Explainable Colon Cancer Stage Prediction with Multimodal Biodata through the Attention-based Transformer and Squeeze-Excitation Framework

e12597 Background: Early-stage HER2-positive breast cancer patients who achieve pathological complete response (pCR) after neoadjuvant therapy generally experience favorable survival outcomes. However, only 40-60% of HER2-positive patients achieve pCR. This study developed a deep learning model using histopathology images to predict neoadjuvant therapy response in HER2-positive breast cancer across different regimens, aiming to guide personalized treatment choices. Methods: In this multi-centered retrospective study, we recruited 402 HER2-positive breast cancer patients from four hospitals: 320 patients from two centers were divided into the discovery cohort (n=223) and internal testing cohort (n=97), while 82 patients from the other two centers served as external validation cohorts (n=21 and n=61). We collected pre-treatment H&E-stained histopathology images, clinical information, neoadjuvant treatment regimens, and established a pCR prediction model, named Histomics-Clinic-Regimen Integrated pCR Prediction Model (HIPPM), which integrates CAMEL2 and FCNN approaches. Model performance was evaluated using area under the curve (AUC), sensitivity (SE), specificity (SP), and accuracy. Additionally, we utilized the weight matrix from the first fully connected layer to assess the relative importance of each clinical variable. Results: The initial model based solely on histopathology images demonstrated the following performance: internal testing cohort (AUC 0.71, SE 0.73, SP 0.70, accuracy 70.6%), external validation cohort 1 (AUC 0.95, SE 1.00, SP 0.94, accuracy 95.2%), and validation cohort 2 (AUC 0.61, SE 0.37, SP 0.85, accuracy 53.3%). After integrating clinical and regimen data into the HIPPM model, performance improved significantly: internal testing cohort (AUC 0.85, SE 0.73, SP 0.91, accuracy 88.2%), external validation cohort 1 (AUC 0.94, SE 1.00, SP 0.94, accuracy 95.2%), and validation cohort 2 (AUC 0.74, SE 0.82, SP 0.67, accuracy 73.3%). Model analysis revealed that T stage, HER2 expression, N stage, and targeted therapy regimen had the highest weights, while Ki67 expression, age, ER expression, and chemotherapy regimen had lower weights. Based on HIPPM, we developed a predictive tool that allows patients to upload biopsy images and clinical information pre-treatment to predict pCR probability for 12 virtual regimens and recommend the optimal drug combination. Conclusions: HIPPM effectively predicts neoadjuvant therapy response in HER2-positive breast cancer, aiding in selecting optimal targeted and chemotherapeutic regimens. This model lays the foundation for patient screening and personalized treatment strategies.

Integration of Deep Learning and Sub-regional Radiomics Improves the Prediction of Pathological Complete Response to Neoadjuvant Chemoradiotherapy in Locally Advanced Rectal Cancer Patients.

BackgroundColon cancer (CC) refers to malignant tumor of the digestive tract worldwide and is also among the cancers with high mortality rates.ObjectiveThe aim of this work was to evaluate the diagnostic performance of multislice spiral CT (MSCT), magnetic resonance imaging (MRI), and MSCT + MRI in different stages of colon cancer (CC) (T1-T2, T3, T4). This work compared the differences in sensitivity (Sen), specificity (Spe), accuracy (Acc), and area under the curve (AUC) values among these methods and explored the optimal diagnostic strategy.MethodsA total of 120 patients with CC confirmed by pathological biopsy and 30 individuals suspected of CC but without detected tumors (as controls) were selected. All subjects underwent MSCT, MRI, and combined MSCT + MRI examinations. Statistical analyses of Sen, Spe, Acc, and AUC values were performed.ResultsIn the T1-T2 stage, MSCT had a Sen of 85.2%, Acc of 86.8%, and an AUC value of 0.878; MRI had a Spe of 91.0%, Sen of 81.6%, and an AUC value of 0.865; the combined MSCT + MRI examination had a Sen of 90.6% and an AUC of 0.903. In the T3 stage, MRI had a significantly higher Sen (91.7%) than MSCT (80.0%), with an AUC of 0.887, while the combined MSCT + MRI examination had a Sen of 98.3% and an AUC of 0.942. In the T4 stage, the combined MSCT + MRI examination performed the best, with a Sen of 100% and an AUC of 0.933, and compared with MSCT or MRI alone, the differences were statistically significant (P < 0.05).ConclusionMSCT and MRI each have their own advantages in the diagnosis of different stages of CC. MSCT is suitable for initial screening in the T1-T2 stage, while MRI is more effective in assessing tumor invasiveness in the T3 and higher stages. The combined MSCT + MRI examination can provide more comprehensive diagnostic information, especially in the T4 stage, where it shows the highest Sen and Acc. Selecting the appropriate examination method based on the patient's specific condition and staging needs is of great significance in improving the diagnostic Acc of CC.

