Optimizing SAT-Nano for Text-Guided 3D Medical Segmentation using Dice Semimetric Loss

Zhi Li, YAQI WANG, Shuai Wang

Keywords: Foundation Models;3D Biomedical Image Segmentation; Text-guided Segmentation

Abstract: This paper presents SAT-Nano-JDT, an approach for the CVPR 2025 Text-guided 3D Biomedical Segmentation Challenge, involving fine-tuning the SAT-Nano baseline with JDTLoss (Dice Semimetric Loss). Pre-trained on 10\% of challenge data, SAT-Nano was further trained using a composite loss including JDTLoss, aiming to directly optimize Dice scores and enhance segmentation. On the validation coreset, SAT-Nano-JDT showed mixed results: CT semantic DSC improved to 0.644 (vs. 0.643 baseline) and Microscopy instance DSC TP to 0.310 (vs. 0.292). However, MRI and PET performance did not exceed the baseline. This empirical study explores JDTLoss's utility in refining foundation models, noting the challenges in surpassing strong baselines. The code is available at https://github.com/ricoleehduu/SAT-Nano-JDTLoss.git.

Submission Number: 11