Achieved Area Under The Curve Research Articles

Classification of Diabetic Retinopathy using Pre-Trained Deep Learning Models

Diabetic Retinopathy (DR) stands as the leading cause of blindness globally, particularly affecting individuals between the ages of 20 and 70. This paper presents a Computer- Aided Diagnosis (CAD) system designed for the automatic classification of retinal images into five distinct classes: Normal, Mild, Moderate, Severe, and Proliferative Diabetic Retinopathy (PDR). The proposed system leverages Convolutional Neural Networks (CNNs) employing pre-trained deep learning models. Through the application of fine-tuning techniques, our model is trained on fundus images of diabetic retinopathy with resolutions of 350x350x3 and 224x224x3. Experimental results obtained on the Kaggle platform, utilizing resources comprising 4 CPUs, 17 GB RAM, and 1 GB Disk, demonstrate the efficacy of our approach. The achieved Area Under the Curve (AUC) values for CNN, MobileNet, VGG-16, InceptionV3, and InceptionResNetV2 models are 0.50, 0.70, 0.53, 0.63, and 0.69, respectively.

Multi-task reconstruction network for synthetic diffusion kurtosis imaging: Predicting neoadjuvant chemoradiotherapy response in locally advanced rectal cancer

PurposeTo assess the feasibility and clinical value of synthetic diffusion kurtosis imaging (DKI) generated from diffusion weighted imaging (DWI) through multi-task reconstruction network (MTR-Net) for tumor response prediction in patients with locally advanced rectal cancer (LARC). MethodsIn this retrospective study, 120 eligible patients with LARC were enrolled and randomly divided into training and testing datasets with a 7:3 ratio. The MTR-Net was developed for reconstructing Dapp and Kapp images from apparent diffusion coefficient (ADC) images. Tumor regions were manually segmented on both true and synthetic DKI images. The synthetic image quality and manual segmentation agreement were quantitatively assessed. The support vector machine (SVM) classifier was used to construct radiomics models based on the true and synthetic DKI images for pathological complete response (pCR) prediction. The prediction performance for the models was evaluated by the receiver operating characteristic (ROC) curve analysis. ResultsThe mean squared error (MSE), peak signal-to-noise ratio (PSNR), and structural similarity index measure (SSIM) for tumor regions were 0.212, 24.278, and 0.853, respectively, for the synthetic Dapp images and 0.516, 24.883, and 0.804, respectively, for the synthetic Kapp images. The Dice similarity coefficient (DSC), positive predictive value (PPV), sensitivity (SEN), and Hausdorff distance (HD) for the manually segmented tumor regions were 0.786, 0.844, 0.755, and 0.582, respectively. For predicting pCR, the true and synthetic DKI-based radiomics models achieved area under the curve (AUC) values of 0.825 and 0.807 in the testing datasets, respectively. ConclusionsGenerating synthetic DKI images from DWI images using MTR-Net is feasible, and the efficiency of synthetic DKI images in predicting pCR is comparable to that of true DKI images.

MRI-Based Kinetic Heterogeneity Evaluation in the Accurate Access of Axillary Lymph Node Status in Breast Cancer Using a Hybrid CNN-RNN Model.

Accurate evaluation of the axillary lymph node (ALN) status is needed for determining the treatment protocol for breast cancer (BC). The value of magnetic resonance imaging (MRI)-based tumor heterogeneity in assessing ALN metastasis in BC is unclear. To assess the value of deep learning (DL)-derived kinetic heterogeneity parameters based on BC dynamic contrast-enhanced (DCE)-MRI to infer the ALN status. Retrospective. 1256/539/153/115 patients in the training cohort, internal validation cohort, and external validation cohorts I and II, respectively. 1.5 T/3.0 T, non-contrast T1-weighted spin-echo sequence imaging (T1WI), DCE-T1WI, and diffusion-weighted imaging. Clinical pathological and MRI semantic features were obtained by reviewing histopathology and MRI reports. The segmentation of the tumor lesion on the first phase of T1WI DCE-MRI images was applied to other phases after registration. A DL architecture termed convolutional recurrent neural network (ConvRNN) was developed to generate the KHimage (kinetic heterogeneity of DCE-MRI image) score that indicated the ALN status in patients with BC. The model was trained and optimized on training and internal validation cohorts, tested on two external validation cohorts. We compared ConvRNN model with other 10 models and the subgroup analyses of tumor size, magnetic field strength, and molecular subtype were also evaluated. Chi-squared, Fisher's exact, Student's t, Mann-Whitney U tests, and receiver operating characteristics (ROC) analysis were performed. P < 0.05 was considered significant. The ConvRNN model achieved area under the curve (AUC) of 0.802 in the internal validation cohort and 0.785-0.806 in the external validation cohorts. The ConvRNN model could well evaluate the ALN status of the four molecular subtypes (AUC = 0.685-0.868). The patients with larger tumor sizes (>5 cm) were more susceptible to ALN metastasis with KHimage scores of 0.527-0.827. A ConvRNN model outperformed traditional models for determining the ALN status in patients with BC. 3 TECHNICAL EFFICACY: Stage 2.

