Overview of State-of-the-Art Learning-Based Classification Methods in Medical Imaging.

Area of Science:

• Medical Imaging AI
• Machine Learning in Healthcare
• Biomedical Engineering

Background:

• Medical image classification relies heavily on AI, with rapid advancements.
• Foundation models, vision-language models (VLMs), and efficient pretraining are transforming clinical utility.
• Established models are well-understood; focus is on state-of-the-art innovations.

Purpose of the Study:

• To provide a clinically grounded reference for AI-based image classifiers.
• To summarize state-of-the-art learning paradigms relevant to clinical translation.
• To guide biomedical engineers and clinicians in selecting and validating AI classifiers.

Main Methods:

• Review of current learning paradigms, contrasting classical machine learning (ML) and deep learning (DL).
• Emphasis on advances like medical foundation models, multimodal VLMs, and hybrid architectures.
• Discussion of modality-specific considerations (X-ray, CT, MRI, etc.) and deployment challenges.

Main Results:

• Summarizes key advances: foundation models, multimodal VLMs, hybrid CNN-transformer architectures, diffusion augmentation, self-supervised pretraining, federated learning, and efficient deployment.
• Addresses modality-specific challenges across various imaging types.
• Outlines persistent clinical challenges including data bias, rare conditions, annotation noise, explainability, calibration, and equitable performance.

Conclusions:

• AI classifiers are increasingly central to medical imaging, with rapid evolution.
• Model selection depends on image structure, annotation cost, and workflow integration.
• Addressing challenges like bias and explainability is crucial for equitable and effective clinical AI implementation.