HemaSight AI: A Four-Stage Deep Learning Pipeline for Automated Leukemia Detection and Post-Transplant Monitoring

Abstract: Leukemia is a life-threatening hematological malignancy where early and accurate detection significantly improves patient survival. Conventional diagnosis relies on manual microscopic examination of peripheral blood smears, a process that is time-consuming, observer-dependent, and prone to error. This paper presents HemaSight AI, a four- stage artificial intelligence pipeline for automated leukemia screening and post-transplant monitoring. Stage 1 applies XGBoost to classify complete blood count (CBC) parameters, including 17+ extended markers such as SGOT, SGPT, creatinine, potassium, chimerism, and GVHD status. Stage 2 employs a conditional generative adversarial network (cGAN) to synthesize microscopic blood smear images alongside a ResNet18-based similarity engine that computes a weighted four-metric match (semantic, color, structural, pixel). Stage 3 uses linear regression on longitudinal blood test data to detect early signs of relapse before clinical symptoms appear. Stage 4 integrates Google Gemini via a multilingual chatbot supporting seven languages. The platform is implemented using Next.js, FastAPI, Supabase, and PyTorch, deployed entirely on free-tier infrastructure. Evaluation across 68 test cases achieved a 98.5% pass rate with sub-3-second API response times. Results demonstrate that a low-cost, scalable, and clinically accessible AI tool can be built without proprietary infrastructure.

Keywords: Leukemia detection, XGBoost, conditional GAN, ResNet18, time-series analysis, large language model, medical AI, multilingual chatbot.