GAMMIL: A graph attention-guided multi-scale fusion multiple instance learning model for the WHO grading of meningioma in whole slide images

Abstract

Background and objective

The WHO grading of meningioma is closely linked to patient treatment and prognosis, making the development of an accurate deep learning model based on whole slide images (WSI) of significant clinical importance. Currently, deep learning-based multiple instance learning (MIL) often focuses on single-scale image information, which can result in information loss, or it fuses all features across different scales, potentially introducing noise from non-key regions in multi-scale images. To overcome the above limitations, we propose a novel MIL model.

Methods

In this study, we proposed a Graph Attention-guided Multi-scale fusion Mutiple-Instance Learning (GAMMIL) model. It consists of three channels: 10×, 40×, and fused. In the 40 × channel, it adaptively selects key patches, instead of all patches for information fusion, avoiding the interference of non-key regions and the waste of computing resources. Further, the graph attention-guided module extracts global information across multi-scale images for better information fusion. Finally, the fused features are used to predict the WHO grade of meningioma.

Results

A total of 428 meningioma WSIs were collected from Guizhou Provincial People’s Hospital. The experimental results demonstrate that the GAMMIL model achieved an AUC of 0.911, surpassing other state-of-the-art MIL models. Additionally, the GAMMIL model was validated on two publicly available datasets (TCGA-RCC and CAMELYON16) and a private multi-center clear cell renal cell carcinoma dataset, where it exhibited superior discriminative performance compared to other leading models.

Conclusions

The GAMMIL model effectively simulates the pathologist’s diagnostic process, showing strong classification performance across datasets. Its adaptive patch selection and multi-scale fusion offer valuable insights for cancer pathology.