Exploring Deep Learning and Transfer Learning for Colonic Polyp Classification.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep Learning (DL) and Convolutional Neural Networks (CNNs) excel at feature extraction for pattern classification.
• Automated endoscopic image classification, specifically for colonic polyps, faces challenges due to limited annotated datasets.
• CNNs offer a powerful approach for analyzing complex medical images.

Purpose of the Study:

• To investigate the effectiveness of DL, specifically CNNs, for automated colonic polyp classification.
• To compare training CNNs from scratch versus using pre-trained CNN architectures.
• To evaluate the utility of combining traditional features with CNN-derived features.

Main Methods:

• Explored various CNN configurations for training from scratch.
• Tested distinct pre-trained CNN architectures on 8 HD-endoscopic image databases.
• Compared DL-based features against commonly used classical features for colonic polyp classification.

Main Results:

• Features learned by CNNs trained from scratch are highly relevant for colonic polyp classification.
• "Off-the-shelf" CNN features also demonstrate significant relevance for automated classification.
• Combining classical and "off-the-shelf" CNN features further improves classification performance.

Conclusions:

• Deep learning models, including CNNs trained from scratch and pre-trained models, are effective for automated colonic polyp classification.
• CNN-derived features are valuable for endoscopic image analysis.
• Hybrid approaches integrating classical and CNN features offer enhanced accuracy for colonic polyp detection.