Reliable leukemia detection via transfer-enhanced Bayesian CNNs

Accurate and early detection of Acute Lymphoblastic Leukemia (ALL) is critical for timely intervention and improved patient outcomes. However, the development of reliable deep learning models for hematological image analysis is challenged by limited data availability, dataset bias, and the need for trustworthy predictions in clinical settings. In this study, we propose a Bayesian deep learning framework that integrates transfer learning, data augmentation, and uncertainty quantification for robust classification of leukemic and healthy lymphocytes from peripheral blood smear images. Three widely used convolutional neural network architectures, InceptionV3, VGG16, and ResNet50, pretrained on ImageNet are fine-tuned on the ALL-IDB2 dataset and extended with Monte Carlo dropout to enable Bayesian inference. Model performance is evaluated using 10-fold cross-validation on both original and augmented datasets, with accuracy, sensitivity, specificity, Youden's index, and Brier score used as evaluation metrics. Among the evaluated models, VGG16 demonstrates the most consistent improvements under data augmentation, achieving the highest accuracy (98.65%±0.09), Youden's index (0.97±0.001) and Brier score (0.035±0.010), while ResNet50 shows strong but more moderate gains. In contrast, InceptionV3 exhibits limited sensitivity to augmentation and comparatively lower robustness. Beyond average predictive performance, Bayesian uncertainty analysis reveals that misclassifications and borderline predictions are consistently associated with elevated predictive entropy and mutual information. Saliency map inspection further indicates that high-uncertainty cases correspond to diffuse or non-localized attention patterns, suggesting reliance on spurious contextual features rather than stable morphological cues. These findings highlight the importance of uncertainty-aware predictions for identifying cases that may require expert pathological review. Overall, the proposed framework combines strong diagnostic performance with interpretable uncertainty estimates, supporting its role as a transparent and clinically trustworthy tool for AI-assisted leukemia screening.