Anatomically-Enhanced URO dot AI: Multi-Stage Fine-Tuning of Foundation Models for Precise Urinary Stone Segmentation

Byeongin Moon, Jaehoon Oh, Young Eun Yoon, Dong Keon Lee, Gihun Lee, Mingyeong Park, Heesun Choi, Hansol Choi, JAEHOON JEONG

Keywords: Urinary Stone, Foundation Model, Hierarchical Fine-Tuning, VISTA3D

Registration Requirement: Yes

Abstract: Automated segmentation of urinary stones in non-contrast CT (NCCT) is challenging due to small lesions, class imbalance, and voxel sparsity. We propose a hierarchical fine-tuning framework based on VISTA3D with (1) anatomical mapping to define regions of interest and (2) organ-aware stone segmentation. On 119 test cases, the method achieved 95.69\% stone-level and 96.64\% patient-level sensitivity, demonstrating improved detection performance through anatomical context.

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 37