Machine learning models for predicting in-hospital outcomes after non-surgical treatment among patients with moderate-to-severe traumatic brain injury

AimThis study aims to develop prediction models for in-hospital outcomes after non-surgical treatment among patients with moderate-to-severe traumatic brain injury (TBI). MethodWe conducted a retrospective review of patients hospitalized for moderate-to-severe TBI in our department from 2011 to 2020. Five machine learning (ML) algorithms and the conventional logistic regression (LR) model were employed to predict in-hospital mortality and the Glasgow Outcome Scale (GOS) functional outcomes. These models utilized clinical and routine blood data collected upon admission. ResultsThis study included a total of 196 patients who received only non-surgical treatment after moderate-to-severe TBI. When predicting mortality, ML models achieved area under the curve (AUC) values of 0.921 to 0.994 using clinical and routine blood data, and 0.877 to 0.982 using only clinical data. In comparison, LR models yielded AUCs of 0.762 and 0.730 respectively. When predicting the GOS outcome, ML models achieved AUCs of 0.870 to 0.915 using clinical and routine blood data, and 0.858 to 0.927 using only clinical data. In comparison, the LR model yielded AUCs of 0.798 and 0.787 respectively. Repeated internal validation showed that the contributions of routine blood data for prediction models may depend on different prediction algorithms and different outcome measurements. ConclusionThe study reported ML-based prediction models that provided rapid and accurate predictions on short-term outcomes after non-surgical treatment among patients with moderate-to-severe TBI. The study also highlighted the superiority of ML models over conventional LR models and proposed the complex contributions of routine blood data in such predictions.

High-throughput detection allied with machine learning for precise monitoring of significant serum metabolic changes in Helicobacter pylori infection

High-throughput detection of large-scale samples is the foundation for rapidly accessing massive metabolic data in precision medicine. Machine learning is a powerful tool for uncovering valuable information hidden within massive data. In this work, we achieved the extraction of a single fingerprinting of 1 μL serum within 5 s through a high-throughput detection platform based on functionalized nanoparticles. We quickly obtained over a thousand serum metabolic fingerprintings (SMFs) including those of individuals with Helicobacter pylori (HP) infection. Combining four classical machine learning models and enrichment analysis, we attempted to extract and confirm useful information behind these SMFs. Based on all fingerprint signals, all four models achieved area under the curve (AUC) values of 0.983–1. In particular, orthogonal partial least squares discriminant analysis (OPLS-DA) model obtained value of 1 in both the discovery and validation sets. Fortunately, we identified six significant metabolic features, all of which can greatly contribute to the monitoring of HP infection, with AUC values ranging from 0.906 to 0.985. The combination of these six significant metabolic features can enable the precise monitoring of HP infection in serum, with over 95 % of accuracy, specificity and sensitivity. The OPLS-DA model displayed optimal performance and the corresponding scatter plot visualized the clear distinction between HP and HC. Interestingly, they exhibit a consistent reduction trend compared to healthy controls, prompting us to explore the possible metabolic pathways and potential mechanism. This work demonstrates the potential alliance between high-throughput detection and machine learning, advancing their application in precision medicine.

Intratumoural and peritumoural CT-based radiomics for diagnosing lepidic-predominant adenocarcinoma in patients with pure ground-glass nodules: a machine learning approach

To develop and validate a diagnostic model utilising machine-learning algorithms that differentiates lepidic predominant adenocarcinoma (LPA) from other pathological subtypes in patients with pure ground-glass nodules (pGGNs). This bicentric study was conducted across two medical centres and included 151 patients diagnosed with lung adenocarcinoma based on histopathological confirmation of pGGNs. The training cohort consisted of 99 patients from Institution 1, while the test cohort included 52 patients from Institution 2. Radiomics features were extracted from both tumours and the 2 mm peritumoural parenchyma. The tumoural and peritumoural radiomics were designated as Modeltumoural and Modelperitumoural, respectively. The diagnostic efficacy of various models was evaluated through the receiver operating characteristic (ROC) curve analysis. Subsequently, a machine-learning-based prediction model that combined Modeltumoural, Modelperitumoural, and Modelclinical-radiological was developed to differentiate LPA from other pathological subtypes in patients with pGGNs. Modeltumoural achieved area under the curve (AUC) values of 0.762 and 0.783 in the training and validation sets, respectively. Modelperitumoural attained AUCs of 0.742 and 0.667, and Modelclinical-radiological generated an AUC of 0.727 and 0.739 in the training and validation sets, respectively. Among the machine-learning models evaluated, gradient boosting machines demonstrated the best diagnostic efficacy, with accuracy, AUC, F1 score, and log loss values of 0.885, 0.956, 0.943, and 0.260, respectively. The combined model based on machine learning that incorporated tumour and peritumoural parenchyma, as well as clinical and imaging characteristics, may offer benefits in assessing the pathological subtype of pGGNs.

