CW nnU-Net for Universal Lesion Segmentation Challenge on 3D Computed Tomography

Ching-Wei Wang, Ting-Sheng Su, Hikam Muzakkyi, Yu-Ching Lee

Keywords: Deep Learning, Oncological computed tomography, 3D Lesion Segmentation

Abstract:

3D lesion segmentation of oncological computed tomography (CT) is a crucial step in precisely monitoring changes in lesion/tumor growth, which enables the extraction of meaningful information from medical images, aiding in diagnosis, treatment planning and monitoring of diseases. In this research, we developed a highly efficient and effective CW nnU-Net and ensemble models for 3D lesion segmentation on CT for the Universal Lesion Segmentation (ULS) Challenge, which will be held jointly with 2024 medical imaging with deep learning (MIDL) conference at Paris, France. The proposed approach was built with a reasonably cheap Nvidia RTX 4080 GPU card and outperformed the baseline models in both development and test phase. In the final test phase, the proposed model ranks as the 3rd place among 577 participants worldwide, achieving a Challenge Score of 0.73, Segmentation DICE of 0.70 and Consistency DICE of 0.79. In the development phase, the proposed CW nnU-Net achieved a Challenge Score of 0.81, Segmentation DICE of 0.78 and Consistency DICE of 0.92.
For computational efficiency, CW nnU-Net takes only 3.25s for processing each VOI on the Grand Challenge platform server with a single T4 GPU and less than 2s using a local PC with RTX4080.

Submission Number: 159