The mean DSC across all six models ranged from 0.754 to 0.903. TransUNet demonstrated the best overall (DSC, 0.903; 95% CI: 0.902, 0.913) (HD, 9.934mm; 95% CI: 9.815, 11.902) (mIoU, 0.841; 95% CI: 0.839, 0.850). For individual muscles segmented by TransUNet, the highest DSC was 0.936 (95% CI: 0.934-0.938) for the right multifidus. The mean ICC for all parameters was 0.931. The highest ICC value was 0.999 (p < 0.001). The deep learning tool enables automatic and accurate segmentation of eight lumbar paraspinal muscles. Muscle parameter measurements derived from model's segmentations and manual ground truth exhibit high agreement. Accurate segmentation and quantification of paraspinal muscle parameters facilitates large-scale epidemiological studies of paraspinal muscles and spine-related diseases, overcoming the limitations in manual methods.

Deep leaning in food safety and authenticity detection: An integrative review and future prospects

OBJECTIVE. Deep learning applications in radiology often suffer from overfitting, limiting generalization to external centers. The objective of this study was to develop a high-quality prostate segmentation model capable of maintaining a high degree of performance across multiple independent datasets using transfer learning and data augmentation. MATERIALS AND METHODS. A retrospective cohort of 648 patients who underwent prostate MRI between February 2015 and November 2018 at a single center was used for training and validation. A deep learning approach combining 2D and 3D architecture was used for training, which incorporated transfer learning. A data augmentation strategy was used that was specific to the deformations, intensity, and alterations in image quality seen on radiology images. Five independent datasets, four of which were from outside centers, were used for testing, which was conducted with and without fine-tuning of the original model. The Dice similarity coefficient was used to evaluate model performance. RESULTS. When prostate segmentation models utilizing transfer learning were applied to the internal validation cohort, the mean Dice similarity coefficient was 93.1 for whole prostate and 89.0 for transition zone segmentations. When the models were applied to multiple test set cohorts, the improvement in performance achieved using data augmentation alone was 2.2% for the whole prostate models and 3.0% for the transition zone segmentation models. However, the best test-set results were obtained with models fine-tuned on test center data with mean Dice similarity coefficients of 91.5 for whole prostate segmentation and 89.7 for transition zone segmentation. CONCLUSION. Transfer learning allowed for the development of a high-performing prostate segmentation model, and data augmentation and fine-tuning approaches improved performance of a prostate segmentation model when applied to datasets from external centers.

BackgroundThe development of deep learning (DL) algorithms is a three-step process—training, tuning, and testing. Studies are inconsistent in the use of the term “validation”, with some using it to refer to tuning and others testing, which hinders accurate delivery of information and may inadvertently exaggerate the performance of DL algorithms. We investigated the extent of inconsistency in usage of the term “validation” in studies on the accuracy of DL algorithms in providing diagnosis from medical imaging.Methods and findingsWe analyzed the full texts of research papers cited in two recent systematic reviews. The papers were categorized according to whether the term “validation” was used to refer to tuning alone, both tuning and testing, or testing alone. We analyzed whether paper characteristics (i.e., journal category, field of study, year of print publication, journal impact factor [JIF], and nature of test data) were associated with the usage of the terminology using multivariable logistic regression analysis with generalized estimating equations. Of 201 papers published in 125 journals, 118 (58.7%), 9 (4.5%), and 74 (36.8%) used the term to refer to tuning alone, both tuning and testing, and testing alone, respectively. A weak association was noted between higher JIF and using the term to refer to testing (i.e., testing alone or both tuning and testing) instead of tuning alone (vs. JIF <5; JIF 5 to 10: adjusted odds ratio 2.11, P = 0.042; JIF >10: adjusted odds ratio 2.41, P = 0.089). Journal category, field of study, year of print publication, and nature of test data were not significantly associated with the terminology usage.ConclusionsExisting literature has a significant degree of inconsistency in using the term “validation” when referring to the steps in DL algorithm development. Efforts are needed to improve the accuracy and clarity in the terminology usage.

