A multi source/task deep learning system for non-invasive detection and estimation of coronary artery stenosis

Abstract

Cardiovascular disease is the leading cause of death worldwide, and coronary artery

disease is its most common form. To diagnose it, clinicians rely on X-ray coronary angiog- raphy, but interpreting these images manually is slow and prone to disagreement between

experts. Most deep-learning tools address the three sub-tasks (outlining the vessels, find- ing the narrowing or stenosis, and measuring its severity through quantitative coronary

analysis, QCA) with separate models trained one after another. A few combined models

exist for CT scans, but for X-ray angiography no single network yet produces all three out- puts at once, fuses additional non-image data, and reports its own confidence. This thesis

builds and compares three designs of increasing ambition, termed flows, on a 200-patient dataset. Flow 1 is a baseline in which three separate models run in sequence: a U-Net that outlines the vessels, a detector that finds the stenosis, and a stage that measures it. Because each is trained independently, an error early in the chain propagates and is

amplified, which is the central weakness this baseline exposes. Flow 2 is the core contri- bution: a single model, GCT-Net, that performs all three tasks together. One shared

MaxViT encoder feeds three heads, two that outline the vessels and the stenosis and one that measures the stenosis by focusing only on the detected lesion. A stop-gradient prevents the measuring head from distorting the detection, and the three task losses are balanced automatically rather than by hand. Flow 3 extends GCT-Net into a multi-source model. Alongside the angiogram, it incorporates two inputs a cardiologist also consults, the ECG and the written procedural report, merging them through a lightweight fusion layer (FiLM). It adds four report-derived heads and wraps the measurements in reliable confidence intervals, broadening the output at a small cost in image accuracy. The results show the benefit of joint optimisation: Flow 2 raises the stenosis overlap score (Dice) from 0.465 to 0.610 and lowers the measurement errors from 18.9 pp and 5.7 mm to 11.73 pp and 4.51 mm in a single pass, while Flow 3 trades a little accuracy for broader clinical output. The model is also validated on a separate public dataset (ARCADE), reaching a Dice of 0.523 on 300 unseen images, and is deployed as a web application that converts raw DICOM archives into a structured QCA report.