Marrow Clues in the Detection of Leukemia Using Machine Learning

Acute Lymphoblastic Leukemia (ALL) is the most prevalent form of childhood cancer, accounting for approximately 25% of all pediatric malignancies worldwide. Conventional diagnosis relies on a trained hematopathologist manually examining Giemsa-stained bone marrow smear slides, a process that is both time-consuming and inherently subjective, with inter-observer variability of up to 20% reported in clinical literature. This paper presents Marrow-Find, a fully integrated AI-powered web application designed to automate the classification, visual explanation, and quantification of leukemia cells from bone marrow microscopic images. The system employs a ResNet50 deep convolutional neural network, pre-trained on ImageNet and fine-tuned on the C-NMC 2019 and ALL-IDB benchmark datasets, to classify bone marrow cells into four clinically significant categories: Benign, Pre-B ALL, Pro-B ALL, and Early Pre-B ALL. The model achieves an overall test accuracy of 94.8% with a weighted F1-score of 0.954 and a Cohen's Kappa of 0.931. To address class imbalance, a Conditional Generative Adversarial Network (cGAN) generates realistic synthetic bone marrow cell images for minority classes. Gradient-weighted Class Activation Mapping (Grad-CAM) produces visual heatmap overlays on each prediction, enabling pathologists to verify the morphological features influencing the model's decision. A Watershed-based segmentation algorithm automatically delineates individual cell boundaries and counts blast cells per image. All five components are integrated into a Flask web application with secure user authentication, prediction history, PDF report generation with QR codes, and an AI chatbot, making specialist-level leukemia diagnostics accessible from any standard web browser.