The early detection of cancer stage is a crucial step for effective treatment. In contrast to traditional approaches, RNA -Seq is the current state of the art technique for gene expression estimation. RNA -Seq data have been used in research and in production as input data for several classification and prediction models in many disease including cancer staging. We present a novel cancer stage prediction approach based on gene expression data. Our approach is based on Weighted Graph Convolution Networks (GCN). GCN is the application of deep learning back-propagation on graph structures. In this work, we used correlation between genes to generate a gene network graph. A neural network model with weighted graph convolution layer was trained to predict the cancer stage for cancer patients. We employed the Kidney Renal Clear Cell Carcinoma dataset (TCGA-KIRC) from the Human Cancer Genome Atlas (TCGA) to predict the cancer stage for each patient. The TCGA-KIRC dataset includes 4 cancer stages, I, II, III, and IV. We generated a binary classification problem where stages I and II are classified as “early cancer stage” and stages III and IV are classified as “late cancer stage”. We compared our approach to the state of the art approaches such as random forest and support vector machine. Our approach achieved an accuracy of 82% which outperformed existing approaches with more than 3% increase.

To compare performance of a deep-learning enhanced algorithm for automated detection of diabetic retinopathy (DR), to the previously published performance of that algorithm, the Iowa Detection Program (IDP)-without deep learning components-on the same publicly available set of fundus images and previously reported consensus reference standard set, by three US Board certified retinal specialists. We used the previously reported consensus reference standard of referable DR (rDR), defined as International Clinical Classification of Diabetic Retinopathy moderate, severe nonproliferative (NPDR), proliferative DR, and/or macular edema (ME). Neither Messidor-2 images, nor the three retinal specialists setting the Messidor-2 reference standard were used for training IDx-DR version X2.1. Sensitivity, specificity, negative predictive value, area under the curve (AUC), and their confidence intervals (CIs) were calculated. Sensitivity was 96.8% (95% CI: 93.3%-98.8%), specificity was 87.0% (95% CI: 84.2%-89.4%), with 6/874 false negatives, resulting in a negative predictive value of 99.0% (95% CI: 97.8%-99.6%). No cases of severe NPDR, PDR, or ME were missed. The AUC was 0.980 (95% CI: 0.968-0.992). Sensitivity was not statistically different from published IDP sensitivity, which had a CI of 94.4% to 99.3%, but specificity was significantly better than the published IDP specificity CI of 55.7% to 63.0%. A deep-learning enhanced algorithm for the automated detection of DR, achieves significantly better performance than a previously reported, otherwise essentially identical, algorithm that does not employ deep learning. Deep learning enhanced algorithms have the potential to improve the efficiency of DR screening, and thereby to prevent visual loss and blindness from this devastating disease.

