Reliable leukemia detection via transfer-enhanced Bayesian CNNs

Area of Science:

• Medical Imaging
• Computational Biology
• Artificial Intelligence in Medicine

Background:

• Early detection of Acute Lymphoblastic Leukemia (ALL) is crucial for effective treatment and patient survival.
• Deep learning for hematological image analysis faces challenges like limited data, bias, and the need for trustworthy clinical predictions.
• Existing models often lack robust uncertainty quantification, hindering clinical adoption.

Purpose of the Study:

• To develop and evaluate a Bayesian deep learning framework for robust classification of leukemic and healthy lymphocytes from peripheral blood smear images.
• To integrate transfer learning, data augmentation, and uncertainty quantification for improved diagnostic accuracy and reliability.
• To assess the clinical trustworthiness of AI predictions by analyzing uncertainty estimates.

Main Methods:

• Utilized three pre-trained Convolutional Neural Network (CNN) architectures (InceptionV3, VGG16, ResNet50) fine-tuned on the ALL-IDB2 dataset.
• Implemented Monte Carlo dropout for Bayesian inference and uncertainty quantification.
• Employed 10-fold cross-validation with data augmentation, evaluating performance using accuracy, sensitivity, specificity, Youden's index, and Brier score.

Main Results:

• The VGG16 model, enhanced with data augmentation, achieved the highest accuracy (98.65%±0.09), Youden's index (0.97±0.001), and lowest Brier score (0.035±0.010).
• Bayesian uncertainty analysis revealed that misclassifications correlated with high predictive entropy and mutual information.
• Saliency maps indicated that high-uncertainty predictions were linked to non-localized attention patterns, suggesting reliance on spurious features.

Conclusions:

• The proposed Bayesian deep learning framework offers strong diagnostic performance for ALL detection from blood smear images.
• Uncertainty quantification is vital for identifying ambiguous cases requiring expert pathological review, enhancing clinical trust.
• This AI-assisted approach provides a transparent and reliable tool for leukemia screening.