Multi-Site Pathological Images-Guided ViT for Oral Squamous Cell Carcinoma Recurrence Prediction

Overview

Background: Head and neck squamous cell carcinoma is one of the most common cancers. Postoperative recurrence is a risk factor for poor prognosis and decreased survival rate. There is a lag and passivity in the diagnosis and monitoring of postoperative recurrence in clinical diagnosis and treatment. The application of artificial intelligence to explore and develop a model for predicting postoperative recurrence of oral squamous cell carcinoma is expected to solve clinical difficulties and guide the formulation of postoperative monitoring and diagnosis and treatment plans. Materials and Methods： We recruited patients diagnosed with oral squamous cell carcinoma who had received surgical treatment, collected patient follow-up information and postoperative pathological images, enhanced and standardized pathological images, and extracted pathological features through visual transformation model (ViT). The features of pathological images were fused into a multi-layer perceptron model (MLP) for training, verification and testing, and the predictive performance of the model was evaluated by various indexes. Results: Among the 1325 patients enrolled, 275 relapsed, accounting for 20.8%. The optimized ViT-Small model has a validation AUC of 94.79% (90% accuracy) and a test AUC of 95.68% (91.5% accuracy), and outperforms other models on both validation and test sets. Conclusion: ViT-Small model have high predictive performance, which is expected to predict postoperative recurrence, guide the formulation of clinical diagnosis and treatment plan.

As the most common tumor in the world, the five-year relative survival rate of head and neck squamous cell carcinoma is about 68%, and the five-year relative survival rate of distant tumors is 40%. More than 60% of patients were diagnosed as locally advanced, the rate of postoperative recurrence was as high as 40%, and the five-year relative survival rate was less than 50%. Tumor metastasis and postoperative recurrence were risk factors for poor prognosis and decreased survival rate, and tumor recurrence occurred almost within 2 years after surgery. The treatment of locally advanced patients often includes postoperative chemoradiotherapy, immunotherapy, etc., and the treatment of early patients is often mainly surgical resection. The problem of over-treatment and missed diagnosis and treatment may occur when guiding postoperative treatment according to clinical or pathological stages, and the risk judgment of postoperative recurrence lacks objective basis and quantitative standard. In clinical diagnosis and treatment, whether adjuvant treatment is needed after surgery should be considered in combination with the risk of recurrence and pathological risk factors. For the monitoring and diagnosis of postoperative recurrence, regular follow-up of patients, imaging and pathological examination and clinical experience judgment of the attending physician are often required, which has lag and passivity. Prediction of postoperative recurrence, optimization of follow-up plan and timely development of personalized postoperative treatment will effectively reduce postoperative recurrence. The application of artificial intelligence in the medical field has provided non-invasive diagnosis and support for the diagnosis and screening of oral cancer. Previously, we deeply learned the clinical diagnosis and treatment information and follow-up data of oral cancer patients to establish a model to predict postoperative recurrence of oral cancer. MLP model can predict postoperative recurrence of oral cancer to a certain extent. In order to focus on postoperative pathology, this study establishes a postoperative recurrence prediction model by deep learning of pathological risk factors, which is expected to improve the prediction efficiency and simplify the process, thus guiding the monitoring plan.Deep learning pathological images and the establishment of postoperative recurrence prediction model of oral squamous cell carcinoma through digital pathology have high predictive performance, which is expected to predict postoperative recurrence, guide the formulation of clinical diagnosis and treatment plan, develop personalized treatment plan for patients, reduce postoperative recurrence and improve survival rate.