FeDETR: a Federated Approach for Stenosis Detection in Coronary Angiography

TL;DR

FeDETR addresses the privacy constraints of multi-institution coronary angiography data by federating only the backbone of a DETR-based stenosis detector across hospitals, while each node trains its local DETR on key angiography frames. The central server aggregates backbone weights via FedAvg, enabling privacy-preserving, non-IID data collaboration and improved detection performance over local and some federated baselines. Evaluated on 1001 examinations from five hospitals with FFR/iFR labels, FeDETR achieves competitive accuracy and better detection of mild/high-severity stenoses, confirming robustness in real clinical settings. The work highlights a first federated detection-transformer approach for medical imaging and points to future work on key-frame selection and broader contrast-enhanced imaging applications.

Abstract

Assessing the severity of stenoses in coronary angiography is critical to the patient's health, as coronary stenosis is an underlying factor in heart failure. Current practice for grading coronary lesions, i.e. fractional flow reserve (FFR) or instantaneous wave-free ratio (iFR), suffers from several drawbacks, including time, cost and invasiveness, alongside potential interobserver variability. In this context, some deep learning methods have emerged to assist cardiologists in automating the estimation of FFR/iFR values. Despite the effectiveness of these methods, their reliance on large datasets is challenging due to the distributed nature of sensitive medical data. Federated learning addresses this challenge by aggregating knowledge from multiple nodes to improve model generalization, while preserving data privacy. We propose the first federated detection transformer approach, FeDETR, to assess stenosis severity in angiography videos based on FFR/iFR values estimation. In our approach, each node trains a detection transformer (DETR) on its local dataset, with the central server federating the backbone part of the network. The proposed method is trained and evaluated on a dataset collected from five hospitals, consisting of 1001 angiographic examinations, and its performance is compared with state-of-the-art federated learning methods.

Figures (3)

• Figure 1: Each node uses its own dataset, consisting of key frames extracted from coronary angiographies videos, to train a detection transformer (DETR). At the beginning of each round, the central server sends the aggregated backbone to each node, which uses it to extract features from the key frames. After each round, each node sends its locally trained backbone back to the central server, which aggregates them all.
• Figure 2: Given a key frame, the CNN backbone learns a 2D representation, which is then flattened by the model. Before passing these learned features to the transformer encoder, the model also extends them with positional encoding. The transformer decoder outputs some learned positional embeddings and feeds them into a feed-forward network that predicts classes and their corresponding bounding boxes.
• Figure 3: Figure \ref{['fig:stenosis']} shows representative examples of low severity stenoses (illustrated in the top row) and mild or high severity stenoses (depicted in the bottom row). Figure \ref{['fig:heatmap']} presents an averaged heatmap, which is derived from the collective bounding box positions across all samples.