Meeting Banner
Abstract #0262

Meta-Learning for Robust Medical Image Segmentation: A Gradient-Similarity Reweighting Approach to Mitigate Noisy Labels

Abdulkhalek Al-Fakih1, Abbas Mohamed Rezk1, Abdulla Shazly1, Kanghyun Ryu2, and Mohammed A. Al-masni1
1Department of Artificial Intelligence and Data Science, Sejong University, seoul, Korea, Republic of, 2Artificial Intelligence and Robotics Institute, Korea Institute of Science and Technology, seoul, Korea, Republic of

Synopsis

Keywords: Diagnosis/Prediction, Segmentation, Meta-Learning, Gradient Similarity, Label Noise, Medical Image Segmentation, Brain Tumor Segmentation, BraTS-PEDs

Motivation: Automated and robust segmentation of medical image requires many data and clean labels.

Goal(s): Our goal was to develop a reweighting approach to mitigate noisy labels via gradient similarity.

Approach: We propose a novel meta-learning approach that dynamically reweights training samples based on their reliability by measuring cosine similarity between per-sample gradients of noisy data and clean meta-data, optimizing loss reweighting.

Results: Comparative analysis showed that the proposed method improved glioblastoma segmentation significantly, with overall Dice scores (DSC) increasing from 0.768% to 0.792%. Notably, in the pediatric task, our method significantly improves segmentation performance by 2.7% in DSC.

Impact: This advancement addresses noisy data and limited data availability in medical image segmentation, enabling more accurate and reliable predictions.

How to access this content:

For one year after publication, abstracts and videos are only open to registrants of this annual meeting. Registrants should use their existing login information. Non-registrant access can be purchased via the ISMRM E-Library.

After one year, current ISMRM & ISMRT members get free access to both the abstracts and videos. Non-members and non-registrants must purchase access via the ISMRM E-Library.

After two years, the meeting proceedings (abstracts) are opened to the public and require no login information. Videos remain behind password for access by members, registrants and E-Library customers.

Click here for more information on becoming a member.

Keywords