Weakly supervised deep learning-based classification for histopathology of gliomas: a single center experience

Area of Science:

• Neuropathology
• Artificial Intelligence
• Oncology

Background:

• Accurate and prompt histopathological diagnosis of tumors is crucial.
• Artificial intelligence (AI) systems are increasingly used for tumor diagnosis.
• Weakly supervised deep learning (WSDL) offers potential for enhancing diagnostic capabilities.

Purpose of the Study:

• To investigate the efficacy of WSDL in aiding glioma diagnosis.
• To develop and evaluate a WSDL model for classifying glioma types and grades.
• To assess the model's ability to infer IDH status.

Main Methods:

• Whole slide images (WSIs) from West China Hospital (WCH) and The Cancer Genome Atlas (TCGA) were analyzed.
• WSIs were processed using OpenSlide and DeepZoom, with color normalization via Reinhard method.
• A WSDL model (ResNet-50 with attention) was developed and evaluated using ten-fold cross-validation and AUC calculations.

Main Results:

• The WSDL model achieved high AUC values for differentiating glioma grades and types in both TCGA and WCH datasets.
• Specific AUCs included differentiating astrocytomas, oligodendrogliomas, all gliomas, and glioma types.
• The model showed strong performance in inferring IDH status from TCGA data (AUC = 0.9488).

Conclusions:

• WSDL is an effective and reliable tool for neuropathological diagnosis of gliomas.
• The developed AI model serves as a valuable auxiliary tool for clinicians.
• This approach demonstrates significant potential for improving the accuracy and efficiency of glioma diagnosis.