Classification of Volumetric Images Using Multi-Instance Learning and Extreme Value Theorem.

Area of Science:

• Medical Imaging Analysis
• Computational Pathology
• Artificial Intelligence in Medicine

Background:

• Volumetric imaging is crucial for medical diagnosis.
• Convolutional Neural Networks (CNNs) for volumetric image analysis face challenges due to limited annotated data and GPU memory constraints.
• Existing methods struggle with the detailed annotation requirements for volumetric data.

Purpose of the Study:

• To propose a novel method for volumetric image classification that addresses data and memory limitations.
• To adaptively select positive instances from positive bags during training.
• To improve the efficiency and performance of deep learning models in medical image analysis.

Main Methods:

• Volumetric image classification framed as a multi-instance classification problem.
• Novel adaptive positive instance selection using extreme value theory to model feature distributions.
• Method validated on retinal OCT, pulmonary 3D-CT (emphysema detection), and 2D histopathology (cancer detection) datasets.

Main Results:

• The proposed method achieved performance comparable to fully supervised methods across all tested tasks.
• State-of-the-art performance was reached in classifying retinal OCT images for fluid build-up, detecting emphysema in pulmonary CT scans, and identifying cancerous regions in histopathology images.
• The technique effectively models feature distributions to identify pathologies without requiring extensive local annotations.

Conclusions:

• The proposed multi-instance learning approach offers a viable solution for volumetric image classification with limited annotated data.
• Extreme value theory-based instance selection enhances classifier performance and generalizability.
• This method presents a significant advancement for AI-driven medical diagnostics, achieving expert-level accuracy.