Deep Learning-Based Framework for the Assessment of Radiation Dermatitis in Nasopharyngeal Carcinoma (NPC) Patients
R Ni1*, G Ren1, V Tam1, Z Dai2, X Wang3, S Lee1, J Cai1, (1) The Hong Kong Polytechnic University, Hong Kong, China, (2) Guangdong Provincial Hospital of Chinese Medicine, Guangzhou, 44, CN, (3) The Second Affiliated Hospital of Guangzhou University of Chinese Medicine, Guangzhou, 44, CN
Presentations
TU-EF-TRACK 4-5 (Tuesday, 7/27/2021) 3:30 PM - 5:30 PM [Eastern Time (GMT-4)]
Purpose: Radiation dermatitis (RD) is a common, unpleasant side-effect of patients receiving radiotherapy. Its severity is graded manually through visual inspection in clinical practice, which is resource-demanding and leads to inter-rater variation. To overcome this difficulty, this study aims to develop an end-to-end automatic RD grading framework based on deep learning (DL) techniques.
Methods: A dataset of 1120 photographs of the head and neck region was built for NPC patients undergoing radiotherapy. These photographs were graded by qualified assessors based on Radiation Therapy Oncology Group (RTOG) guidance. We further classified RD Grade 0&1 as mild cases and RD Grade 2 and above as severe cases. An end-to-end RD grading framework was developed by combining DL-based segmentation and RD severity classification models. Models were trained separately for the front and side view. First, the neck region was segmented from the camera-captured photographs via U-Net. Secondly, a convolutional neural network (CNN) classifier (VGG16-BN with pretrained) was applied to the segmented region for RD severity classification. The performance of the proposed DL methods was qualitatively and quantitatively evaluated for both segmentation and classification parts.
Results: Preliminary results of segmentation showed that the averaged Dice Similarity Coeffcients (DSCs) were 91.2% and 90.8% for front and side view, respectively. For RD severity classification, the overall accuracy of 111 test photographs was 93%. Our method accurately classified 94% (82/87) of the mild cases and 92% (22/24) of the severe cases. Mild cases were predicted with precision of 0.98, recall of 0.94 and F1-score of 0.96; severe cases with precision of 0.81, recall of 0.92 and F1-score of 0.86.
Conclusion: We demonstrated a DL-based end-to-end framework for automatic grading of RD severity in NPC patients. This method holds great potential to facilitate the monitoring and management of RD in NPC patients and improve patients’ quality of life.
Handouts
Keywords
Not Applicable / None Entered.
Taxonomy
TH- Response Assessment: Modeling: Machine Learning
Contact Email