Deep Learning Techniques for Automatic MRI Cardiac Multi-structures Segmentation and Diagnosis: Is the Problem Solved?

Description

Delineation of the left ventricular cavity,myocardium, and right ventricle from cardiac magneticresonance images (multi-slice 2-D cine MRI) is a commonclinical task to establish diagnosis. The automationof the corresponding tasks has thus been the subjectof intense research over the past decades. In this paper,we introduce the "Automatic Cardiac Diagnosis Challenge"dataset (ACDC), the largest publicly available and fully annotateddataset for the purpose of cardiac MRI (CMR) assessment.The dataset contains data from 150 multi-equipmentsCMRI recordings with reference measurements and classificationfrom two medical experts. The overarching objectiveof this paper is to measure how far state-of-the-art deeplearning methods can go at assessing CMRI, i.e., segmentingthe myocardium and the two ventricles as well as classifyingpathologies. In the wake of the 2017 MICCAI-ACDCchallenge, we report results from deep learning methodsprovided by nine research groups for the segmentation taskand four groups for the classificationtask. Results show thatthe best methods faithfully reproduce the expert analysis,leading to a mean value of 0.97 correlation score for theautomatic extraction of clinical indices and an accuracy of0.96 for automatic diagnosis. These results clearly openthe door to highly accurate and fully automatic analysis ofcardiac CMRI. We also identify scenarios for which deeplearning methods are still failing. Both the dataset anddetailed results are publicly available online, while the platformwill remain open for new submissions.

Abstract

International audience

Additional details