Recently, researchers have been studying methods to introduce deep learning into automated optical inspection (AOI) systems to reduce labor costs. However, the integration of deep learning in the industry may encounter major challenges such as sample imbalance (defective products that only account for a small proportion). Therefore, in this study, an anomaly detection neural network, dual auto-encoder generative adversarial network (DAGAN), was developed to solve the problem of sample imbalance. With skip-connection and dual auto-encoder architecture, the proposed method exhibited excellent image reconstruction ability and training stability. Three datasets, namely public industrial detection training set, MVTec AD, with mobile phone screen glass and wood defect detection datasets, were used to verify the inspection ability of DAGAN. In addition, training with a limited amount of data was proposed to verify its detection ability. The results demonstrated that the areas under the curve (AUCs) of DAGAN were better than previous generative adversarial network-based anomaly detection models in 13 out of 17 categories in these datasets, especially in categories with high variability or noise. The maximum AUC improvement was 0.250 (toothbrush). Moreover, the proposed method exhibited better detection ability than the U-Net auto-encoder, which indicates the function of discriminator in this application. Furthermore, the proposed method had a high level of AUCs when using only a small amount of training data. DAGAN can significantly reduce the time and cost of collecting and labeling data when it is applied to industrial detection.

e12533 Background: Temporal radiomic features (TRF) extracted from dynamic contrast-enhanced breast MR (DCE-MR), which provide important information about tumor heterogeneity, offer a non-invasive approach to predict high-risk groups and could potentially serve as a cost-effective alternative to the genetic assay. This study explored how TRF from breast MR could be integrated to predict the high-risk group of OncotypeDX (ODX). Methods: In 173 patients with breast cancer [low-risk, 144 (83.2%); high-risk, 29 (16.8%)], TRF such as dynamic signal intensity changes and texture variations were derived from the imaging sequences. Hierarchical clustering was applied to reduce feature redundancy and identify significant predictors. Machine learning algorithms such as random forest, SVM, logistic regression, and KNN were utilized with 7-fold cross validation model. Results: Cross-validation revealed that models with TRF were consistently better than those without it across 4 different machine learning algorithms. Random forest AUC improved from 0.48 to 0.56, support vector machines from 0.46 to 0.63, logistic regression from 0.51 to 0.69, and K-nearest neighbors from 0.61 to 0.73 (Table). Conclusions: Incorporating TRF from DCE-MR images into a machine learning model improved the predictive performance for high-risk groups of ODX compared to using only the reference RF. Comparison of classifier performance using radiomic and temporal radiomic features. Random forest SVM Logistic regression KNN Features RF RF + TRF RF RF + TRF RF RF + TRF RF TRF AUC Fold1 0.38 (0.13, 0.65) 0.41 (0.14, 0.69) 0.68 (0.38, 0.94) 0.54 (0.23, 0.84) 0.39 (0.12, 0.71) 0.50 (0.18, 0.85) 0.60 (0.22, 0.90) 0.55 (0.27, 0.82) AUC Fold2 0.62 (0.35, 0.86) 0.52 (0.24, 0.85) 0.71 (0.45, 0.92) 0.65 (0.40, 0.87) 0.31 (0.09, 0.57) 0.55 (0.28, 0.84) 0.54 (0.29, 0.78) 0.69 (0.42, 0.91) AUC Fold3 0.57 (0.30, 0.82) 0.68 (0.38, 0.93) 0.46 (0.10, 0.75) 0.50 (0.12, 0.86) 0.49 (0.13, 0.83) 0.76 (0.53, 0.96) 0.43 (0.00, 0.82) 0.70 (0.48, 0.90) AUC Fold4 1.00 (1.00, 1.00) 0.83 (0.67, 0.96) 0.08 (0.00, 0.21) 0.79 (0.62, 0.96) 1.00 (1.00, 1.00) 1.00 (1.00, 1.00) 0.77 (0.61, 0.90) 0.75 (0.56, 0.92) AUC Fold5 0.23 (0.00, 0.50) 0.30 (0.00, 0.63) 0.83 (0.62, 0.98) 0.44 (0.13, 0.83) 0.36 (0.08, 0.62) 0.44 (0.13, 0.82) 0.62 (0.27, 0.89) 0.66 (0.46, 0.83) AUC Fold6 0.23 (0.00, 0.50) 0.66 (0.39, 0.89) 0.35 (0.00, 0.78) 0.76 (0.57, 0.95) 0.70 (0.44, 0.91) 0.70 (0.37, 1.00) 0.62 (0.13, 0.95) 0.85 (0.62, 1.00) AUC Fold7 0.35 (0.08, 0.67) 0.48 (0.14, 0.91) 0.11 (0.00, 0.32) 0.75 (0.53, 0.91) 0.33 (0.08, 0.67) 0.86 (0.68, 1.00) 0.70 (0.45, 0.90) 0.90 (0.77, 1.00) AUC Average 0.48 0.56 0.46 0.63 0.51 0.69 0.61 0.73 Data were expressed as Area Under the Curve (AUC) and 95% Confidence Intervals (CI). RF, Radiomic features; TRF, Temporal radiomic features; SVM, Support Vector Machine; KNN, K-Nearest Neighbors, AUC, Area Under the ROC Curve.

