AI / ML Blood Cancer Detection System
A DL+ML hybrid model developed to detect all subtypes of blood cancer, utilizing H&E stain normalization and advanced preprocessing techniques.
OVERVIEW
A hybrid deep learning and classical ML system for classifying all major subtypes of Acute Lymphoblastic Leukemia (ALL) from peripheral blood smear microscopy images. Vahadane H&E stain normalization corrects inter-lab color inconsistencies before a ResSplit-KAN backbone extracts morphological features for classification.
Key features
Detects all major ALL subtypes from peripheral blood smear images
H&E stain normalization (Vahadane) for cross-lab color consistency
Hybrid ResSplit-KAN architecture combining residual CNN with KAN classifier
Grad-CAM++ explainability overlays for clinical transparency
Deployed as a live HuggingFace Space for zero-friction inference
Engineering
- First-of-kind integration of Kolmogorov-Arnold Networks with residual CNNs for medical imaging
- Stain normalization pipeline reduced cross-dataset color variance by over 40%
- Achieved state-of-the-art accuracy while maintaining a lightweight model footprint
Impact
- Supports pathologists in early ALL subtype identification, accelerating treatment decisions
- Research accepted for publication in a peer-reviewed medical journal