Abstract:

In the decision-making process for thyroid cancer surgery, the accurate preoperative assessment of lymph node metastasis poses a challenging issue. To minimize unnecessary surgeries and enhance patient quality of life, precise prediction of lymph node metastasis in thyroid cancer is of practical significance. In this study non-negative latent factor model and PEFT(Parameter Efficient Fine Tuning) technique are used to solve the problem of small scale of medical data and missing clinical data. The non-negative latent factor model was used to complete the clinical data to improve the reliability and accuracy of the data. By introducing PEFT technology to fine-tune large pre-trained models, the computational cost is significantly reduced. The results show that the latent factor model is superior to the traditional method under different missing proportions, and the PEFT method has higher training accuracy and lower training time on two different data sets. By comparing the comprehensive performance of local data set and public data set, the effectiveness of the proposed method is verified. The proposed method reduces the computational cost and has higher interpretability while maintaining high prediction accuracy, and provides an efficient and feasible scheme for the application of pre-trained large models in medical tasks.

Key words: latent factor model, efficient parameter fine-tuning, PEFT, medical image processing, data filling, transfer learning, lack of medical data, multimodal