RegSegNet: A Joint Registration Segmentation Network for Automatic Liver Segmentation from Non-contrast 3D SPECT-CT Images

Xuan Loc Pham, Manh Ha Luu, Hong Son Mai, Ngoc Ha Le, Bram van Ginneken, Alessa Hering

Keywords: Joint Registration-Segmentation, Non-contrast SPECT-CT, Liver Cancer

Abstract: 3D SPECT-CT images play a vital role in the treatment process for liver cancer. However, in many cases, the CT scan taken alongside SPECT is non-contrast, making the liver segmentation task a tough challenge for both AI models and radiologists. Previous methods often faced trade-offs between accuracy and runtime. This study introduces RegSegNet, a deep learning model that utilizes image registration to effectively guide the liver segmentation in non-contrast SPECT-CT images. The proposed method is trained and evaluated on a dataset consisting of 60 liver cancer patients. Experimental results show that RegSegNet significantly outperforms baseline methods in terms of runtime while maintaining comparable accuracy.

Submission Number: 59