Assessing generalizability and uncertainty in deep learning‑based brain MR image reconstruction from healthy to pathological data

Le résumé fourni par la source

Magnetic resonance (MR) imaging is a common clinical imaging technique. Recent work has demonstrated that deep learning (DL)-based MR reconstruction methods may allow reduced scan times; however, the generalizability of such techniques is often poorly characterized. We investigated the generalizability of the end-to-end variational network (E2E-VarNet) architecture using data from healthy individuals and patients (termed healthy and pathological data, respectively). We trained six models using different proportions of healthy and pathological data. All models achieved similar reconstruction performance on a healthy data test set. When tested using pathological data, the models trained on mixed healthy/pathological data achieved comparable performance, while the model trained on exclusively healthy data performed significantly worse. All models, but especially the model trained without pathological data, tended to ”in-paint” pathological regions with structures reminiscent of healthy anatomy. Monte Carlo dropout was used to estimate voxel-wise reconstruction uncertainty. Pathological regions had higher uncertainty, particularly in models with limited pathology exposure during training. These findings motivate the development of methods that promote model generalizability and establish accuracy when reconstructing pathology data across models.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.