Demystifying Deep Learning Decisions in Leukemia Diagnostics Using Explainable AI.

Area of Science:

• Hematology
• Computational Biology
• Medical Imaging

Background:

• Conventional leukemia diagnosis relies on expert interpretation of peripheral blood smears and bone marrow assessments.
• These methods face challenges due to biological and imaging variability, necessitating advanced diagnostic tools.
• Existing molecular techniques like LDI-PCR, molecular cytogenetics, and array-CGH supplement diagnosis but are complex.

Purpose of the Study:

• To develop and validate an AI pipeline for accurate and interpretable leukemia diagnosis.
• To integrate Convolutional Neural Networks (CNNs) and Explainable AI (XAI) for transparent diagnostic rationale.
• To establish a comprehensive benchmark dataset for evaluating AI models in leukemia classification.

Main Methods:

• A unified benchmark dataset of 66,550 images covering various leukemia types (ALL, AML, CLL, CML) and healthy controls was curated.
• Multiple CNN backbones (DenseNet-121, MobileNetV2, etc.) were fine-tuned and evaluated using accuracy and F1-score metrics.
• Explainable AI techniques (LIME, Grad-Cam) were employed to provide transparent rationale for the AI's diagnostic decisions.

Main Results:

• MobileNetV2 achieved 97.9% accuracy/F1 on the five-class leukemia classification task.
• DenseNet-121 also demonstrated high performance with 97.66% F1-score, showing strong nucleus-centric explanations.
• XAI analyses successfully localized critical cellular morphology, aligning model saliency with clinical indicators.

Conclusions:

• The proposed AI pipeline, integrating CNNs and XAI, achieves state-of-the-art accuracy in leukemia diagnosis.
• Explainable AI methods provide crucial interpretability, corroborating diagnostic findings with clinical relevance.
• This approach offers a more accurate and transparent alternative to conventional diagnostic workflows for hematologic malignancies.