Go to CVPR 2025 Workshop MedSegFM homepageMobileSeg3D: A Lightweight Framework for Multi-Modality 3D Medical Image Segmentation
Keywords: Multi-Modality Learning, Lightweight Architectures, Prompt-based Segmentation, 3D Medical Image Segmentation
TL;DR: Our method combines efficient MobileNet encoders with prompt generation from image intensities to enable fast, annotation-sparse segmentation across diverse 3D medical modalities.
Abstract: The growing availability of complex 3D medical imaging data, including CT, MRI, PET, ultrasound, and microscopy, has increased the demand for segmentation models that are accurate, efficient, and robust across imaging modalities. Although recent 3D architectures such as SAM-Med3D, SegVol, and VISTA3D have shown promising results, they often struggle with modality generalization, interactive refinement, and input variability. In this work, we present a lightweight and modular segmentation framework designed to address these challenges. The architecture integrates encoder variants and bottleneck bypass connections to better preserve spatial and modality-specific information. To handle weak or missing annotations, we introduce an intensity-based thresholding strategy that generates bounding box prompts in the absence of detailed labels. We also explore MobileNet-based backbones, which have been underutilized in 3D medical segmentation, and demonstrate that they outperform heavier models such as SegVol in low-resource and modality-diverse scenarios. Our approach achieves competitive segmentation accuracy while remaining computationally efficient and well-suited for interactive refinement. Experiments on the CVPR BiomedSegFM dataset confirm the model's strong generalization across modalities and robust performance during iterative use. On the official validation leaderboard, our method achieved an average DSC score of 0.50 and ranked 4th overall among participating teams. Our code is publicly available here: https://github.com/lexorcvpr/lexor-cvpr-2025
Submission Number: 5
Loading