Click here to

Session: Deep Learning for Image Segmentation [Return to Session]

Deep Learning Method for Objective Segmentation of Radiation-Induced Dermatitis in Skin Photographs

Y Park1*, S Choi2, M Cho2, J Son2, M Han1, J Kim3, H Kim1, D Kim4, J Kim1, C Hong1, (1) Yonsei University College of Medicine, Seoul, , KR, (2) Yongin Severance Hospital, Yongin, ,KR, (3) Yonsei University College of Medicine, Gangnam Severance Hospital, Seoul, Seoul, KR, (4) Yonsei Cancer Center, Seoul, 11, KR,


TU-D1030-IePD-F2-2 (Tuesday, 7/12/2022) 10:30 AM - 11:00 AM [Eastern Time (GMT-4)]

Exhibit Hall | Forum 2

Purpose: To develop and validate the objective segmentation model for radiation-induced dermatitis (RID) using skin photographs and to investigate whether the use of skin-dose distribution features can improve the performance of that model.

Methods: Seventy-three RID datasets consisting of skin photographs and skin-dose distribution were obtained from patients treated radiotherapy. All skin photographs were acquired in the treatment room using a 3D depth camera. The skin-dose distribution was generated using radiotherapy plan files and converted into a 2D intensity map from rigid registration with skin photographs. The ground truth of RID segmentation was manually delineated from skin photographs by an experienced radiation oncologist and medical physicists, as faint or severe RID. A U-net architecture was constructed as segmenting model of RID. The model training was performed two times with different input constructions, with and without skin-dose distribution data. The dice similarity coefficient (DSC) and balanced accuracy were applied to evaluate the segmentation performance. Evaluation metrics of the two models were compared using a paired t-test.

Results: The optimal performance of the RID segmentation network was observed in the network trained with the skin-dose distribution. The average of DSC and balanced accuracy in the segmentation model with skin-dose data were 0.58 and 0.77 in faint RID, and 0.67 and 0.79 in severe RID, respectively, which is significantly higher than evaluation metrics of the model without skin-dose distribution.

Conclusion: This study indicates the potential of objective diagnosis for RID by segmenting skin regions using a deep learning model. The dosimetric factor of the skin region could aid in improving the performance of the RID segmentation model whereas the correlation between skin dose distribution and the morphology of RID needs to be analyzed in further study.

Funding Support, Disclosures, and Conflict of Interest: This work was supported by the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (No. NRF-2020R1C1C1005713).


Segmentation, Computer Vision


IM/TH- Image Segmentation Techniques: Machine Learning

Contact Email