Annotation-free glioma grading from pathological images using ensemble deep learning.

Area of Science:

• Neuro-oncology
• Computational pathology
• Artificial intelligence in medicine

Background:

• Glioma grading is crucial for treatment decisions, but differentiating between grades II and III remains a pathological challenge.
• Single deep learning models exhibit limited accuracy for fine-grained glioma classification.
• Annotation-free grading of pathological images is highly desirable.

Purpose of the Study:

• To develop and evaluate an ensemble deep learning (DL) model for annotation-free glioma grading, specifically distinguishing between grades II and III.
• To improve the accuracy and reliability of glioma grading compared to traditional single DL models.

Main Methods:

• Established multiple tile-level DL models using the ResNet-18 architecture.
• Developed ensemble DL models by combining component classifiers for patient-level grading.
• Utilized logistic regression (LR) with a 14-component DL classifier (LR-14) for the final ensemble model.
• Included whole-slide images from 507 low-grade glioma (LGG) subjects from The Cancer Genome Atlas (TCGA).

Main Results:

• The 30 individual DL models achieved an average area under the curve (AUC) of 0.7991 for patient-level glioma grading.
• The ensemble LR-14 model demonstrated a mean patient-level accuracy of 0.8011 and an AUC of 0.8945.
• The ensemble approach significantly reduced model variation, indicated by a median between-model cosine similarity of 0.9524.

Conclusions:

• The proposed LR-14 ensemble DL model achieves state-of-the-art performance for glioma grade II and III classification.
• This annotation-free method offers a promising advancement in pathological image analysis for neuro-oncology.
• Ensemble learning effectively enhances the diagnostic accuracy of deep learning in challenging histopathological grading tasks.