Computer-Aided Diagonosis for Colorectal Cancer using Deep Learning with Visual Explanations.

Area of Science:

• Gastroenterology
• Artificial Intelligence
• Medical Imaging

Background:

• Colorectal cancer prevention relies on detecting and removing neoplasms.
• Current endoscopic imaging techniques are effective but operator-dependent.
• Deep learning offers a potential solution to improve diagnostic consistency.

Purpose of the Study:

• To develop and evaluate a computer-aided diagnostic (CAD) system for colorectal adenoma detection using deep learning.
• To compare the diagnostic performance of the CAD system against expert endoscopists.

Main Methods:

• Collected and categorized 3000 colonoscopic images (normal, low-grade dysplasia, high-grade dysplasia, adenocarcinoma).
• Implemented and fine-tuned three convolutional neural networks (CNNs): Inception-v3, ResNet-50, and DenseNet-161.
• Utilized transfer learning, data augmentation, and 10-fold cross-validation; compared CNN-CAD performance to endoscopist groups.

Main Results:

• The CNN-CAD system achieved a 92.48% classification accuracy.
• The CNN-CAD system demonstrated superior performance across all criteria compared to expert endoscopists.
• Class Activation Mapping (CAM) visualizations confirmed the model's focus on relevant pathological regions.

Conclusions:

• Deep learning-based CAD systems can effectively distinguish colorectal adenoma pathology.
• The developed CNN-CAD system shows potential to enhance diagnostic accuracy in colonoscopy, exceeding human expert performance.
• This technology could significantly improve colorectal cancer prevention strategies.