MammoDetect - Kaggle Medgemma Impact Challenge Entry
AI-powered mammography analysis using Google's MedGemma 1.5
Mammography analysis app built on Google's MedGemma 1.5 4B vision-language model. Runs N independent inference rounds on the same image and scores agreement across findings, borrowing from how radiologists do double reads in practice.
High-resolution mammography review needs both diagnostic rigor and practical tooling. Single-pass model outputs were not enough for a workflow inspired by clinical double reads.
Owned the product framing, experimentation strategy, and end-to-end application design for the Kaggle challenge submission.
System Constraints
Large medical images with important detail in small regions
Need for explainable, repeatable output structures
Multiple inference backends with different latency and quality tradeoffs
Research-grade positioning without overstating clinical readiness
What I Built
A mammography analysis app centered on multi-round inference and agreement scoring instead of one-shot outputs.
A tiling and prioritization strategy that preserved detail while keeping inference tractable on available hardware.
Structured BI-RADS reporting and bilingual workflows to make the output easier to evaluate and compare.
Outcome / Takeaway
Created a more trustworthy analysis flow by borrowing from the logic of double-reading protocols.
Made backend flexibility a product feature by supporting local, hosted, and optimized inference paths.
Turned a model experiment into a coherent research product with clear user and safety framing.
In medical AI, confidence design matters. The interface needs to communicate uncertainty and agreement just as clearly as it communicates findings.