Ensemble Deep Learning Model for Glioma Histopathological Image Classification

Overview of Glioma

  • Glioma Definition: A primary brain tumor derived from glial cells.
  • WHO Classification: Gliomas are graded I to IV, with grade IV categorized as Glioblastoma Multiforme (GBM).
  • Cancer Genome Atlas (TCGA): Provides data on diffuse glioma, divided into two subtypes:
    • GBM: Grade IV
    • LGG: Grades II and III

Importance of Histopathological Images

  • Gold Standard in Diagnosis: Histopathological images are essential for accurate diagnosis in computational pathology.
  • Challenges: Human evaluation of these images can be subjective, necessitating automatic image analysis to improve accuracy and efficiency.

Deep Learning for Image Classification

  • Advancements: Significant research in using deep learning to analyze histopathological images, particularly through convolutional neural networks (CNNs) and more recently, transformer-based models.
  • Feature Extraction: Deep learning models can capture complex features from histopathological images, aiding in automatic classification.
    • Whole Slide Images (WSIs): Difficult to represent with traditional models, hence deep learning has shown great promise.

Proposed Method

  • Ensemble Deep Learning: Combining multiple deep learning models for classification tasks.
  • Application: Focused on binary classification for glioma subtypes (GBM vs. LGG).
  • Experimental Data: Utilized 22,592 patch images from TCGA for model training and testing.
  • Results:
    • CNN-based accuracy: 0.959 (AUC = 0.988)
    • Transformer-based accuracy: 0.979 (AUC = 0.997)
    • Demonstrated significant performance improvements over single deep learning models.

Methodology Steps

  1. Feature Extraction: Using various deep learning models to obtain unique features from histopathological images.
  2. Feature Selection: Importance of features was evaluated using Random Forest to prevent dimensionality issues while maintaining critical features.
  3. Classification: Machine learning models (SVM, LightGBM, MLP) used for final classification based on selected features.
    • Example Results:
      • MLP achieved better AUC metrics than individual models, indicating the efficiency of the ensemble approach.

Experimental Features and Models

  • Dataset: High-resolution WSI images of diffuse glioma patients analyzed.
  • Preprocessing: Images divided into 256x256 patch images to handle resolution challenges.
  • Deep Learning Models: Includes CNN and transformer-based architectures like ResNet, VGG, BEiTV2, and Swin Transformer.

Results Summary

  • Top Performing CNN Models:
    • VGG11 bn
    • ResNet50d
    • ConvNeXtV2
  • Top Performing Transformer Models:
    • GCViT
    • BEiTV2
    • Swin TransformerV2

Conclusion

  • An ensemble deep learning approach significantly enhances classification accuracy for glioma histopathological image classification via improved feature extraction and selection methods.
  • Future Work: Analyze multimodal datasets beyond histopathological images to enhance diagnostic capabilities.