Enhanced SAM-Med3D: A Robust Solution for 3D Medical Image Segmentation with Advanced Post-processing

Jiacheng Lin, zhengdong, Zhihui Ma, Yulin Xiao, Hongke Fu, YiRen Pan

Keywords: 3D Medical Image Segmentation;SAM-Med3D;Post-processing;Coreset;Limited Data Training ;Deep Learning

TL;DR: This paper presents an enhanced version of SAM-Med3D for the CVPR 2025 3D Medical Image Segmentation Challenge, specifically targeting the \textbf{coreset track} which utilizes only 10\% of the available training data.

Abstract:

This paper presents an enhanced version of SAM-Med3D for the CVPR 2025 3D Medical Image Segmentation Challenge, specifically targeting the coreset track which utilizes only 10\% of the available training data. Our approach builds upon the foundation of SAM-Med3D, a state-of-the-art 3D medical image segmentation model, and introduces a novel 5-step post-processing pipeline designed to maximize performance under limited data constraints.

Our method combines strategic training optimizations with advanced post-processing techniques including region filtering, hole filling, morphological operations, Gaussian smoothing, and overlap resolution. The approach maintains the interactive capabilities of SAM-Med3D while adapting to the data-limited scenario of the coreset track. On the validation set, our enhanced model achieves an average DSC of 0.28 without post-processing and 0.27 with post-processing across multiple medical imaging modalities (CT, MRI, PET, Ultrasound, and Microscopy). While the overall performance shows mixed results, our analysis reveals that the post-processing pipeline demonstrates some improvement for specific modalities such as PET imaging, highlighting the complexity of developing universal enhancement strategies in data-constrained environments.

Our contribution lies in developing a practical approach for 3D medical image segmentation that can work effectively with minimal training data, making it particularly relevant for scenarios where annotated medical data is scarce. The systematic post-processing pipeline provides a framework for improving segmentation quality in data-constrained environments.

Submission Number: 4