Predicting isocitrate dehydrogenase status in glioma using hierarchical attention-based deep 3D multiple instance...

Area of Science:

• Neuro-oncology
• Medical Imaging Analysis
• Machine Learning in Medicine

Background:

• Isocitrate dehydrogenase (IDH) status is a critical prognostic biomarker for central nervous system tumors, influencing diagnosis and treatment.
• Current methods for determining IDH status in gliomas are invasive, relying on surgical biopsies.
• The 2021 WHO classification highlights the importance of IDH status in glioma management.

Purpose of the Study:

• To develop a non-invasive method for predicting IDH status in gliomas.
• To evaluate the efficacy of a novel Hierarchical Attention-Based Multiple Instance Learning (HAB-MIL) framework using preoperative MRI.
• To reduce the reliance on invasive surgical procedures for IDH status determination.

Main Methods:

• A retrospective cohort of 345 glioma patients from Xi'an Jiaotong University and 495 from the TCIA dataset were included.
• A Hierarchical Attention-Based Multiple Instance Learning (HAB-MIL) framework was developed, incorporating positional encoding for 3D lesion representation.
• Model performance was assessed using five-fold cross-validation, ROC curves, AUC, sensitivity, and specificity.

Main Results:

• The HAB-MIL framework achieved high performance with AUCs of 0.917 (TCIA) and 0.892 (Xi'an Jiaotong University).
• The model's performance is comparable to state-of-the-art methods on the TCIA dataset.
• Multiple instance learning demonstrates significant potential for accurate IDH prediction in gliomas.

Conclusions:

• The HAB-MIL framework enables IDH classification using conventional preoperative MRI, eliminating the need for pixel-level annotations.
• This approach significantly reduces the annotation burden for clinicians.
• The study presents a promising non-invasive tool for glioma IDH status assessment.