Cardio Vision AI
A research prototype integrating trained AI models for echocardiography, CCTA, and 12 lead ECG analysis with explainability, evidence aggregation, and MedGemma assisted reporting.
CardioVision AI is a research prototype that brings three cardiovascular AI pipelines into a single application. It analyzes echocardiography, CCTA, and 12 lead ECG data using trained deep learning models and presents model outputs together with explainability and structured evidence. The system combines a FastAPI backend, React frontend, local case management, model inference pipelines, XAI, and optional local MedGemma based narrative reporting and clinical question answering. A single file Streamlit interface is also provided as an open demonstration surface with built in sample cases. The system is designed as research software and is not a clinically validated or regulatory approved medical device.
Method
CardioVision AI uses modality specific models rather than forcing different medical data types into a single model. Echocardiography: UNet++ with an EfficientNet B3 encoder for four class cardiac structure segmentation using patient level CAMUS splits. CCTA: Small3DUNet for coronary lumen segmentation after volumetric resampling and intensity normalization. Inference uses a 96³ sliding window with 50% overlap and a 0.60 segmentation threshold. ECG: A 1D residual CNN for five class multi label screening using 12 lead PTB XL ECG signals. Explainability: Grad CAM is used for CCTA, while input gradient based attribution is used for echocardiography and ECG. Evidence layer: Cross modal evidence aggregation is deterministic software and does not constitute a learned multimodal fusion model. Language layer: MedGemma is used locally to generate narrative explanations and answer case level questions using computed findings as context.
Results
0.9044
Dice
Echocardiography test Dice on the CAMUS held out split of 75 patients and 300 image/mask pairs.
0.8282
IoU
Echocardiography test IoU on the CAMUS held out split of 75 patients and 300 image/mask pairs.
0.5996
Dice
CCTA test Dice across 3 held out test cases from the MedHK23/CCA dataset. The test set is too small to support a reliable general performance estimate.
0.9125
Macro AUROC
12 lead ECG test macro AUROC across 2,793 held out patients from PTB XL.
0.3614
HYP Precision
ECG hypertrophy class precision at the shipped threshold on the held out test set, demonstrating the substantially weaker performance of this class compared with the other ECG classes.
A locally deployed multimodal cardiovascular AI platform integrating CCTA, echocardiography, and clinical data with segmentation, explainable prediction, and medical AI assisted case level Q&A.