ABSTRACTBackgroundMost T1 gastric cancer (GC) harbor lymph node metastasis (LNM) at a rate of < 20%; however, owing to the difficulty in accurately diagnosing LNM preoperatively, many patients with T1 GC undergo unnecessary invasive radical gastrectomy with lymphadenectomy. In the present study, we established an epigenetic liquid biopsy assay for the preoperative diagnosis of LNM in T1 GC.MethodsA comprehensive biomarker discovery was performed by analyzing genome‐wide DNA methylation profiling. We obtained 277 clinical specimens, including 177 surgical tissues and 100 pre‐operative plasmas. DNA methylation biomarkers were trained and validated using quantitative methylation‐specific polymerase chain reaction (qMSP) assays.ResultsWe identified six novel differentially methylated regions, including at least two differentially methylated CpG probes (|Delta‐beta| > 0.12 and p < 0.05) within 100 bp, through genome‐wide biomarker discovery. A DNA methylation panel was generated using qMSP assays in clinical tissue specimens, with an area under the curve (AUC) of 0.80. This panel was validated in an independent clinical cohort, and a combined model, which integrated the DNA methylation model with preoperative computed tomography ‐based findings, was established through multivariate logistic regression analyses (AUC: 0.84). Finally, we translated this model into a liquid biopsy, and this cell‐free DNA (cfDNA) methylation model exhibited robust performance for LNM identification in T1 GC (AUC: 0.86) and allowed 44% of patients to avoid unnecessary invasive operations, without missing any LNM‐positive patients.ConclusionsWe have successfully developed a cfDNA methylation signature‐based liquid biopsy diagnostic assay that allows for robust and less‐invasive LNM detection in patients with T1 GC.

Traditional cardiovascular (CV) biomarkers (high-sensitivity troponinT [hsTnT] and N-terminal pro-B-type natriuretic peptide [NT-proBNP]) are important to monitor cancer patients' cardiac function and to assess prognosis. Newer CV biomarkers (mid-regional pro-adrenomedullin [MR-proADM], C-terminal pro-arginine vasopressin [copeptin], and mid-regional pro-atrial natriuretic peptide [MR-proANP]) might outperform traditional biomarkers. Overall, 442 hospitalized cancer patients without significant CV disease or current infection were enrolled (61 ± 15 years, 52% male, advanced cancer stage: 85%) and concentrations of CV biomarkers were analysed. Differences in echocardiographic, clinical, laboratory parameters were assessed. Patients were followed for up to 69 months for all-cause mortality. In univariable analyses, MR-proADM, hsTnT, copeptin, MR-proANP, and NT-proBNP predicted all-cause mortality. In multivariable analyses (adjusted for sex, age, Eastern Cooperative Oncology Group performance status, estimated glomerular filtration rate [eGFR], C-reactive protein, anti-cancer therapy, reason for hospitalization, cancer stage and type), only MR-proADM remained an independent predictor of mortality (MR-proADM per 1 ln: hazard ratio [HR] 2.27, 95% confidence interval [CI] 1.47-3.50], p < 0.001). MR-proADM had the highest area under the curve (AUC) using receiver operating characteristic analysis (AUC [95% CI] 0.74 [0.69-0.79]; hsTnT: AUC 0.69; copeptin: AUC 0.66; MR-proANP: AUC 0.63; NT-proBNP: AUC 0.62). Optimal cut-point for mortality prediction with MR-proADM was 0.94 nmol/L (HR 2.43 [95% CI 1.92-3.06], p < 0.001). Patients with MR-proADM >0.94 nmol/L were older, more often had cancer stage IV, showed reduced performance status, eGFR, haemoglobin, diastolic left ventricular function, and elevated systolic pulmonary artery pressure. MR-proADM is an independent predictor of mortality in advanced stage, hospitalized cancer patients without significant CV disease or current infection. The optimal MR-proADM cut-point for mortality prediction was 0.94 nmol/L with hazards for mortality being approximately 2.5 times higher. There was a continuous increase in mortality risk with stepwise increase of MR-proADM concentrations. Elevated concentrations of MR-proADM were also associated with reduced performance status and mildly reduced left ventricular diastolic function as well as higher age and more often cancer stage IV.

