Grad-CAM based deep learning analytics for image-level colon disease classification based on graph neural networks...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate colonoscopic image classification is crucial for early colorectal disease detection.
• Deep learning models, including transformers and graph neural networks (GNNs), offer advanced methods for analyzing image data.
• These models can capture global context and relational structures vital for disease characterization.

Purpose of the Study:

• To evaluate transformer-based and graph-based deep learning frameworks for endoscopic colon disease classification.
• To compare the performance of Vision Transformers (ViT) against CNN-GNN pipelines.
• To assess the effectiveness of different graph construction strategies within GNNs.

Main Methods:

• Experiments utilized the Kvasir V2 dataset.
• A Vision Transformer (ViT) was selectively fine-tuned.
• A CNN-GNN pipeline integrated ViT embeddings with various graph construction methods (cosine similarity, k-NN, epsilon-radius) and GNN architectures.
• Performance metrics included accuracy, precision, recall, and macro-F1 score; Grad-CAM was used for interpretability.

Main Results:

• The fine-tuned Vision Transformer achieved 94.6% accuracy and a 0.94 macro-F1 score.
• The optimal graph-based approach, using ViT embeddings with an epsilon graph and GIN aggregation, reached 92% accuracy and a 0.92 macro-F1 score.
• Both transformer and graph-based methods demonstrated strong classification capabilities.

Conclusions:

• Transformer-based models show superior discriminative power for colon disease classification.
• Graph-based relational modeling provides competitive results when utilizing high-quality image embeddings.
• These findings highlight the potential of advanced deep learning techniques in improving endoscopic diagnostics.