Papillary thyroid carcinoma lymph node metastasis prediction method based on Transform-MIL

Abstract

The invention discloses a papillary thyroid carcinoma lymph node metastasis prediction method based on Transform-MIL. The method comprises the following steps: S1, extracting the characteristics of patch in WSI by using a lightweight ViT network; s2, selecting a key path by adopting a clustering-based strategy; s3, constructing a Transform-MIL model, learning relationships among instances from multiple aspects through a multi-head self-attention mechanism, and embedding instance-level features into packet representation; s4, combining a thyroid papillary set and a lymph node metastasis data set, and helping a Transform-MIL model to learn and predict lymph node metastasis through mutual knowledge distillation; according to the method, the Transform-MIL model is constructed, the instance-level features are better embedded into the packet representation, the morphological similarity between the tumor cells and the lymph node metastasis cells is fully utilized, and the knowledge of the relationship between the two data sets is transmitted by taking the attention map as a medium, so that the prediction accuracy of the lymph node metastasis histopathology image is improved.

Description

technical field [0001] The invention relates to the field of biotechnology, in particular to a method for predicting lymph node metastasis of thyroid papillary carcinoma based on Transformer-MIL. Background technique [0002] For cancer patients, lymph node metastasis determines the extent of their lymph node dissection and is one of the main independent prognostic factors. Accurately predicting the status of lymph nodes in cancer patients before surgery is of great significance to avoid overtreatment and reduce postoperative complications. Many studies have shown that preoperative CT radiomics can help realize individualized prediction of lymph node status in cancer patients, but these studies often use tumor radiomics features or combine a small number of clinicopathological features, such as lymph node status in CT reports , serum biomarkers, TNM staging, etc. [0003] As histopathological image analysis has become an important means of tumor diagnosis, more and more hi...

AI Technical Summary

Problems solved by technology

However, effective deployment of MIL in the field of histopathology images is still a challenging problem nowadays, and currently proposed MIL methods usually select multiple instances with the highest and lowest scores based on instance prediction, but in the histopathology image analysis task , the number of positive and negative patches is extremely unbalanced, which leads to the fact that the standard multiple instance (SMI) assumption cannot be satisfied, and the extreme instances selected according to the prediction score cannot represent WSI well

Images