DNA methylation is a key epigenetic alteration in tumorigenesis, but its diagnostic value in early-stage lung cancer remains unclear. In this study, tissue and plasma samples from patients with lung cancer or benignity were analyzed. Methylation profiles were obtained using bisulfite sequencing and compared with selected lung cancer-specific markers. Diagnostic prediction models were constructed using these markers, with their efficacy assessed by sensitivity, specificity and area under the curve (AUC). In the tissue cohort, 276 markers were found to be significantly differentially methylated in lung cancer (FDR < 0.05). A diagnostic prediction model using six markers showed promising performance in both the training cohort (sensitivity = 90%; specificity = 97%; AUC = 0.988) and the validation cohort (sensitivity = 92%; specificity = 94%; AUC = 0.977). In the plasma cohort, a diagnostic prediction model using nine markers achieved a sensitivity of 98% and specificity of 100% (AUC = 0.998) in the training cohort, a sensitivity of 81% and specificity of 59% (AUC = 0.791) in the validation cohort. Furthermore, we observed a significant correlation between delta methylation changes in tissue and plasma in the paired patient cohort. Additional analysis based on methylation haplotypes identified 1222 differentially methylated regions in tissue samples, mainly enriched in DNA replication-related pathways. Additionally, correlations between DNA methylation and clinical characteristics revealed significant differential methylation patterns between smokers and non-smokers . Thus, DNA methylation in both tissue and plasma holds potential as a biomarker for the early diagnosis of lung cancer.

Purpose: To predict pathologic complete response (pCR) in breast cancer patients undergoing neoadjuvant chemotherapy (NAC), from baseline and early-treatment DCE-MRI scans, in the context of the ACRIN 6698/I-SPY 2 BMMR2 challenge. Materials and Methods: The BMMR2 dataset consists of 191 patients undergoing NAC for locally advanced breast cancer as part of the ACRIN 6698/I-SPY 2 trial. DCE-MRI was obtained at time points T0 (pre-NAC), T1 (3 weeks), and T2 (12 weeks). The BMMR2 challenge provided the MRI scans, tumor annotations, and limited clinical and demographic information. The data were split 60/40; using the 60% training data, the task was to develop models to predict pCR; the competition was for best area under the curve (AUC) when applied to the 40% unseen test data. Using the publicly available CaPTk software we calculated 3 types of radiomic features within the segmented tumor volume: 1) texture of the signal enhancement ratio (SER) kinetic map of T0 images; 2) texture of the difference between the T1 kinetic maps (PE, WIS, WOS, and SER) and corresponding T0 maps; 3) texture of the difference between the T1 kinetic maps and the corresponding T0 maps, with T1 scans deformably registered to T0 scans. ComBat harmonization was applied to the extracted features to account for MRI acquisition differences. We computed the tumor longest diameter, functional tumor volume (FTV), and clinical tumor size each at T0 and T1. We modeled pCR via logistic regression. Using the training data alone, with the criteria of performance in univariable modeling and low collinearity, we selected radiomic features and clinical, demographic, and size covariates. We then performed PCA on the combined set of selected radiomic features and size covariates. We evaluated multivariable models including the selected clinical covariates in combination with the first few PCs via cross-validated AUC (5-fold, 200 repetitions) on the training data. The best models were submitted for independent evaluation on the unseen test data of the BMMR2 challenge. Results: Of the available clinical covariates, only hormone receptor (HR)± and human epidermal growth factor receptor 2 (HER2)± had any association with pCR. We retained these in all models, and performed PCA on the set combining the best-performing features and the size variables FTV at T0, FTV at T1, and longest diameter at T1. Models based on the first few PCs, HR, and HER2, had training AUCs in 0.78–0.81. Our best-performing model had an AUC on test data of 0.84, using the covariates PCs 1–5, HR, and HER2 (Table 1). Conclusions: Our preliminary results suggest that radiomic phenotyping of changes in tumor heterogeneity can accurately predict pCR early in the course of NAC. Future analysis with larger samples from ISPY-2 could also examine the effect of different therapies, including targeted therapy and immunotherapy. Table 1: Performance of candidate logistic regression models on training and test data. AUC: Area under receiver operating characteristic curve. * Mean 5-fold cross-validated AUC across 200 replicates. † Competition best-performing predictions. Citation Format: Eric A. Cohen, Rhea D. Chitalia, Snekha Thakran, Walter C. Mankowski, Alex Anh-Tu Nguyen, Hannah Horng, Elizabeth S. McDonald, Michael Feldman, Angela DeMichele, Despina Kontos. Title: Characterizing Changes in Tumor Heterogeneity via Radiomic Phenotyping for Predicting Response to Neoadjuvant Chemotherapy for Locally Advanced Breast Cancer: Preliminary Results from the ACRIN 6698/I-SPY 2 trial [abstract]. In: Proceedings of the 2022 San Antonio Breast Cancer Symposium; 2022 Dec 6-10; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2023;83(5 Suppl):Abstract nr PD16-08.

