367 Deep Learning-Based Classification of AML Subtypes Using Cytomorphology Images: A Pathology-Centered Workflow with Grad-CAM Interpretability
Saanie Sulley · American Journal of Clinical Pathology · 2025
Abstract Introduction/Objective Acute Myeloid Leukemia (AML) classification remains a diagnostic challenge due to overlapping morphological features across genetic subtypes. Deep learning models trained on cytomorphology images offer a scalable solution, but class imbalance and interpretability remain key obstacles. Methods/Case Report We trained a convolutional neural network (CNN) using an image-only dataset of AML cytomorphology images spanning five diagnostic classes: NPM1, CBFB-MYH11, RUNX1-RUNX1T1, PML-RARA, and control. The model was optimized using a class-balanced sampling strategy and focal loss to mitigate class imbalance. Model interpretability was assessed using Grad-CAM heatmaps, and predictions were overlaid on original images. Dimensionality reduction techniques (PCA, t-SNE, UMAP) were used to explore learned feature distributions. Results The model achieved a validation accuracy of 55% with robust sensitivity for minority classes. Grad-CAM visualizations consistently highlighted diagnostically relevant nuclear and cytoplasmic regions. UMAP and t-SNE projections revealed distinct subtype clusters, supporting biological relevance of learned features. Classification reports and confusion matrices confirmed predictive strengths in PML-RARA and CBFB-MYH11 subtypes. Conclusion This image-only deep learning pipeline demonstrates promising performance for AML subtype classification, with added interpretability through visual analytics. These findings support further development of AI-assisted hematopathology workflows and demonstrate the potential of deep learning in morphologic precision diagnostics.