What is it about?
Since the discovery of X-rays 130 years ago, medical imaging has become essential for diagnosing diseases without surgery. This paper reviews how artificial intelligence is now being taught to analyze these scans, helping doctors identify conditions like tumors more efficiently. We trace the history of imaging technology, explain how computer systems learn to interpret medical data, and outline the steps needed to make these tools safe and reliable for everyday patient care.
Featured Image
Photo by Navy Medicine on Unsplash
Why is it important?
This review is timely, arriving in 2025 — the 130th anniversary of Roentgen's X-ray discovery — at a pivotal moment when medical imaging is transitioning from traditional deep learning to foundation models and diffusion-based approaches. What makes this work unique is its dual focus: we systematically trace the physical evolution of imaging modalities (X-ray, CT, MRI, PET, ultrasound) alongside the parallel evolution of computer vision architectures, explicitly connecting hardware advances to algorithmic adaptations. Two significant contributions distinguish this survey: a) we provide the first comprehensive catalog of 30+ publicly available medical imaging datasets with licensing terms, task descriptions, and state-of-the-art performance baselines in a single reference, and b) we critically evaluate emerging paradigms—including Segment Anything Model adaptations, diffusion probabilistic segmentation, and hybrid CNN-Transformer architectures—highlighting not just their accuracy gains but also their clinical deployment challenges, such as uncertainty calibration and domain shift. By bridging the gap between imaging physics, clinical workflow constraints, and machine learning methodology, this work offers a unified roadmap for researchers, clinicians, and engineers aiming to develop robust, trustworthy AI tools for real-world healthcare settings.
Perspectives
As one of the authors, writing this review felt like standing at a unique crossroads—honoring 130 years of medical imaging history while helping to chart its AI-driven future. What made this work personally meaningful was the opportunity to bridge two worlds I care deeply about: the clinical realities of diagnostic imaging and the rapidly evolving landscape of deep learning. I also hope this paper serves as a gentle corrective to the "hype cycle" sometimes surrounding medical AI. It was important to us to acknowledge not just what works, but what remains hard: annotation bottlenecks, domain shift in real-world deployment, and the ethical weight of algorithmic decisions in patient care. If a student reads this and feels empowered to tackle one of these open challenges—or if a clinician gains confidence in evaluating an AI tool for their practice—then the effort was worthwhile. Finally, on a personal note: my own research journey into hyperspectral imaging for pathogen identification began with questions this survey tries to answer—Which modality fits which task? Which architecture generalizes beyond the lab? I hope this paper saves others time, sparks new collaborations, and ultimately contributes to tools that make healthcare more precise, accessible, and human-centered. If it does even a little of that, I'll consider it a success.
Dr. Aleksandr Sinitca
Sankt-Peterburgskij gosudarstvennyj elektrotehniceskij universitet LETI
Read the Original
This page is a summary of: Computer Vision-Based Medical Imaging Techniques: Past, Present, and Future, IEEE Access, January 2026, Institute of Electrical & Electronics Engineers (IEEE),
DOI: 10.1109/access.2026.3654393.
You can read the full text:
Contributors
The following have contributed to this page
