The challenge
Medical-imaging AI has a higher bar than producing a model that performs well on a handful of examples.
The goal was not just to train models. It was to build the engineering foundation required to turn medical-imaging models into an assistive product.
What was built
Prepared and curated medical-imaging datasets and training workflows for different imaging tasks.
Built PyTorch-based classification systems for chest X-ray and mammography, plus localization models designed to identify relevant regions rather than only an image-level prediction.
Tools and measures
- PyTorch
- Classification
- Object localization
- Quantitative evaluation
Current development benchmarks
The results are treated as engineering measurements, not a finish line. Model development continues as datasets, evaluation sets, and system requirements improve.
- Chest X-ray classification: 95.5% overall F1
- Dual-view mammography classification: 85% F1
- Tumor localization: 65.1% mAP
From model to product
A useful medical-AI system needs more than a trained checkpoint. The work extended into backend APIs, Google Cloud deployment, integration of model outputs into the web product, development of the Mknoon website, and infrastructure connecting ML components to the user-facing application.
Reliability and human oversight
The product is positioned as assistive decision support, not autonomous diagnosis.
Outputs should remain reviewable by qualified humans; performance should be evaluated separately for each task; confidence and failure cases need to be understood; and new model versions should be compared against established evaluation sets.
Why this project matters
The project demonstrates the complete engineering path behind a high-stakes AI product and reinforces a principle behind Relytic: where mistakes matter, performance needs to be measurable and human oversight needs to be designed into the system.