Detecting train driveshaft damages using accelerometer signals and Differential Convolutional Neural Networks

Maintaining railway axles is crucial to prevent catastrophic failures and enormous human and economic costs. In recent years, there has been a growing interest in the railway industry to adopt condition monitoring techniques to enhance the safety and efficiency of the rail transport system, which maintenance is currently based on periodic inspections. In this context, this work presents a technique for real-time crack diagnosis on railway axles, based on advanced 2D-Convolutional Neural Network (CNN) architectures applied to time–frequency representations of vibration signals. One of the critical novelties is introducing a differential CNN structure that captures the system’s statistical properties, enabling generalisation between different mechanical sets and conditions. The proposed system has been trained with data corresponding to a unique wheelset assembly, showing that the model is able to diagnose cracks on the three different wheelset tested in operation under 32 different combinations of conditions, such as load, speed, sense of rotation and vibration direction. Four different crack levels have been introduced, representing the maximum one a 0.08% of the axle diameter, and the method proposed achieved Area Under the Curve (AUC) score of 0.85, significantly outperforming results obtained with other architectures proposed in the state-of-the-art, the score of the next below is 0.76. The results demonstrate the effectiveness and practicality of this approach to accurately classify the four crack levels tested within a condition monitoring system for non-stationary conditions, that would enable reliable real-time diagnosis, thus paving the way towards a more robust and efficient railway axle maintenance system.

Abstract 5377: An integrative prognostic machine learning model in mantle cell lymphoma

Abstract Background: Mantle cell lymphoma (MCL) is an uncommon B-cell lymphoma. The clinical course is highly variable: some patients have aggressive disease and relapse after treatment, while others have indolent disease or respond exceptionally to frontline therapy. Prognostication of MCL patients is dynamic and continues to evolve as novel therapies develop. Current prognostic indicators, such as the MCL international prognostic index (MIPI), were primarily designed with patients treated with chemo-immunotherapies. Using machine learning (ML) and molecular data, we provide a novel predictive method to improve upon conventional clinical markers. Methods: We studied 785 MCL patients diagnosed at MD Anderson since 2014 and retrospectively classified them as “aggressive MCL” (n=311): relapsed or refractory to frontline treatment, and “mild MCL” (n=474): those who did not relapse after the first treatment (exceptional response) or had indolent disease never requiring treatment. After data extraction and feature engineering, 195 baseline features comprised of clinical, genomic, pathology, and cytogenetic data were integrated into an extreme gradient-boosted ensemble ML model (XGBoost). The dataset containing all patients was split (75/25) into a training and test set. Hyperparameters for the model were tuned using a grid-based, space-filling (Latin hypercube) technique and resampled 10-fold cross-validation from the training set. Training, validation, and testing sets were split using stratification of the classification variable to avoid class imbalance. Results: Our integrative model achieved area under the curve (AUC) = .82 and accuracy = .76 on the test set and outperformed an XGBoost model using only clinical features (AUC = .78, accuracy = .68). Additionally, the fully integrated model improved on metrics from a similar multivariate logistic model including all patients (AUC = .72, accuracy =.72). Univariate logistic models were fit on the classification variable using the MIPI and other prognostic indices. The integrated ML model significantly outperformed the MIPI (AUC = .62, accuracy = .60) and other indices in predicting patient class. Clinical, pathological, cytogenetic, and genomic data were all represented as impactful features in a variable importance plot (VIP) and Shapley (SHAP) additive values of the fully integrated ML model. This model was used to launch a rest application programming interface (API) in which important features could be entered and a prediction returned. Conclusion: Our study demonstrates that the current paradigm of using limited features in disease prognostics should be replaced with more advanced ML models that utilize genomic and other molecular data. Future work will include expanding the features included in the model and using the rest API to construct a graphical user interface accessible to clinicians and other researchers to make treatment decisions in precision oncology. Citation Format: Holly A. Hill, Preetesh Jain, Michael L. Wang, Ken Chen. An integrative prognostic machine learning model in mantle cell lymphoma. [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2023; Part 1 (Regular and Invited Abstracts); 2023 Apr 14-19; Orlando, FL. Philadelphia (PA): AACR; Cancer Res 2023;83(7_Suppl):Abstract nr 5377.

