1011 - An Intelligent Grading Model for Radiation Dermatitis Based on Image Recognition Technology

Purpose/Objective(s): Radiation dermatitis (RD) remains the most common side effect in radiation therapy (RT), with its assessment being a dynamic process. This study developed an intelligent grading model using image recognition technology to standardize RD severity classification. The primary objectives were to automate the grading of RD and validate its clinical utility in improving diagnostic consistency and efficiency compared to manual evaluation.

Materials/Methods: A dataset of 824 RD images was collected from four tertiary hospitals in Chongqing, China, between January and November 2024. Two nurses specialized in radiotherapy nursing independently annotated the images according to the RT Oncology Group (RTOG) acute radiation injury grading criteria. The latest object detection framework, YOLOv11, was then employed to develop an efficient medical image classification system by integrating multi-task architecture adaptation and transfer learning strategies. The model was validated on the RD dataset, and its robustness was assessed through five-fold cross-validation.

Results: The intelligent grading model achieved an overall accuracy of 84.14% in the test set. The Grade 1 classification model demonstrated excellent performance across multiple metrics, including sensitivity, positive predictive value (PPV), and negative predictive value (NPV), with an AUC value of 0.965. The Grade 2 classification model exhibited the highest specificity at 0.984, although it had a lower sensitivity of 0.312, resulting in an AUC value of 0.845. The Grade 3 classification model showed improvements in sensitivity and PPV, achieving an AUC value of 0.908. The Grade 4 classification model achieved a balanced performance in sensitivity and specificity, with an AUC value of 0.979, the highest among all levels.

Conclusion: The intelligent grading model for RD developed in this study demonstrated promising performance in enhancing the diagnostic accuracy and efficiency of RD assessment. This model has the potential to aid healthcare providers in making informed decisions for patient care and treatment planning. Future studies are warranted to further validate the model's performance in diverse clinical settings and explore its integration into clinical workflows.