A 3D-optimized AI imaging model for the skin tumor burden assessment of cutaneous T-cell lymphoma

Summary: Abstract Body: Cutaneous T-cell lymphoma (CTCL) is characterized by widespread skin patches that may progress to plaques and tumors, necessitating precise tumor burden assessment for staging, prognosis, and treatment guidance. However, existing methods, including the widely accepted modified Severity Weighted Assessment Tool (mSWAT), present significant challenges in routine practice due to their time-consuming nature and inter-observer variability. This study developed an AI model, mSWAT-Net, to estimate mSWAT scores using clinical images of CTCL patients. The model integrates three segmentation sub-modules for lesion type, body region, and overlap area segmentation, each trained on annotated images using UNet or DeeplabV3+ backbones. Notably, the overlap area segmentation sub-module addressed errors related to double-counting overlap areas in photos captured from different angles, utilizing 3,904 annotated images generated from 61 3D human images. Subsequently, mSWAT-Net was trained and validated on clinical photos with performance evaluated against junior dermatologists, experienced specialists, and objective ground truth derived from 3D imaging of CTCL patients. The three sub-modules achieved Jaccard Indices of 0.669, 0.767, and 0.738 for lesion type, body region, and overlap area segmentation, respectively. Across 2,410 images from 131 imaging series, mSWAT-Net achieved intraclass correlation coefficients (ICCs) of 0.917 and 0.858 in internal and temporal validation, demonstrating comparable performance to experienced CTCL specialists and surpassing junior dermatologists (ICC: 0.917 vs. 0.777). When compared to a more convincing ground truth derived from 3D patient imaging, the model achieved an ICC of 0.812. Follow-up datasets further confirmed mSWAT-Net’s ability to closely align with bedside mSWAT scores, reliably tracking skin tumor burden and measuring treatment response. These findings highlight mSWAT-Net's potential as an accurate, automated tool for CTCL tumor burden assessment, patient follow-up, and remote consultations. Huizhong Wang¹, Zihao Liu², Haihao Pan¹, Tingting Jiang², Yang Wang¹ 1. Peking University First Hospital, Beijing, Beijing, China. 2. Peking University, Beijing, Beijing, China. Bioinformatics, Computational Biology, and Imaging