Classification of Hematological Malignancies Using Peripheral Blood Smear Images
The Core Challenge
Manual microscopic examination of peripheral blood smears to diagnose leukemia and other hematological malignancies is labor-intensive, subject to observer fatigue, and hindered by extreme dataset class imbalance where rare pathological malignant subtypes are severely underrepresented.
Methodological Contribution
Proposes a dual-stage framework using Generative Adversarial Networks (GANs) for high-fidelity synthetic microscopic cell generation to balance training distributions, combined with a deep CNN classifier to achieve robust multi-class leukemia subtype discrimination.
Methodological Pipeline & Tensor Transformations
Microscopic Image Preprocessing
Illumination correction, color normalization, and white blood cell segmentation from blood smear slides.
GAN-Driven Data Generation
Conditional generator network synthesizes rare malignant cell variants under discriminator guidance.
Balanced Representation Learning
Convolutional feature encoder is trained on the augmented, class-balanced dataset to learn fine-grained morphologies.
Multi-Class Malignancy Classification
Outputs calibrated malignancy classifications across acute lymphocytic leukemia (ALL), AML, and healthy controls.
Reported Evaluation Results
Scientific Guardrails & Validation Boundaries
Future iterations explore diffusion-based augmentation architectures and direct integration with digital pathology scanner APIs.
Translation to Enterprise Production
Demonstrates advanced generative AI applications in production computer vision: utilizing GANs to systematically resolve extreme data sparsity, ensuring classifiers maintain high recall on critical rare events in mission-critical deployments.
Access Verified Publication Records
All citations, publisher identifiers, and code references adhere to strict verification standards.