Related Experiment Videos
A deep learning framework for recognizing skin changes secondary to chronic venous insufficiency in clinical
Ko Eun Kim1, Yerin Lee2, Hyunyoung Kang2
1Department of Dermatology, Korea University Guro Hospital, Seoul, Republic of Korea.
Background:
Chronic venous insufficiency (CVI) produces heterogeneous lower-extremity skin changes that often mimic inflammatory dermatoses, complicating the differentiation between venous etiologies and conditions requiring dermatologic care. Coexisting venous signs further confound visual interpretation, leading to diagnostic variability. To address this unmet clinical need, we developed and externally validated a pose-guided, high-resolution deep learning pipeline to classify CVI-related skin lesions from routine clinical photographs.
Materials And Methods:
A deep learning framework was developed using a multicentre dataset (8672 images) from Korea University Guro Hospital, Chung-Ang University Gwangmyeong Hospital and Heartwell Vein Clinic. The pipeline integrated Sapiens-based pose estimation for limb localization and a U-Net model refined with Segment Anything Model 2 (SAM2) for skin segmentation. Multiple backbone architectures were evaluated, and ResNet-34 was selected based on optimal performance, taking refined masks and RGB images as input. Performance was evaluated on an internal test set, an independent clinical dataset (n = 183), and the Fitzpatrick17k dataset.
Results:
ResNet-34 emerged as the optimal architecture, achieving an internal AUROC of 0.922 with balanced sensitivity (0.679) and specificity (0.943) at 1024 × 1024 resolution. Pose-guided skin segmentation improved discrimination compared to using raw images (AUROC 0.920) or skin masking alone (AUROC 0.905). Multicentre training enhanced generalizability, yielding an external AUROC of 0.877 and maintaining high performance on the Fitzpatrick17k dataset (sensitivity 0.918).
Conclusions:
Our pose-guided, high-resolution pipeline enables robust, objective detection of CVI-related skin lesions. By demonstrating high generalizability across multicentre and public datasets, this framework serves as a reliable decision-support tool to standardize clinical phenotyping and facilitate consistent triage and management of chronic venous disease across diverse care settings.