BackgroundPretreatment prediction of stage in patients with cervical cancer (CC) is vital for tailoring treatment strategy. This study aimed to explore the feasibility of a model combining reduced field-of-view (rFOV) diffusion-weighted imaging (DWI)-derived radiomics with clinical features in staging CC.MethodsPatients with pathologically proven CC were enrolled in this retrospective study. The rFOV DWI with b values of 0 and 800 s/mm2 was acquired and the clinical characteristics of each patient were collected. Radiomics features were extracted from the apparent diffusion coefficient maps and key features were selected subsequently. A clinical–radiomics model combining radiomics with clinical features was constructed. The receiver operating characteristic curve was introduced to evaluate the predictive efficacy of the model, followed by comparisons with the MR-based subjective stage assessment (radiological model).ResultsNinety-four patients were analyzed and divided into training (n = 61) and testing (n = 33) cohorts. In the training cohort, the area under the curve (AUC) of clinical–radiomics model (AUC = 0.877) for staging CC was similar to that of radiomics model (AUC = 0.867), but significantly higher than that of clinical model (AUC = 0.673). In the testing cohort, the clinical–radiomics model yielded the highest predictive performance (AUC = 0.887) of staging CC, even without a statistically significant difference when compared with the clinical model (AUC = 0.793), radiomics model (AUC = 0.846), or radiological model (AUC = 0.823).ConclusionsThe rFOV DWI-derived clinical–radiomics model has the potential for staging CC, thereby facilitating clinical decision-making.

Juvenile Idiopathic Arthritis (JIA) frequently affects children's hips, causing severe progression, but early hip synovitis lacks obvious symptoms and is hard to detect via conventional ultrasound, delaying diagnosis. magnetic resonance imaging (MRI), though accurate, is costly and inaccessible for routine use. This study aims to develop an automatic identification system for the early diagnosis of hip synovitis in JIA through the integration of deep learning and radiomics techniques. A YOLO-JIA model specifically designed for the automatic segmentation of hip ultrasound images was developed. Radiomic features were extracted from these segmented regions. Subsequently, feature selection was performed using the analysis of variance (ANOVA) test followed by least absolute shrinkage and selection operator (LASSO) regression. Based on the selected features, a Random Forest (RF) classification model was constructed and evaluated separately on an internal and an external validation set. The YOLO-JIA model demonstrated high precision (0.98) and recall (1.00) in object detection tasks, with a mean average precision at 50-95% (mAP50-95) for mask (M) reaching 0.86. The RF classification model achieved an area under the curve (AUC) of 0.88 on the internal validation set and 0.81 on the external validation set. Decision curve analysis further confirmed the clinical utility of our proposed system. Finally, the models were integrated and deployed locally. This study successfully developed a system for the early diagnosis of JIA hip synovitis based on deep learning and radiomics. The system offers an effective and reliable means for early screening, enhancing diagnosis rates, and ultimately reducing the risk of severe joint damage in JIA patients.

