Random-split evaluation substantially overestimated performance relative to a resolution-defined acquisition-shift stress test, and entropy-based selective prediction improved reliability by identifying a high-confidence subset for automated prediction while deferring the remainder to human review.
Abstract
Background: Deep-learning systems for endoscopic image classification are commonly evaluated with random data splits, which may overestimate performance under acquisition shift; uncertainty-aware selective prediction may improve reliability by allowing a model to abstain on uncertain cases. Methods: We evaluated binary abnormality detection in two endoscopic imaging domains: laryngeal contact-endoscopy narrow-band imaging (CE-NBI; 210 patients, patient-level) and gastrointestinal endoscopy (HyperKvasir; 6746 images, image-level, predominantly white-light). ImageNet-pretrained ResNet-50 deep ensembles were assessed under a resolution-defined acquisition-shift stress test; for the gastrointestinal data, a random stratified split was additionally used as an in-distribution reference. We evaluated discrimination, calibration, decision-curve analysis, and entropy-based selective prediction. Results: In the gastrointestinal dataset, random-split evaluation produced high performance (AUROC 0.994, 95% CI 0.990–0.996; AUPRC 0.991). Under resolution shift on the same data, performance fell to AUROC 0.723 (95% CI 0.710–0.736; AUPRC 0.690; sensitivity 0.417; specificity 0.874); the two intervals do not overlap. Selective prediction improved reliability among retained cases: under resolution shift, accuracy rose from 0.683 at full coverage to 0.852 (95% CI 0.835–0.871) at 25% coverage (balanced accuracy 0.646 → 0.798). Predictive entropy was significantly higher for incorrect than for correct predictions in both regimes (Mann–Whitney p = 7.1 × 10−77 with rank-biserial |r| = 0.30 under shift). In the laryngeal cohort, no statistically significant differences were detected among four architectures (ROC-AUC 0.844–0.901; all pairwise DeLong p > 0.05). Conclusions: Random-split evaluation substantially overestimated performance relative to a resolution-defined acquisition-shift stress test, and entropy-based selective prediction improved reliability by identifying a high-confidence subset for automated prediction while deferring the remainder to human review. Target-domain recalibration substantially restores calibration under shift (ECE 0.172 → 0.036 with temperature scaling; → 0.017 with isotonic regression) but does not recover discrimination; selective prediction is complementary, mitigating residual confident-wrong predictions. An encoder-transfer experiment showed asymmetric cross-domain utility; features learned on the larger gastrointestinal cohort transferred to the laryngeal cohort (AUROC 0.80 vs. in-domain 0.89), whereas the reverse direction did not transfer (0.53 vs. 0.72). Prospective multi-center validation remains required before clinical deployment. All code, fold definitions, random seeds, and a reproducible protocol are publicly released.
The proposed framework establishes a reliable and lightweight baseline for automated gastrointestinal disease detection and demonstrates that ConvNeXt-Tiny effectively captures disease-relevant visual patterns in endoscopic images while maintaining consistent performance across varying training conditions.
Muhammad Faqih, O. Q. Aziz, Ajib Hanani· Jurnal Ilmu Komputer dan Inf...· 0 citations
This study introduces a hybrid quantum– classical deep learning framework that leverages quantum variational filtering in tandem with transfer learning via ViT-B16 and EfficientNet-B3 to enhance feature representation and demonstrates promising robustness and generalization in limited-data environments.
S. Behuria, Sujata Swain, Anjan Bandyopadhyay et al.· Artificial Intelligence in H...· 0 citations
Background/Objectives: Deep learning approaches have demonstrated strong performance in breast histopathology image classification; however, reliable generalization across heterogeneous acquisition environments remains challenging due to domain shift. In clinical practice, missed malignant cases are particularly critical because they may directly affect diagnostic decisions and patient outcomes. This study systematically investigates the behavior of modern deep learning architectures and adaptation strategies under realistic cross-domain conditions, with particular emphasis on malignant case detection and false-negative reduction. Methods: Three modern architectures—ConvNeXt-Tiny, Swin-Tiny, and MaxViT-Tiny—were initially trained on a large-scale breast histopathology dataset and subsequently evaluated on the BreaKHis dataset using strict patient-level separation to avoid information leakage. Three transfer settings were investigated: direct zero-shot transfer, head-only adaptation, and full fine-tuning. Performance was evaluated independently across four magnification levels (40×, 100×, 200×, and 400×) using accuracy, precision, sensitivity, F1-score, ROC–AUC, PR–AUC, and false-negative rates. Results: Direct zero-shot transfer produced substantial performance degradation across all architectures, with mean false-negative rates ranging from 75.85% to 90.11%, highlighting the limited transferability of source-domain representations under heterogeneous acquisition conditions. Both adaptation strategies substantially improved performance and reduced missed malignant cases to below 10%. Swin-Tiny under head-only adaptation achieved the most favorable malignant detection profile, reaching a mean sensitivity of 97.36% while reducing the average false-negative rate to 2.64%. In contrast, MaxViT-Tiny achieved the highest mean ROC–AUC value (0.849) after full fine-tuning, although this did not correspond to the lowest false-negative burden. Conclusions: The findings demonstrate that maximizing global discrimination performance does not necessarily correspond to optimal malignant detection under cross-domain conditions. Sensitivity and missed-case analysis provide complementary information beyond conventional discrimination metrics and may support more informed model assessment. Furthermore, the proposed methodology provides a reproducible framework for investigating adaptation performance in AI-assisted breast histopathology systems.
