An academic clinical decision-support prototype for analyzing paired stress and rest myocardial perfusion images. It compares three transfer-learning CNNs, combines their outputs through ensemble inference, generates structured reports and stores scan history. It is a research prototype, not a clinically validated medical device.
The Engineering Challenge
System Architecture & Topology
Paired stress and rest images are validated by the web layer, processed by an isolated TensorFlow service, combined into an ensemble output and converted into a structured report with a deterministic fallback.
Strategic Architectural Decisions
Decision 1: Key Trade-off↓
Model paired stress and rest images as a six-channel input instead of treating each scan as unrelated.
Decision 2: Key Trade-off↓
Compare multiple pretrained CNN backbones and expose individual model outputs alongside the ensemble.
Decision 3: Key Trade-off↓
Separate the TensorFlow inference service from the Next.js application for independent deployment.
Decision 4: Key Trade-off↓
Use a deterministic rule-based reporting fallback when Groq is unavailable.
Decision 5: Key Trade-off↓
Persist scans and reports in Supabase while retaining a local browser fallback for demonstration.
Decision 6: Key Trade-off↓
Clearly label the application as an academic prototype and document privacy, authorization and clinical-validation gaps.
