Cardiac magnetic resonance imaging (MRI) plays a pivotal role in cardiovascular diagnostics, offering detailed anatomical and functional insights. Despite its utility, the current workflow for cardiac MRI-spanning data acquisition, reconstruction, report generation, and quality assurance (QA)-is often disjointed, time-consuming, and prone to subjective interpretation. This project aims to revolutionize the workflow by developing an End-to-End Foundation Model for Cardiac MRI, leveraging Vision-Language Models (VLMs), advanced segmentation tools, and automated QA systems.
Research Objectives:
Aim 1: Develop a state-of-the-art framework integrating a VLM with segmentation tools to automate comprehensive cardiac MRI report generation.
The VLM will undergo self-supervised pretraining on a large cardiac MRI dataset, followed by fine-tuning for MR sequence classification, segmentation, and report generation.
A sequence classifier will ensure accurate identification of MRI sequences, enabling appropriate inputs for segmentation and reporting.
Advanced segmentation tools will delineate cardiac chambers and myocardium, extracting critical metrics such as valve function and ejection fraction.
The VLM will generate structured, clinically relevant reports combining quantitative measurements and descriptive findings to aid diagnoses and decision-making.
Aim 2: Implement robust QA for VLM-generated reports using a custom RadGraph and human assessments.
A cardiac MRI-specific RadGraph will assess report accuracy, consistency, and completeness. Human QA will include Turing Tests to evaluate report quality, workflow efficiency analyses, and studies on radiologist performance with and without VLM-generated reports.
Scientific Rationale: The proposed work addresses key limitations in cardiac MRI workflows by integrating advanced AI tools for automation and QA. This project is poised to significantly enhance diagnostic accuracy, streamline and clinical workflows.