L. Stanescu, Cosmin Stoica Spahiu· Diagnostics· 0 citations
Dermoscopic skin lesion classification is a task of major clinical importance but is computationally expensive, making it inaccessible in resource-constrained healthcare settings. In this paper, we introduce a computationally efficient skin lesion classification framework for seven classes using EfficientNet-B0, complemented by Monte Carlo (MC) Dropout for uncertainty quantification. Our approach was trained and tested on the HAM10000 dataset containing 10,015 dermoscopic images across seven classes. To address the severe 67:1 class imbalance, we employ WeightedRandomSamplerand class-weighted cross-entropy loss as complementary corrections acting at the batch-composition level and the gradient-magnitude level respectively. By performing T=50 stochastic forward passes during inference, we decompose predictive uncertainty into aleatoric and epistemic components and apply an entropy-based referral threshold that flags uncertain predictions for specialist review. To validate spatial interpretability, Gradient-weighted Class Activation Mapping (Grad-CAM) is applied and quantitatively evaluated via Intersection over Union (IoU) against ISIC segmentation masks, yielding a mean IoU of 0.61 across all accepted predictions. Our experiments achieve a test macro AUROC of 0.9404and macro F1-score of 0.7308, with six of seven classes exceeding 70% per-class accuracy (melanocytic nevi: 69.8%). Referring the 30% most uncertain predictions to a clinician raises accepted-subset AUROC from 0.9404 to 0.9568 (+1.64%). The framework is competitive with ResNet-50 and DenseNet-121 at one-fifth the parameter count, and the only lightweight method in the comparison providing calibrated uncertainty estimates. Inference latency benchmarks on an NVIDIA Jetson Nano (edge CPU mode) are reported to contextualize deployment feasibility.
Princy Randhawa, S. Suddala, M. Hemal et al.· BioMedInformatics· 0 citations
GOALS
To compare a vision transformer with 2 convolutional neural network architectures for multiclass lesion classification in capsule endoscopy images.
BACKGROUND
Manual review of capsule endoscopy is time-consuming and subject to interobserver variability. Deep learning can automate lesion recognition; however, most prior capsule endoscopy work evaluates a small number of classes, and systematic comparisons between transformer and convolutional architectures across many lesion categories are limited.
STUDY
Two publicly available data sets (SEE-AI and Kvasir-Capsule) were merged and preprocessed to create a 21-class image data set (∼58,000 frames). Images were resized to 224 × 224 pixels and split using stratified sampling into training (n=40,587), validation (n=8696), and test (n=8696) sets. A pretrained Vision Transformer, DenseNet121, and ResNet50 were fine-tuned using categorical cross-entropy loss and Adam optimization, with early stopping. Performance was assessed using accuracy, macroaveraged precision, recall, F1 Score, and the area under the receiver operating characteristic curve.
RESULTS
On the independent test set, the vision transformer achieved 92.2% accuracy with macroaveraged precision/recall/F1-score of 0.92 and an area under the receiver operating characteristic curve of 0.99. DenseNet121 achieved 74.0% accuracy (F1-score 0.78; area under the receiver operating characteristic curve 0.85). ResNet50 achieved 38.0% accuracy (F1-score 0.40; area under the receiver operating characteristic curve 0.55).
CONCLUSIONS
In this merged 21-class capsule endoscopy image data set, the vision transformer achieved higher frame-level classification performance than DenseNet121 and ResNet50 under the present experimental conditions. Importantly, the data set was split at the image level rather than at the patient or procedure level, because frames from the same examination may be correlated; therefore, performance figures likely reflect benchmark results on this frame-level public data set and should not be interpreted as estimates of patient-level generalization or as evidence of definitive architectural superiority. These findings support further evaluation of transformer-based approaches, but grouped reanalysis, external validation, and workflow-oriented studies are required before clinical implementation.
Sri Harsha Boppana, S. Komati, Aditya Chandrashekar et al.· Journal of Clinical Gastroen...· 0 citations
Deep learning has shown strong potential in medical image analysis, but most existing methods rely on large-scale annotations and a closed-world assumption that rarely holds in clinical practice. Although Generalized Category Discovery (GCD) has advanced rapidly on natural images, it remains underexplored in medical imaging. To address this issue, we propose MedXplore, a unified framework for reliable and unbiased medical GCD, optimizing from both perceptual and decision levels. Specifically, at the perceptual level, taking a frequency domain perspective, Frequency-SNR Adaptive Attention and Consistency (FAAC) performs learnable full-spectrum filtering and global-local energy contrast activation to not only highlight local abnormal signals relative to the global context, but also provide reliable semantic anchors for patch consistency learning. At the decision level, Adaptive Cosine-Angular Margin (ACAM) adjusts angular margins using semantic difficulty and feature confidence to balance intra-class compactness and inter-class separability. Together, the two modules improve lesion-sensitive representation learning and mitigate old-class bias. Experiments on multiple benchmarks show an average \textbf{8.5\%} gain in \textit{All} accuracy over the strongest competing methods. On Kvasir, MedXplore reduces false-old errors from 14.50\% to 0.80\%, demonstrating strong robustness under severe old-new ambiguity.