The development of deep learning (DL) algorithms is a three-step process-training, tuning, and testing. Studies are inconsistent in the use of the term "validation", with some using it to refer to tuning and others testing, which hinders accurate delivery of information and may inadvertently exaggerate the performance of DL algorithms. We investigated the extent of inconsistency in usage of the term "validation" in studies on the accuracy of DL algorithms in providing diagnosis from medical imaging. We analyzed the full texts of research papers cited in two recent systematic reviews. The papers were categorized according to whether the term "validation" was used to refer to tuning alone, both tuning and testing, or testing alone. We analyzed whether paper characteristics (i.e., journal category, field of study, year of print publication, journal impact factor [JIF], and nature of test data) were associated with the usage of the terminology using multivariable logistic regression analysis with generalized estimating equations. Of 201 papers published in 125 journals, 118 (58.7%), 9 (4.5%), and 74 (36.8%) used the term to refer to tuning alone, both tuning and testing, and testing alone, respectively. A weak association was noted between higher JIF and using the term to refer to testing (i.e., testing alone or both tuning and testing) instead of tuning alone (vs. JIF <5; JIF 5 to 10: adjusted odds ratio 2.11, P = 0.042; JIF >10: adjusted odds ratio 2.41, P = 0.089). Journal category, field of study, year of print publication, and nature of test data were not significantly associated with the terminology usage. Existing literature has a significant degree of inconsistency in using the term "validation" when referring to the steps in DL algorithm development. Efforts are needed to improve the accuracy and clarity in the terminology usage.

Deep learning has gained immense popularity in scientific computing, and its algorithms are widely used in complex problem-solving industries. Every deep learning algorithm use different types of neural networks to perform indented tasks. Deep learning (DL) algorithms have emerged from different machine learning and soft computing methodologies. Since then, a number of deep learning (DL) algorithms have been recently introduced in the scientific community and applied in various application fields. Today, the use of DLs has become indispensable due to their intelligence, effective learning, accuracy and reliability in model creation. However, a comprehensive list of DL algorithms has not yet been presented in the scientific literature. This article lists the most popular DL algorithms and their application areas. Deep learning uses ANN artificial neural networks to perform convoluted calculations on huge amounts of data. It is a type of machine learning based on the structure and function of the human brain. Deep learning algorithms train machines by learning from examples. Industries such as healthcare, e-commerce, entertainment and advertising often use deep learning.

Previous deep learning (DL) algorithms for brain metastasis (BM) detection and segmentation have not been commonly used in clinics because they produce false-positive findings, require multiple sequences, and do not reflect physiological properties such as necrosis. The aim of this study was to develop a more clinically favorable DL algorithm (RLK-Unet) using a single sequence reflecting necrosis and apply it to automated treatment response assessment. A total of 128 patients with 1339 BMs, who underwent BM magnetic resonance imaging using the contrast-enhanced 3D T1 weighted (T1WI) turbo spin-echo black blood sequence, were included in the development of the DL algorithm. Fifty-eight patients with 629 BMs were assessed for treatment response. The detection sensitivity, precision, Dice similarity coefficient (DSC), and agreement of treatment response assessments between neuroradiologists and RLK-Unet were assessed. RLK-Unet demonstrated a sensitivity of 86.9% and a precision of 79.6% for BMs and had a DSC of 0.663. Segmentation performance was better in the subgroup with larger BMs (DSC, 0.843). The agreement in the response assessment for BMs between the radiologists and RLK-Unet was excellent (intraclass correlation, 0.84). RLK-Unet yielded accurate detection and segmentation of BM and could assist clinicians in treatment response assessment.

Detection of Lisfranc malalignment leading to the instability of the joint, particularly in subtle cases, has been a concern for foot and ankle care providers. X-ray radiographs are the mainstay in the diagnosis of these injuries; thus, improving the performance of clinicians in interpreting radiographs can noticeably affect the quality of health care in these patients. Here we assessed the performance of deep learning algorithms on weightbearing radiographs for detection of Lisfranc joint malalignment in patients with Lisfranc instability. In a retrospective study, 640 patients with Lisfranc malalignment leading to instability were recruited plus 640 individuals with uninjured feet and healthy Lisfranc joint as the control group. All radiographs were screened by orthopaedic surgeons. Two deep learning models were trained, validated, and tested (in a ratio 80:10:10) using a single-view (anteroposterior) and 3-view (anteroposterior, lateral, oblique) radiographs. The performances of the models were reported as sensitivity, specificity, positive and negative predictive values, accuracy, F score, and area under the curve (AUC). No significant differences were observed between the patients and the controls regarding age, gender, race, and body mass index. The best deep learning algorithm outperformed our human interpreters (<1% vs ~10% misdiagnosis), 94.8% sensitivity, 96.9% specificity, 98.6% accuracy, 95.8% F score, and 99.4% AUC. Deep learning methods have shown promising potential in acting as an assistant interpreter of radiographic images in patients with Lisfranc malalignment. Developing these algorithms can hasten and improve the accuracy of diagnosis and reduce further costs and burdens on the patients and health care system. Level III, case-control Machine Learning study.

