RadAI: Gaze-Intention Fusion Pipeline for Attention-Guided Educational Videos in Chest X-Ray Interpretation

원문 초록

Chest radiograph interpretation errors often arise from failures in visual search: although trainees may correctly identify an abnormality once pointed out, they frequently fail to fixate on the relevant region, while expert search strategies (where radiologists look, in what sequence, and for how long) remain largely implicit. Eye movement modeling examples (EMME), which replay a recorded expert's gaze over the image, have been shown to improve novice detection of findings on chest radiographs, but producing one requires an eye-tracking session on that specific case, so teaching sets are limited to recorded sessions by experts. In this work, RadAI is presented, the first end-to-end framework for transforming predicted radiologist gaze into narrated instructional videos, enabling scalable gaze-guided training without collecting new eye-tracking data. Given only a chest X-ray and its report, RadAI predicts an expert fixation sequence, associates each interval of that sequence with the diagnostic finding it pursues, extracts the corresponding image region, and renders the result as an EMME-style video in which each highlighted finding is narrated with synthesized speech. RadAI is evaluated on 97 cases from the EGD-CXR dataset, assessing whether the highlighted regions correspond to annotated abnormality locations. We position this work as a framework for scalable generation of educational content for residents and medical students in the field of radiology.