Di-san junyi daxue xuebao (May 2022)
Development of intelligent delineation model for clinical target volume of nasopharyngeal carcinoma based on integrated clinical health care system
Abstract
Objective To established an intelligent delineation model of clinical target volume (CTV) for nasopharyngeal carcinoma (NPC) using artificial intelligence (AI) technology to improve the efficiency and the accuracy of CTV delineation among various radiotherapy institutes and clinicians with different technical levels. Methods Thirty NPC patients (20 males and 10 females, at a median age of 52 years) with early stage (T1~2N2) and small lymph nodes (≤ 2 cm) admitted in our Army Medical Center of PLA from September 2020 to May 2021 were collected in this study. CTV was delineated according to the International Guideline of CTV Delineation for NPC in 2017, and a deep learning model was trained to obtain AI+CTV model. The model was subsequently tested with manually segmented CTV in 10 NPC cases (7 males and 3 females, at a median age of 54 years) in our medical center and 6 cases (4 males and 2 females, at a median age of 46 years) from the teaching hospitals. Dice similarity coefficient (DSC), Hausdorff distance (HD), average symmetric surface distance (ASSD), maximum symmetric surface distance (MSSD) and relative absolute volume difference (RAVD) were analyzed to evaluate the feasibility and accuracy of AI delineation. Results Our established AI+CTV model obtained quite high accuracy, with an average DSC value of 0.865, and median DSC values of high-risk areas (CTV1) and low-risk areas (CTV2) of 0.886 and 0.842, respectively. The median HD, ASSD, MSSD and RAVD of CTV1 and CTV2 were all quite low, that is, 7.789 and 6.144 mm, 1.683 and 1.467 mm, 20.52 and 14.03 mm, 16.30% and 15.14%, respectively. The results of external verification showed that CTV1 delineation had higher consistency than CTV2. CTV1 of 6 patients and CTV2 of 5 patients were well delineated (DSC value was greater than 0.7) except CTV2-DSC value of 1 case was close to 0.7. Conclusion Our AI+CTV model based on small samples is basically feasible for the clinical application of CTV delineation in NPC patients with early stage (T1~2N2) and small lymph nodes (≤2 cm).
Keywords