To design and evaluate an automated deep learning method for segmentation and analysis of cardiac MRI T1 maps with use of synthetic T1-weighted images for MRI relaxation-based contrast augmentation. This retrospective study included MRI scans acquired between 2016 and 2019 from 100 patients (mean age ± SD, 55 years ± 13; 72 men) across various clinical abnormalities with use of a modified Look-Locker inversion recovery, or MOLLI, sequence to quantify native T1 (T1native), postcontrast T1 (T1post), and extracellular volume (ECV). Data were divided into training (n = 60) and internal (n = 40) test subsets. "Synthetic" T1-weighted images were generated from the T1 exponential inversion-recovery signal model at a range of optimal inversion times, yielding high blood-myocardium contrast, and were used for contrast-based image augmentation during training and testing of a convolutional neural network for myocardial segmentation. Automated segmentation, T1, and ECV were compared with experts with use of Dice similarity coefficients (DSCs), correlation coefficients, and Bland-Altman analysis. An external test dataset (n = 147) was used to assess model generalization. Internal testing showed high myocardial DSC relative to experts (0.81 ± 0.08), which was similar to interobserver DSC (0.81 ± 0.08). Automated segmental measurements strongly correlated with experts (T1native, R = 0.87; T1post, R = 0.91; ECV, R = 0.92), which were similar to interobserver correlation (T1native, R = 0.86; T1post, R = 0.94; ECV, R = 0.95). External testing showed strong DSC (0.80 ± 0.09) and T1native correlation (R = 0.88) between automatic and expert analysis. This deep learning method leveraging synthetic contrast augmentation may provide accurate automated T1 and ECV analysis for cardiac MRI data acquired across different abnormalities, centers, scanners, and T1 sequences.Keywords: MRI, Cardiac, Tissue Characterization, Segmentation, Convolutional Neural Network, Deep Learning Algorithms, Machine Learning Algorithms, Supervised Learning Supplemental material is available for this article. © RSNA, 2022.

Background Machine learning (ML) algorithm support vector machine (SVM) performed better than Thrombolysis in Myocardial Infarction (TIMI) score for ASIAN STEMI patients. However, Deep Learning (DL) effectiveness in the multiethnic ASIAN population has yet to be determined. DL has automatic learning of the feature from a given dataset without the need to conduct feature selection. However, the selected features by the algorithm is black box. Identifying features associated with mortality is essential to recognize characteristics of patients with high risk for better patient management. Purpose To develop a DL algorithm for in-hospital mortality in multiethnic STEMI patients using predictors identified from the SVM algorithm. To investigate DL performance constructed using predictors from SVM feature extraction and expert-recommended predictors. Methods We constructed four algorithms; a) DL and SVM algorithms with predictors identified from the SVM variable importance b) DL and SVM using predictors based on expert recommendation. We used registry data from the National Cardiovascular Disease Database of 11397 patient's. Fifty parameters including demographics, cardiovascular risk, medications and clinical variables were considered. The Area under the curve (AUC) is the performance evaluation metric. Algorithms were validated against the TIMI and tested using the same validation data. SVM variable importance with backward elimination was used to select and rank important variables. Results DL algorithms outperform SVM and TIMI on the validation dataset; i) DL with SVM selected predictors (15 predictors, AUC = 0.97), ii) DL with expert-recommended predictors (16 predictors, AUC = 0.96), iii) SVM with selected predictors (15 predictors, AUC = 0.92), iv) SVM with expert-recommended predictors (AUC = 0.89) and TIMI (AUC = 0.82). Common predictors across SVM feature selection, expert-recommendation and TIMI are: age, heart rate, Killip class, fasting blood glucose, systolic blood pressure, comorbid diseases and ST-elevation. SVM feature selection also identified diuretics, PCI and pharmacotherapy drugs as predictors that improve mortality prediction in STEMI patients. Our findings suggest that the TIMI score underestimates patients risk of mortality. DL algorithm using selected predictors classified 35% of nonsurvival patients as high risk (risk probabilities &amp;gt;50%) compared to only 12.7% nonsurvival patients by TIMI (score &amp;gt;5) (Figure below). Conclusions In the ASIAN population, patients with STEMI can be better classified using the DL algorithm compared to the ML and TIMI score. Combining ML feature selection with DL allows the identification of distinct factors in a unique ASIAN population for better mortality prediction than relying solely on an expert recommendation as it is a very subjective approach. Continuous validation on population-specific algorithms using DL and ML is needed before implementing in a real clinical setting. Funding Acknowledgement Type of funding sources: Public grant(s) – National budget only. Main funding source(s): Technology Development Fund 1 TIMI performance on validation setDL performance on validation set

