Prediction of the short-term efficacy and recurrence of photodynamic therapy in the treatment of oral leukoplakia based on deep learning

Background: The treatment of oral leukoplakia (OLK) with aminolaevulinic acid photodynamic therapy (ALAPDT) is widespread. Nonetheless, there is variation in efficacy. Therefore, this study constructed a model for predicting the short-term efficacy and recurrence of OLK after ALA-PDT. Methods: The short-term efficacy and recurrence of ALA-PDT were calculated by statistical analysis, and the relevant influencing factors were analyzed by Logistic regression and COX regression model. Finally, prediction models for total response (TR) rate, complete response (CR) rate and recurrence in OLK patients after ALA-PDT treatment were established. Features from pathology sections were extracted using deep learning autoencoder and combined with clinical variables to improve prediction performance of the model. Results: The logistic regression analysis showed that the non-homogeneous (OR: 4.911, P: 0.023) OLK and lesions with moderate to severe epithelial dysplasia (OR: 4.288, P: 0.042) had better short-term efficacy. The area under receiver operating characteristic curve (AUC) of CR, TR and recurrence predict models after the ALA-PDT treatment of OLK patients is 0.872, 0.718, and 0.564, respectively. Feature extraction revealed an association between inflammatory cell infiltration in the lamina propria and recurrence after PDT. Combining clinical variables and deep learning improved the performance of recurrence model by more than 30 %. Conclusions: ALA-PDT has excellent short-term efficacy in the management of OLK but the recurrence rate was high. Prediction model based on clinicopathological characteristics has excellent predictive effect for short-term efficacy but limited effect for recurrence. The use of deep learning and pathology images greatly improves predictive value of the models.