2026-10-06

Rationale And Objectives: This study investigated the clinical utility of a machine learning (ML) model based on ultrasoundomics features (UFs) from primary foci to predict pathological complete response (pCR) in patients with breast cancer (BC) undergoing neoadjuvant chemotherapy (NAC). Materials And Methods: Ultrasound images from 333 patients with BC treated with NAC were retrospectively analyzed. Patients in the internal training cohort and external validation cohort were from two different hospitals, respectively. Tumor regions of interest (ROI) were delineated, and outward segmentation of the peritumor area was performed, using the ROI edge as a baseline and expanding it by 3 mm and 5 mm, respectively. This resulted in three distinct segmentation images: ROI (intratumoral region), ROI (intratumoral region + 3-mm peritumoral region), and ROI (intratumoral region + 5-mm peritumoral region). Three sets of UFs were extracted, normalized using the Z-score method, and further refined through Spearman correlation analysis and LASSO for feature selection. Eight ML models (R0), based on key UFs in ROI (intratumoral region), were compared: support vector machine, K-nearest neighbor, random forest, Naïve Bayes, logistic regression, AdaBoost (AB), multilayer perceptron, and light gradient boosting machine learning. The optimal ML models were evaluated using diagnostic performance metrics, including AUC, accuracy, sensitivity, and specificity. The best-performing algorithm was then applied to construct ML models (R3 and R5) incorporating ROI (intratumoral region + 3-mm peritumoral region) and ROI (intratumoral region + 5-mm peritumoral region). Predictive efficacy and clinical utility were assessed through decision curve analysis (DCA) and Shapley Additive exPlanation (SHAP), which quantified feature importance and facilitated model visualization and interpretation. Results: Among the 333 patients, 119 (35.7%) achieved pCR, while 214 (64.3%) did not. Among the eight ML algorithms, the R0 model based on AB demonstrated the best performance, with an AUC of 0.811 (Internal training cohort) and 0.632 (External validation cohort). The R3 model also showed improved performance, with an AUC of 0.830 (Internal training cohort) and 0.699 (External validation cohort). The R5 model yielded the highest predictive accuracy, with an AUC of 0.871 (Internal training cohort) and 0.779 (External validation cohort). SHAP analysis was conducted for R5, and SHAP waterfall charts enable single-sample predictions to validate the model's clinical utility. DCA revealed that the combination of intratumoral region and 5-mm peritumoral region in R5 offered superior clinical benefit. Conclusion: ML models integrating intratumoral and peritumoral ultrasoundomics offer valuable assistance in assessing NAC efficacy in patients with BC.

Background: Accurate assessment of aortic stenosis (AS) requires integration of both structural and functional information characterized by visual traits as well as quantitation of gradients. Existing artificial intelligence models use solely either structural or functional information. Objectives: The authors aimed to develop and validate an automated multiview deep learning framework to assess AS severity. Methods: The authors developed EchoNet-AS, an open-source end-to-end integrated approach combining video-based convolutional neural networks to assess valve motion with segmentation models to automate measurement of aortic valve peak velocity and classify AS severity. Results: EchoNet-AS was trained on 210,193 images from 16,076 studies from KPNC (Kaiser Permanente Northern California) and validated on 1,588 held-out test studies and a temporally distinct cohort of 19,202 studies. The final model was also externally validated on 2,415 studies from SHC (Stanford Healthcare) and 9,038 studies from CSMC (Cedars-Sinai Medical Center). Combining assessments from multiple echocardiographic videos and Doppler measurements, EchoNet-AS achieved excellent discrimination of severe AS with an area under the receiver-operating characteristic curve of 0.964 (95% CI: 0.952-0.973) in the KPNC held-out cohort and 0.986 (0.983-0.989) in the temporally distinct cohort, which was superior to models using single views or only Doppler measurements. The performance was consistently robust in distinct external cohorts with an area under the receiver-operating characteristic curve of 0.985 (0.975-0.992) at SHC and 0.989 (0.986-0.992) at CSMC. Conclusions: EchoNet-AS synthesizes information from both B-mode videos and Doppler images to accurately assess AS severity. Its strong performance generalizes robustly to external validation cohorts and shows potential as an automated clinical decision support tool.

Muscle-invasive bladder cancer (MIBC) exhibits significant heterogeneity in clinical outcomes and treatment responses. While gene expression profiling has provided important biological and prognostic insight, transcriptomic assays are not routinely incorporated into standard clinical practice, in part due to cost. AI models have been demonstrated to predict gene expression from H&E-stain whole-slide images (WSIs) potentially offering a scalable approach for gathering transcriptomic information for patient stratification and biomarker discovery. We developed and tested models to predict expression of 857 genes directly from H&E-stained whole-slide images (WSI). The best performing final model was a dual-ensemble combining predictions from models trained on features extracted with different foundation models. The dual-ensemble reached 776/857 predictable genes with an average correlation of 0.43 (cross-validation), 0.45 (holdout) and 0.40 (external Frankfurt cohort). Predicted expression strongly correlates with expression of clinically relevant markers (like NECTIN4). Estimated hazard ratios of predicted expression correlated strongly with those of measured expression and genes associated with overall survival for both predicted and measured expression in both cohorts were identified. AI-predicted basal vs. luminal subtypes were associated with overall survival and aligned with the ground-truth (accuracies > 0.8). Findings demonstrate that gene expressions inferred from routine staining can support biomarker discovery and patient stratification in MIBC.

Background: Oral and esophageal squamous cell carcinomas (OSCC/ESCC) are commonly diagnosed late since histopathology is subjective and robust molecular tools are lacking. Leveraging interpretable machine learning, we aimed to develop an objective and platform-agnostic molecular measure of malignancy that overcomes the technical and geographic barriers of conventional genomic testing. Methods: Using a discovery-to-generalization framework with 2,541 samples, we developed and validated a single-sample transcriptomic scoring system, EsoralPrint. We feature engineered 34 intra-sample ordinal gene pairs (isoPairs) that capture consistent expression-order reversals between malignant and healthy states. It was trained on oral tissues and externally validated across diverse global cohorts spanning the upper aerodigestive tract, technical platforms, and geographic regions. Independent single-cell RNA sequencing was used to provide post-hoc mechanistic explainability. Results: In the discovery cohort, EsoralPrint distinguished patients histologically graded as low-risk hyperplasia who nonetheless progressed to cancer (log-rank p < 0.0001); we treat this result as exploratory pending independent confirmation. In external validations, EsoralPrint reached 95.0% accuracy and 96.7% F1-score across OSCC cohorts (460 OSCC, 162 non-OSCC) and 91.6% accuracy and 91.6% F1-score in esophageal tissue unseen during discovery and training (574 ESCC, 614 healthy mucosa). Further, a malignancy gradient reflecting field cancerization effects in resection boundaries was observed, and a transition zone (scores 0.25-0.75) was defined where molecular alterations preceded morphological changes. Conclusion: EsoralPrint provides a platform-independent and quantitative complement to histological grading in oral and esophageal squamous malignancy, with the potential to give clinicians lead time for intervention. It can be accessed via https://github.com/aiphaqua/EsoralPrint.