(ExMod) model for medical image segmentation using scribble annotations

Alshewaier, Hateef, Qin, Yipeng ORCID: https://orcid.org/0000-0002-1551-9126 and Sun, Xianfang ORCID: https://orcid.org/0000-0002-6114-0766 2024. (ExMod) model for medical image segmentation using scribble annotations. Presented at: The 5th International Conference on Medical Imaging and Computer-Aided Diagnosis, Manchester, UK, 19-21 November 2024.

Preview

PDF - Accepted Post-Print Version Download (1MB) | Preview

Abstract

Medical image segmentation presents significant challenges due to the high cost of acquiring precise annotations. The task becomes even more difficult when using weak annotations, such as scribbles, as these annotations provide only limited information about the region of interest. Scribble annotations, however, are easier to acquire in practice, making them a more feasible option. Despite this, training neural networks for segmentation based solely on scribble annotations remains complex. We propose an innovative Expansion and Modification (ExMod) neural network architecture to tackle the challenges inherent in weakly supervised medical image segmentation. While scribble-based supervision has been explored in prior works, ExMod introduces a unique set of enhancements tailored to overcome the limitations of existing methods. Built upon the U-Net framework, our architecture stands out by incorporating multiple advancements designed to boost segmentation accuracy under weak supervision. ExMod introduces additional convolutional layers for richer feature extraction and batch normalization layers to improve training stability and convergence. These modifications lead to superior segmentation performance, particularly when using only scribble annotations. Compared to existing scribble-based methods, ExMod captures intricate image structures more effectively, offering better accuracy with fewer annotations and setting a new benchmark for weakly supervised segmentation. The proposed method was tested on two datasets, i.e., MSCMRseg and ACDC.

Item Type:	Conference or Workshop Item (Paper)
Status:	In Press
Schools:	Computer Science & Informatics
Subjects:	Q Science > QA Mathematics > QA75 Electronic computers. Computer science
Date of First Compliant Deposit:	22 November 2024
Date of Acceptance:	1 November 2024
Last Modified:	30 Nov 2024 02:30
URI:	https://orca.cardiff.ac.uk/id/eprint/174243

Actions (repository staff only)

Edit Item