Biomimetic Transfer Learning-Based Complex Gastrointestinal Polyp Classification.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Investigates Artificial Intelligence (AI) for automatic gastrointestinal (GI) polyp classification in endoscopic images.
• Focuses on biomimetic convolutional neural networks (CNNs) and Transfer Learning for enhanced diagnostic accuracy.
• Aims to support early detection of colorectal cancer.

Purpose of the Study:

• To evaluate the efficacy of various CNN architectures for classifying GI polyps.
• To assess the performance of optimized ResNet50, DenseNet121, and MobileNetV2 models.
• To determine the real-time applicability and diagnostic support potential of AI models.

Main Methods:

• Utilized the Kvasir dataset (4000 annotated endoscopic images, 8 polyp categories).
• Pre-processed images using normalization, resizing, and data augmentation.
• Trained and evaluated ResNet50, DenseNet121, and MobileNetV2 CNN models using standard performance metrics.

Main Results:

• ResNet50 achieved the highest validation accuracy (90.5%), followed by DenseNet121 (87.5%) and MobileNetV2 (86.5%).
• Models demonstrated good generalization with minimal training-validation accuracy differences.
• Average inference time was under 0.5 seconds, indicating real-time potential. Confusion matrix analysis revealed challenges with visually similar polyp classes.

Conclusions:

• Deep learning-based CNN architectures combined with Transfer Learning effectively classify endoscopic images.
• AI models show significant potential in supporting medical diagnostics for GI polyps.
• Model-assisted diagnostics can help overcome challenges in distinguishing subtle features in gastrointestinal imagery.