A Straightforward and Efficient PET Pipeline for Cross-Modality Abdominal Organ Segmentation

Changrui Li

Keywords: Abdominal organs segmentation, Unsupervised domain adaption, Style translation, PET scans

Abstract: Accurate multi-organ segmentation is essential for medical image analysis but remains challenging across modalities with varying image quality. In this work, we address cross-modality segmentation from CT to MRI and CT to PET, and design two tailored frameworks for these distinct tasks. On the public validation set, the CT-to-MR model achieves an average DSC of 79.56\% and NSD of 86.52\%, while the CT-to-PET model reaches DSC of 79.47\% and NSD of 64.44\%, demonstrating stable performance across modalities. A key contribution of this study is the PET segmentation pipeline, which adopts a straightforward yet effective design: unlabeled PET scans are first processed through a style translation module to reduce modality discrepancies, followed by direct segmentation using a dedicated SegNet. Unlike conventional semi-supervised strategies, this simplified pipeline reduces the number of training stages while still achieving strong segmentation accuracy. In addition, its streamlined structure offers notable computational advantages, with an average inference time of 10.34 seconds per case and GPU memory usage capped at 3.3 GB for all PET scans.

Submission Number: 15