Automated classification of polyps using deep learning architectures and few-shot learning.

Area of Science:

• Medical imaging analysis
• Artificial intelligence in gastroenterology
• Computational pathology

Background:

• Colorectal cancer (CRC) is a major global health concern, with colonoscopy being a key prevention method.
• Accurate classification of colon polyps is crucial for determining appropriate treatment, but manual classification can be challenging.
• Existing classification systems, such as NICE and Paris, aid in polyp characterization, but their application can be complex.

Purpose of the Study:

• To develop and evaluate two novel automated polyp classification systems to assist gastroenterologists.
• To improve the accuracy and efficiency of polyp classification based on established NICE and Paris criteria.
• To address the challenges of data scarcity in medical machine learning for polyp analysis.

Main Methods:

• Developed a two-step transformer network-based system for Paris classification, involving polyp detection, cropping, and classification.
• Designed a few-shot learning algorithm utilizing Deep Metric Learning for NICE classification, creating an embedding space for efficient classification with limited data.
• Implemented and validated both systems on relevant datasets, including public data for Paris classification.

Main Results:

• Achieved 89.35% accuracy for Paris classification, setting a new state-of-the-art benchmark.
• Attained 81.13% accuracy for NICE classification, demonstrating the effectiveness of few-shot learning in data-scarce scenarios.
• Provided system explainability through neural network activation heat maps and ablation studies.

Conclusions:

• Introduced two automated polyp classification systems designed to support gastroenterologists in clinical practice.
• Achieved state-of-the-art results for Paris classification and demonstrated the feasibility of few-shot learning for NICE classification.
• Addressed critical data scarcity issues in medical AI, paving the way for more robust diagnostic tools.