# Image Feature Fusion of Hyperspectral Imaging and MRI for Automated Subtype Classification and Grading of Adult Diffuse Gliomas According to the 2021 WHO Criteria

## TL;DR

This paper introduces a new method combining hyperspectral imaging and MRI to accurately classify and grade adult diffuse gliomas according to the latest WHO criteria.

## Contribution

The novel Hyperspectral Attention Fusion Network (HAFNet) integrates HSI and MRI data for improved glioma diagnosis.

## Key findings

- HAFNet achieved a macro-averaged AUC of 0.9886 and 98.66% classification accuracy.
- Multimodal integration outperformed HSI-only models with a significant p-value (<0.001).
- MRI radiomic features enhance discrimination beyond spectral information alone.

## Abstract

Background: Current histopathology- and molecular-based gold standards for diagnosing adult diffuse gliomas (ADGs) have inherent limitations in reproducibility and interobserver concordance, while being time-intensive and resource-demanding. Although hyperspectral imaging (HSI)-based computer-aided pathology shows potential for automated diagnosis, it often yields suboptimal accuracy due to the lack of complementary spatial and structural tumor information. This study introduces a multimodal fusion framework integrating HSI with routinely acquired preoperative magnetic resonance imaging (MRI) to enable automated, high-precision ADG diagnosis. Methods: We developed the Hyperspectral Attention Fusion Network (HAFNet), incorporating residual learning and channel attention to jointly capture HSI patterns and MRI-derived radiomic features. The dataset comprised 1931 HSI cubes (400–1000 nm, 300 spectral bands) from histopathological patches of six major World Health Organization (WHO)-defined glioma subtypes in 30 patients, together with their routinely acquired preoperative MRI sequences. Informative wavelengths were selected using mutual information. Radiomic features were extracted with the PyRadiomics package. Model performance was assessed via stratified 5-fold cross-validation, with accuracy and area under the curve (AUC) as primary endpoints. Results: The multimodal HAFNet achieved a macro-averaged AUC of 0.9886 and a classification accuracy of 98.66%, markedly outperforming the HSI-only baseline (AUC 0.9267, accuracy 87.25%; p < 0.001), highlighting the complementary value of MRI-derived radiomic features in enhancing discrimination beyond spectral information. Conclusions: Integrating HSI biochemical and microstructural insights with MRI radiomics of morphology and context, HAFNet provides a robust, reproducible, and efficient framework for accurately predicting 2021 WHO types and grades of ADGs, demonstrating the significant added value of multimodal integration for precise glioma diagnosis.