Systematic radiomics analysis based on multiparameter MRI to preoperatively predict the expression of Ki67 and histological grade in patients with bladder cancer.

Bladder cancer is among the most prevalent urothelial malignancies. Radiomics-based preoperative prediction of Ki67 and histological grade will facilitate clinical decision-making. This retrospective study recruited 283 bladder cancer patients between 2012 and 2021. Multiparameter MRI sequences included: T1WI, T2WI, diffusion-weighted imaging (DWI), and dynamic contrast-enhanced (DCE) imaging. The radiomics features of intratumoral and peritumoral regions were extracted simultaneously. Max-Relevance and Min-Redundancy (mRMR) and least absolute shrinkage and selection operator (LASSO) algorithms were employed to select the features. Six machine learning-based classifiers were adopted to construct the radiomics models, and the best was chosen for the model construction. The mRMR and LASSO algorithms were more suitable for Ki67 and histological grade, respectively. Additionally, Ki67 had a higher proportion of intratumoral features, while peritumoral features accounted for a greater proportion of the histological grade. Random forests performed the best in predicting both pathological outcomes. Consequently, the multiparameter MRI (MP-MRI) models achieved area under the curve (AUC) values of 0.977 and 0.852 for Ki67 in training and test sets, respectively, and 0.972 and 0.710 for the histological grade. Radiomics holds the potential to predict multiple pathological outcomes of bladder cancer preoperatively and are expected to provide clinical decision-making guidance. Furthermore, our work inspired the process of radiomics research. This study demonstrated that different feature selection techniques, segmentation regions, classifiers, and MRI sequences will affect the performance of the model. We systematically demonstrated that radiomics can predict histological grade and Ki67.

MitoSomatic: a tool for accurate identification of mitochondrial DNA somatic mutations without paired controls.

Mitochondrial DNA (mtDNA) somatic mutations play important roles in the initiation and progression of cancer. Although next-generation sequencing (NGS) of paired tumor and control samples has become a common practice to identify tumor-specific mtDNA mutations, the unique nature of mtDNA and NGS-associated sequencing bias could cause false-positive/-negative somatic mutation calling. Additionally, there are clinical scenarios where matched control tissues are unavailable for comparison. Therefore, a novel approach for accurately identifying somatic mtDNA variants is greatly needed, particularly in the absence of matched controls. In this study, the ground truth mtDNA variants orthogonally validated by triple-paired tumor, adjacent nontumor, and blood samples were used to develop mitoSomatic, a random forest-based machine learning tool. We demonstrated that mitoSomatic achieved area under the curve (AUC) values over 0.99 for identifying somatic mtDNA variants without paired control in three tumor types. In addition, mitoSomatic was also applicable in nontumor tissues such as adjacent nontumor and blood samples, suggesting the flexibility of mitoSomatic's classification capability. Furthermore, analysis of triple-paired samples identified a small group of variants with uncertain somatic/germline origin, whereas application of mitoSomatic significantly facilitated the prediction of their possible source. Finally, a control-free evaluation of the public pan-cancer NGS dataset with mitoSomatic revealed a substantial number of variants that were probably misclassified by conventional tumor-control comparison, further emphasizing the usefulness of mitoSomatic in application. Taken together, our study demonstrates that mitoSomatic is valuable for accurately identifying somatic mtDNA variants in mtDNA NGS data without paired controls, applicable for both tumor and nontumor tissues.

Urinary marker panels for aggressive prostate cancer detection

Majority of patients with indolent prostate cancer (PCa) can be managed with active surveillance. Therefore, finding biomarkers for classifying patients between indolent and aggressive PCa is essential. In this study, we investigated urinary marker panels composed of urinary glycopeptides and/or urinary prostate-specific antigen (PSA) for their clinical utility in distinguishing non-aggressive (Grade Group 1) from aggressive (Grade Group ≥ 2) PCa. Urinary glycopeptides acquired via data-independent acquisition mass spectrometry (DIA-MS) were quantitatively analyzed, where prostatic acid phosphatase (ACPP), clusterin (CLU), alpha-1-acid glycoprotein 1 (ORM1), and CD antigen 97 (CD97) were selected to be evaluated in various combinations with and without urinary PSA. Targeted parallel reaction monitoring (PRM) assays of the glycopeptides from urinary ACPP and CLU were investigated along with urinary PSA for the ability of aggressive PCa detection. The multi-urinary marker panels, combined via logistic regression, were statistically evaluated using bootstrap resampling and validated by an independent cohort. Majority of the multi-urinary marker panels (e.g., a panel consisted of ACPP, CLU, and Urinary PSA) achieved area under the curve (AUC) ranged from 0.70 to 0.85. Thus, multi-marker panels investigated in this study showed clinically meaningful results on aggressive PCa detection to separate Grade Group 1 from Grade Group 2 and above warranting further evaluation in clinical setting in future.

