Authors:

Sarthak Mangla

Date Created:

Not specified

Course Title:

Professor:

Not specified

About Paper:

Deep Learning techniques are now widely utilized to automate various segmentation and classification tasks in medical imaging. Specifically, performing semantic segmentation on cardiac magnetic resonance (CMR) has proved to save time and effort in the long run. In addition, pretraining models with self-supervision learning methods have been widely successful in bolstering performance. We aim to combine these two methodologies in a novel framework to improve general CMR segmentation accuracy. Our approach involves leveraging a convolutional U-net to construct a general foundational model built on top of a large collection of CMR datasets. We utilize self-supervision learning techniques like contrastive learning via Barlow Twins, image inpainting, and rotation prediction to acquire robust semantic representations of CMR images for the foundation model. This model will then serve as a starting point for subsequent fine-tuning on specific cardiac analysis tasks in diseases like Duchenne muscular dystrophy (DMD). Our previous work specifically focused on automating the segmentation of MRI images for DMD strain calculations. With no pretraining, the DMD images had a test-set Mean Squared Error (MSE) of 0.19 ± 0.11 (cm2; mean ± SD) and a test-set mean Dice score of 0.89 ± 0.08 for the endocardial boundary and a 0.93 ± 0.05 for the epicardial boundary. While the foundational model is still being incorporated, it has demonstrated promise in initial tests. We hope to make this foundational model open-source as a baseline for CMR understanding, enabling further contributions from the community.

Source:

Purdue University / 2023

Topics:

No topics listed

Co-authors:

Sarthak Mangla