Simple SummaryUterine cervical cancer (CC) is a leading cause of cancer-related deaths in women worldwide. Pelvic magnetic resonance imaging (MRI) allows the assessment of local tumor extent and guides the choice of primary treatment. MRI tumor segmentation enables whole-volume radiomic tumor profiling, which is potentially useful for prognostication and individualization of therapy in CC. Manual tumor segmentation is, however, labor intensive and thus not part of routine clinical workflow. In the current work, we trained a deep learning (DL) algorithm to automatically segment the primary tumor in CC patients. Although the achieved segmentation performance of the trained DL algorithm is slightly lower than that for human experts, it is still relatively good. This study suggests that automated MRI primary tumor segmentations by DL algorithms without any human interaction is possible in patients with CC.Uterine cervical cancer (CC) is the most common gynecologic malignancy worldwide. Whole-volume radiomic profiling from pelvic MRI may yield prognostic markers for tailoring treatment in CC. However, radiomic profiling relies on manual tumor segmentation which is unfeasible in the clinic. We present a fully automatic method for the 3D segmentation of primary CC lesions using state-of-the-art deep learning (DL) techniques. In 131 CC patients, the primary tumor was manually segmented on T2-weighted MRI by two radiologists (R1, R2). Patients were separated into a train/validation (n = 105) and a test- (n = 26) cohort. The segmentation performance of the DL algorithm compared with R1/R2 was assessed with Dice coefficients (DSCs) and Hausdorff distances (HDs) in the test cohort. The trained DL network retrieved whole-volume tumor segmentations yielding median DSCs of 0.60 and 0.58 for DL compared with R1 (DL-R1) and R2 (DL-R2), respectively, whereas DSC for R1-R2 was 0.78. Agreement for primary tumor volumes was excellent between raters (R1-R2: intraclass correlation coefficient (ICC) = 0.93), but lower for the DL algorithm and the raters (DL-R1: ICC = 0.43; DL-R2: ICC = 0.44). The developed DL algorithm enables the automated estimation of tumor size and primary CC tumor segmentation. However, segmentation agreement between raters is better than that between DL algorithm and raters.

Accurate automatic segmentation of the prostate in magnetic resonance images (MRI) is a challenging task due to the high variability of prostate anatomic structure. Artifacts such as noise and similar signal intensity of tissues around the prostate boundary inhibit traditional segmentation methods from achieving high accuracy. We investigate both patch-based and holistic (image-to-image) deep-learning methods for segmentation of the prostate. First, we introduce a patch-based convolutional network that aims to refine the prostate contour which provides an initialization. Second, we propose a method for end-to-end prostate segmentation by integrating holistically nested edge detection with fully convolutional networks. Holistically nested networks (HNN) automatically learn a hierarchical representation that can improve prostate boundary detection. Quantitative evaluation is performed on the MRI scans of 250 patients in fivefold cross-validation. The proposed enhanced HNN model achieves a mean ± standard deviation. A Dice similarity coefficient (DSC) of [Formula: see text] and a mean Jaccard similarity coefficient (IoU) of [Formula: see text] are used to calculate without trimming any end slices. The proposed holistic model significantly ([Formula: see text]) outperforms a patch-based AlexNet model by 9% in DSC and 13% in IoU. Overall, the method achieves state-of-the-art performance as compared with other MRI prostate segmentation methods in the literature.