18F-FDG PET/CT radiomic analysis for classifying and predicting microvascular invasion in hepatocellular carcinoma and intrahepatic cholangiocarcinoma.

BackgroundMicrovascular invasion (MVI) is a critical risk factor for early recurrence of hepatocellular carcinoma (HCC) and intrahepatic cholangiocarcinoma (ICC). The aim of this study was to explore the contribution of 18F-fluorodeoxyglucose positron emission tomography/computed tomography (18F-FDG PET/CT) radiomic features for the preoperative prediction of HCC and ICC classification and MVI.MethodsIn this retrospective study, 127 (HCC: ICC =76:51) patients with suspected MVI accompanied by either HCC or ICC were included (In HCC group, MVI positive: negative =46:30 in ICC group, MVI positive: negative =31:20). Results-driven feature engineering workflow was used to select the most predictive feature combinations. The prediction model was based on supervised machine learning classifier. Ten-fold cross validation on training cohort and independent test cohort were constructed to ensure stability and generalization ability of models.ResultsFor HCC and ICC classification, radiomics predictors composed of two PET and one CT feature achieved area under the curve (AUC) of 0.86 (accuracy, sensitivity, specificity was 0.82, 0.78, 0.88, respectively) on test cohort. For MVI prediction, in HCC group, our MVI prediction model achieved AUC of 0.88 (accuracy, sensitivity, specificity was 0.78, 0.88, 0.60 respectively) with three PET features associated with tumor stage on test cohort. In ICC group, the phenotype composed of two PET features and carbohydrate antigen 19-9 (CA19-9) achieved AUC of 0.90 (accuracy, sensitivity, specificity was 0.77, 0.75, 0.80, respectively).Conclusions18F-FDG PET/CT radiomic features integrating clinical factors have potential in HCC and ICC classification and MVI prediction, while PET features have dominant predictive power in model performance. The prediction model has value in providing a non-invasive biomarker for an earlier indication and comprehensive quantification of primary liver cancers.

Machine Learning to Delineate Surgeon and Clinical Factors That Anticipate Positive Surgical Margins After Robot-Assisted Radical Prostatectomy.

Purpose: Automated performance metrics (APMs), derived from instrument kinematic and systems events data during robotic surgery, are validated objective measures of surgeon performance. Our previous studies showed that APMs are strong outcome predictors of urinary continence after robot-assisted radical prostatectomy (RARP). We now use machine learning to investigate how surgeon performance (i.e., APMs) and clinical factors can predict positive surgical margins (PSMs) after RARP. Methods: We prospectively collected data of patients undergoing RARP at our institution from 2016 to 2019. Random Forest model predicted PSMs based on 15 clinical factors and 38 APMs from 11 standardized RARP steps. Out-of-bag Gini impurity index determined the top 10 variables of importance (VOI). APMs in the top 10 VOI were assessed for confounding effects by extracapsular extension (ECE) and pathologic T (pT) through Poisson regression with Generalized Estimating Equation. Results: 55/236 (23.3%) cases had PSMs. Of the 55 cases with PSMs, 9 (16.4%) were pT2 and 46 (83.6%), pT3. The full model, including clinical factors and APMs, achieved area under the curve (AUC) 0.74. When assessing clinical factors or APMs alone, the model achieved AUC 0.72 and 0.64, respectively. The strongest PSM predictors were ECE and pT stage, followed by APMs in specific steps. After adjusting for ECE and pT stage, most APMs remained as independent predictors of PSM. Conclusion: Using machine learning methods, we found that the strongest predictors of PSMs after RARP are nonmodifiable, disease-driven factors (ECE and pT). While APMs provide minimal additional insight into when PSMs may occur, they are nonetheless capable of independently predicting PSMs based on objective measures of surgeon performance.