This study presents a robust approach for the classification of ovarian cancer subtypes through the integration of deep learning and k-nearest neighbor (KNN) methods. The proposed model leverages the powerful feature extraction capabilities of EfficientNet-B0, utilizing its deep features for subsequent fine-grained classification using the fine-KNN approach. The UBC-OCEAN dataset, encompassing histopathological images of five distinct ovarian cancer subtypes, namely, high-grade serous carcinoma (HGSC), clear-cell ovarian carcinoma (CC), endometrioid carcinoma (EC), low-grade serous carcinoma (LGSC), and mucinous carcinoma (MC), served as the foundation for our investigation. With a dataset comprising 725 images, divided into 80% for training and 20% for testing, our model exhibits exceptional performance. Both the validation and testing phases achieved 100% accuracy, underscoring the efficacy of the proposed methodology. In addition, the area under the curve (AUC), a key metric for evaluating the model’s discriminative ability, demonstrated high performance across various subtypes, with AUC values of 0.94, 0.78, 0.69, 0.92, and 0.94 for MC. Furthermore, the positive likelihood ratios (LR+) were indicative of the model’s diagnostic utility, with notable values for each subtype: CC (27.294), EC (9.441), HGSC (12.588), LGSC (17.942), and MC (17.942). These findings demonstrate the effectiveness of the model in distinguishing between ovarian cancer subtypes, positioning it as a promising tool for diagnostic applications. The demonstrated accuracy, AUC values, and LR+ values underscore the potential of the model as a valuable diagnostic tool, contributing to the advancement of precision medicine in the field of ovarian cancer research.

BackgroundGlomerulonephritis (GN) encompasses a heterogeneous group of kidney diseases, often presenting with subclinical manifestations in children, leading to frequent missed diagnoses. Renal biopsy, while considered the gold standard, is invasive, prone to sampling errors, and time-consuming, thus hindering rapid diagnosis. This study aimed to develop a noninvasive diagnostic model for childhood GN using renal ultrasound images through the integration of deep learning and radiomics techniques.MethodsUltrasound images were acquired from children undergoing ultrasound-guided biopsy. A total of 469 renal ultrasound images were selected and divided into training and validation sets at a ratio of 8:2 to train a U-Net model for precise kidney image segmentation. Using radiomics, a comprehensive set of radiomic features were extracted from the segmented kidney regions. The extracted features were categorized based on GN types: IgA nephropathy (127 cases), minimal change disease (83 cases), and Henoch–Schönlein purpura nephritis (103 cases). These categories were further randomly split into training and validation sets at a ratio of 8:2. Within the training set, analysis of variance (ANOVA) was used for feature selection, followed by supervised Least Absolute Shrinkage and Selection Operator (LASSO) regression for dimensionality reduction, resulting in the selection of 37 features. These features were then integrated with a random forest algorithm to develop a GN classification model. The model's performance was comprehensively evaluated using the validation set.ResultsThe segmentation model exhibited remarkable performance during training, achieving an accuracy of 95.19% in the validation set. Thirty-seven features were identified through feature selection, leading to the development of a robust classification model. Evaluation on the validation set revealed high accuracy and predictive power across different GN categories, with Area Under the Curve (AUC) values ranging from 0.91 to 0.98.ConclusionsThe combined use of deep learning and radiomics techniques utilizing renal ultrasound images demonstrates significant potential for classifying childhood GN subtypes. This noninvasive approach holds promise for improving diagnostic efficiency and patient outcomes in GN.