Main Session
Sep
28
PQA 04 - Breast Cancer, Patient Reported Outcomes/QoL/Survivorship, Functional Radiation Medicine, Hematologic Malignancies, Palliative Care, and International/Global Oncology
2871 - Deep Learning-Based Cross-Center Ordinal Grading of Radiation Dermatitis In Breast Cancer Radiotherapy
Presenter(s)
Zihan Shi, - Xijing Hospital, Xian 710032, Shaanxi
Z. Shi1, Q. Huang1, H. Sun2, and L. Zhao3; 1Xijing Hospital, Xian 710032, Shaanxi, China, 2Department of Radiation Oncology, Xijing Hospital, Air Force Medical University, xi'an, Shaanxi, China, 3Department of Radiation Oncology, First Affiliated Hospital of Air Force Medical University, Xi'an, Shaanxi, China
Purpose/Objective(s):
Radiation dermatitis is the most common acute toxicity during breast cancer radiotherapy, and grading remains physician-dependent and potentially subjective. We developed and externally validated a deep learning model for automatic ordinal grading of radiation dermatitis.Materials/Methods:
A total of 833 female breast cancer patients were retrospectively included, and 8,077 clinical skin photographs were annotated according to the RTOG criteria (grades 0–2). An EfficientNet-V2-L backbone was used for feature extraction, and the task was formulated as an ordinal regression problem under joint multi-objective supervision. The dataset was randomly split into training and internal test sets at an 8:2 ratio, and the model’s performance was compared against manual grading on the test set. Subsequently, external validation was performed on an independent cohort.Results:
On the internal test set, the model achieved an overall accuracy of 96.86%, with class accuracies of 97.97%, 96.93%, and 95.51% for grades 0, 1, and 2, respectively, and a recall of 95.51% for grade =2 dermatitis. In an external cohort predominantly composed of grade 2 cases, overall accuracy was 88.28% with a grade 2 recall of 90.74%. In a head-to-head comparison of 300 cases per rater, AI accuracy reached 98.33%, compared with 80.33%, 88.67%, and 83.33% for three physicians.Conclusion:
The proposed ordinal deep learning model demonstrates high internal accuracy, robust external performance despite class imbalance, and superior consistency compared with physician assessment, supporting its potential role in standardized toxicity monitoring during breast radiotherapy.