A Conformational Variant of p53 (U-p53AZ) as Blood-Based Biomarker for the Prediction of the Onset of Symptomatic Alzheimer’s Disease

BackgroundOngoing research seeks to identify blood-based biomarkers able to predict onset and progression of Alzheimer’s disease (AD).ObjectiveThe unfolded conformational variant of p53 (U-p53AZ), previously observed in AD individuals, was evaluated in plasma samples from individuals participating in the Australian Imaging, Biomarkers and Lifestyle (AIBL) cohort for diagnostic and prognostic assessment, validated on a neuropsychological-based diagnosis, over the course of six years.DesignRetrospective Longitudinal Prognostic biomarker study.SettingSingle-center study based on the AIBL cohort.Participants482 participants of the AIBL cohort, aged 60–85 years, without uncontrolled diabetes, vascular disease, severe depression or psychiatric illnesses.MeasurementsThe AlzoSure® Predict test, consisting of immunoprecipitation (IP) followed by liquid chromatography (LC) tandem mass spectrometry (MS/MS), was performed to quantify the AZ 284® peptide as readout of U-p53AZ and compared with an independent neuropsychological diagnosis. The amyloid load via amyloid β-positron emission tomography (Aβ-PET) and supporting clinical information were included where possible.ResultsU-p53AZ diagnostic and prognostic performance was assessed in both time-independent and time-dependent (36, 72 and 90 months following initial sampling) analyses. Prognostic performance of Aβ-PET and survival analyses with different risk factors (gender, Aβ-PET and APOE ε4 allele status) were also performed. U-p53AZ differentiated neuropsychologically graded AD from non-AD samples, and its detection at intermediate/high levels precisely identified present and future symptomatic AD. In both time-independent and time-dependent prognostic analyses U-p53AZ achieved area under the curve (AUC) >98%, significantly higher than Aβ-PET AUCs (between 84% and 93%, P respectively <0.0001 and <0.001). As single factor, U-p53AZ could clearly determine the risk of AD neuropsychological diagnosis over time (low versus intermediate/high U-p53AZ hazard ratio=2.99). Proportional hazards regression analysis identified U-p53AZ levels as a major independent predictor of AD onset.ConclusionsThese findings support use of U-p53AZ as blood-based biomarker predicting whether individuals would reach neuropsychologically-defined AD within six years prior to AD diagnosis. Integration of U-p53AZ in screening processes could support refined participant stratification for interventional studies.

Deep Learning for Glaucoma Detection and Identification of Novel Diagnostic Areas in Diverse Real-World Datasets.

PurposeTo develop a three-dimensional (3D) deep learning algorithm to detect glaucoma using spectral-domain optical coherence tomography (SD-OCT) optic nerve head (ONH) cube scans and validate its performance on ethnically diverse real-world datasets and on cropped ONH scans.MethodsIn total, 2461 Cirrus SD-OCT ONH scans of 1012 eyes were obtained from the Glaucoma Clinic Imaging Database at the Byers Eye Institute, Stanford University, from March 2010 to December 2017. A 3D deep neural network was trained and tested on this unique raw OCT cube dataset to identify a multimodal definition of glaucoma excluding other concomitant retinal disease and optic neuropathies. A total of 1022 scans of 363 glaucomatous eyes (207 patients) and 542 scans of 291 normal eyes (167 patients) from Stanford were included in training, and 142 scans of 48 glaucomatous eyes (27 patients) and 61 scans of 39 normal eyes (23 patients) were included in the validation set. A total of 3371 scans (Cirrus SD-OCT) from four different countries were used for evaluation of the model: the non overlapping test dataset from Stanford (USA) consisted of 694 scans: 241 scans from 113 normal eyes of 66 patients and 453 scans of 157 glaucomatous eyes of 89 patients. The datasets from Hong Kong (total of 1625 scans; 666 OCT scans from 196 normal eyes of 99 patients and 959 scans of 277 glaucomatous eyes of 155 patients), India (total of 672 scans; 211 scans from 147 normal eyes of 98 patients and 461 scans from 171 glaucomatous eyes of 101 patients), and Nepal (total of 380 scans; 158 scans from 143 normal eyes of 89 patients and 222 scans from 174 glaucomatous eyes of 109 patients) were used for external evaluation. The performance of the model was then evaluated on manually cropped scans from Stanford using a new algorithm called DiagFind. The ONH region was cropped by identifying the appropriate zone of the image in the expected location relative to Bruch's Membrane Opening (BMO) using a commercially available imaging software. Subgroup analyses were performed in groups stratified by eyes, myopia severity of glaucoma, and on a set of glaucoma cases without field defects. Saliency maps were generated to highlight the areas the model used to make a prediction. The model's performance was compared to that of a glaucoma specialist using all available information on a subset of cases.ResultsThe 3D deep learning system achieved area under the curve (AUC) values of 0.91 (95% CI, 0.90–0.92), 0.80 (95% CI, 0.78–0.82), 0.94 (95% CI, 0.93–0.96), and 0.87 (95% CI, 0.85–0.90) on Stanford, Hong Kong, India, and Nepal datasets, respectively, to detect perimetric glaucoma and AUC values of 0.99 (95% CI, 0.97–1.00), 0.96 (95% CI, 0.93–1.00), and 0.92 (95% CI, 0.89–0.95) on severe, moderate, and mild myopia cases, respectively, and an AUC of 0.77 on cropped scans. The model achieved an AUC value of 0.92 (95% CI, 0.90–0.93) versus that of the human grader with an AUC value of 0.91 on the same subset of scans (). The performance of the model in terms of recall on glaucoma cases without field defects was found to be 0.76 (0.68–0.85). Saliency maps highlighted the lamina cribrosa in glaucomatous eyes versus superficial retina in normal eyes as the regions associated with classification.ConclusionsA 3D convolutional neural network (CNN) trained on SD-OCT ONH cubes can distinguish glaucoma from normal cases in diverse datasets obtained from four different countries. The model trained on additional random cropping data augmentation performed reasonably on manually cropped scans, indicating the importance of lamina cribrosa in glaucoma detection.Translational RelevanceA 3D CNN trained on SD-OCT ONH cubes was developed to detect glaucoma in diverse datasets obtained from four different countries and on cropped scans. The model identified lamina cribrosa as the region associated with glaucoma detection.

