Survival Prediction with Clinical and Image Data via Transfer Learning in Head and Neck Squamous Cell Carcinoma

Abstract

Head and neck squamous cell carcinoma (HNSCC) is a type of cancer that exclusively affects the mucous membranes in the head and neck area. Accurate prognostication of cancer patients’ survival is essential for various therapeutic objectives, including early disease detection, treatment planning, risk assessment, follow-up care provision, patient counselling, and improving the quality of healthcare. The integration of clinical and imaging data to develop predictive models enables the enhancement of accuracy and clinical utility in survival prediction, ultimately resulting in improved patient treatment and outcomes.Our research focuses on the development of a predictive model that leverages transfer learning techniques to utilize clinical and imaging data obtained from TCIA for predicting the survival rates of patients with HNSCC. Transfer learning allows us to harness pre-trained models, significantly enhancing the efficiency and accuracy of feature extraction and prognostic predictions. Specifically, we employed a manual feature selection technique to assess clinical data, while CapsuleNet was utilized to extract features from imaging data. These features were then combined and inputted into DenseNet 121, a deep learning model fine-tuned through transfer learning, to generate prognostic predictions for patients diagnosed with HNSCC.The proposed model’s performance was evaluated using metrics such as precision, sensitivity, accuracy, and F1 Score. Our analysis demonstrates that the transfer learning-based model achieved an accuracy rate of 98.1%, surpassing the performance of existing models, including CNN and ResNet50, which were assessed using same dataset. This highlights the significant potential of transfer learning in improving prognostic predictions and ultimately enhancing patient outcomes.