Convolutional neural networks for detection of transthyretin amyloidosis in 2D scintigraphy images

BackgroundTransthyretin amyloidosis (ATTR) is a progressive disease which can be diagnosed non-invasively using bone avid [99mTc]-labeled radiotracers. Thus, ATTR is also an occasional incidental finding on bone scintigraphy. In this study, we trained convolutional neural networks (CNN) to automatically detect and classify ATTR from scintigraphy images. The study population consisted of 1334 patients who underwent [99mTc]-labeled hydroxymethylene diphosphonate (HMDP) scintigraphy and were visually graded using Perugini grades (grades 0–3). A total of 47 patients had visual grade ≥ 2 which was considered positive for ATTR. Two custom-made CNN architectures were trained to discriminate between the four Perugini grades of cardiac uptake. The classification performance was compared to four state-of-the-art CNN models.ResultsOur CNN models performed better than, or equally well as, the state-of-the-art models in detection and classification of cardiac uptake. Both models achieved area under the curve (AUC) ≥ 0.85 in the four-class Perugini grade classification. Accuracy was good in detection of negative vs. positive ATTR patients (grade < 2 vs grade ≥ 2, AUC > 0.88) and high-grade cardiac uptake vs. other patients (grade < 3 vs. grade 3, AUC = 0.94). Maximum activation maps demonstrated that the automated deep learning models were focused on detecting the myocardium and not extracardiac features.ConclusionAutomated convolutional neural networks can accurately detect and classify different grades of cardiac uptake on bone scintigraphy. The CNN models are focused on clinically relevant image features. Automated screening of bone scintigraphy images using CNN could improve the early diagnosis of ATTR.

Computed Tomography Imaging-Based Radiogenomics Analysis Reveals Hypoxia Patterns and Immunological Characteristics in Ovarian Cancer.

PurposeThe hypoxic microenvironment is involved in the tumorigenesis of ovarian cancer (OC). Therefore, we aim to develop a non-invasive radiogenomics approach to identify a hypoxia pattern with potential application in patient prognostication.MethodsSpecific hypoxia-related genes (sHRGs) were identified based on RNA-seq of OC cell lines cultured with different oxygen conditions. Meanwhile, multiple hypoxia-related subtypes were identified by unsupervised consensus analysis and LASSO–Cox regression analysis. Subsequently, diversified bioinformatics algorithms were used to explore the immune microenvironment, prognosis, biological pathway alteration, and drug sensitivity among different subtypes. Finally, optimal radiogenomics biomarkers for predicting the risk status of patients were developed by machine learning algorithms.ResultsOne hundred forty sHRGs and three types of hypoxia-related subtypes were identified. Among them, hypoxia-cluster-B, gene-cluster-B, and high-risk subtypes had poor survival outcomes. The subtypes were closely related to each other, and hypoxia-cluster-B and gene-cluster-B had higher hypoxia risk scores. Notably, the low-risk subtype had an active immune microenvironment and may benefit from immunotherapy. Finally, a four-feature radiogenomics model was constructed to reveal hypoxia risk status, and the model achieved area under the curve (AUC) values of 0.900 and 0.703 for the training and testing cohorts, respectively.ConclusionAs a non-invasive approach, computed tomography-based radiogenomics biomarkers may enable the pretreatment prediction of the hypoxia pattern, prognosis, therapeutic effect, and immune microenvironment in patients with OC.

Radiomics and Qualitative Features From Multiparametric MRI Predict Molecular Subtypes in Patients With Lower-Grade Glioma.

BackgroundIsocitrate dehydrogenase (IDH) mutation and 1p19q codeletion status have been identified as significant markers for therapy and prognosis in lower-grade glioma (LGG). The current study aimed to construct a combined machine learning-based model for predicting the molecular subtypes of LGG, including (1) IDH wild-type astrocytoma (IDHwt), (2) IDH mutant and 1p19q non-codeleted astrocytoma (IDHmut-noncodel), and (3) IDH-mutant and 1p19q codeleted oligodendroglioma (IDHmut-codel), based on multiparametric magnetic resonance imaging (MRI) radiomics, qualitative features, and clinical factors.MethodsA total of 335 patients with LGG (WHO grade II/III) were retrospectively enrolled. The sum of 5,929 radiomics features were extracted from multiparametric MRI. Selected robust, non-redundant, and relevant features were used to construct a random forest model based on a training cohort (n = 269) and evaluated on a testing cohort (n = 66). Meanwhile, preoperative MRIs of all patients were scored in accordance with Visually Accessible Rembrandt Images (VASARI) annotations and T2-fluid attenuated inversion recovery (T2-FLAIR) mismatch sign. By combining radiomics features, qualitative features (VASARI annotations and T2-FLAIR mismatch signs), and clinical factors, a combined prediction model for the molecular subtypes of LGG was built.ResultsThe 17-feature radiomics model achieved area under the curve (AUC) values of 0.6557, 0.6830, and 0.7579 for IDHwt, IDHmut-noncodel, and IDHmut-codel, respectively, in the testing cohort. Incorporating qualitative features and clinical factors into the radiomics model resulted in improved AUCs of 0.8623, 0.8056, and 0.8036 for IDHwt, IDHmut-noncodel, and IDHmut-codel, with balanced accuracies of 0.8924, 0.8066, and 0.8095, respectively.ConclusionThe combined machine learning algorithm can provide a method to non-invasively predict the molecular subtypes of LGG preoperatively with excellent predictive performance.

Predicting prognosis of primary pontine hemorrhage using CT image and deep learning

Prognosis of primary pontine hemorrhage (PPH) is important for treatment planning and patient management. However, only few clinical factors were reported to have prognostic value to PPH. Here, we propose a deep learning (DL) model that mines high-dimensional prognostic information from computed tomography (CT) images and combines clinical factors for predicting individualized prognosis of PPH. We proposed a multi-task DL model to learn high-dimensional CT features of hematoma and perihematomal areas for predicting the risk of 30-day mortality, 90-day mortality and 90-day functional outcome of PPH simultaneously. We further explored the combination of the DL model and clinical factors by building a combined model. All the models were trained in a training cohort (n = 219) and tested in an independent testing cohort (n = 35). The DL model achieved area under the curve (AUC) of 0.886, 0.886, and 0.759 in predicting 30-day mortality, 90-day mortality and 90-day functional outcome of PPH in the independent testing cohort, which improved over the previously reported new PPH score and the clinical model. When combining the DL model and clinical factors, the combined model achieved improved performance (AUC = 0.920, 0.941, and 0.894), indicating that DL model mines CT information that complements clinical factors. Through DL visualization technique, we found that the internal structure of hematoma and its expansion to perihematomal regions are important for predicting the prognosis of PPH. This DL model provides an easy-to-use way for predicting individualized prognosis of PPH by mining high-dimensional information from CT images, and showed improvement over clinical factors and present methods.

Machine Learning-Based Radiomics of the Optic Chiasm Predict Visual Outcome Following Pituitary Adenoma Surgery.

Preoperative prediction of visual recovery after pituitary adenoma surgery remains a challenge. We aimed to investigate the value of MRI-based radiomics of the optic chiasm in predicting postoperative visual field outcome using machine learning technology. A total of 131 pituitary adenoma patients were retrospectively enrolled and divided into the recovery group (N = 79) and the non-recovery group (N = 52) according to visual field outcome following surgical chiasmal decompression. Radiomic features were extracted from the optic chiasm on preoperative coronal T2-weighted imaging. Least absolute shrinkage and selection operator regression were first used to select optimal features. Then, three machine learning algorithms were employed to develop radiomic models to predict visual recovery, including support vector machine (SVM), random forest and linear discriminant analysis. The prognostic performances of models were evaluated via five-fold cross-validation. The results showed that radiomic models using different machine learning algorithms all achieved area under the curve (AUC) over 0.750. The SVM-based model represented the best predictive performance for visual field recovery, with the highest AUC of 0.824. In conclusion, machine learning-based radiomics of the optic chiasm on routine MR imaging could potentially serve as a novel approach to preoperatively predict visual recovery and allow personalized counseling for individual pituitary adenoma patients.