Main article

Aisha Mensah*
Institute for Digital Health, King's College London, London, UK, SE1 9RT
aisha.mensah@kcl.ac.uk
Thomas Brückner
Department of Radiology, Charité – Universitätsmedizin Berlin, Berlin, Germany, 10117
Linh Nguyen
Faculty of Medicine, University of Sydney, Sydney, NSW, Australia, 2006

Abstract

The last decade has witnessed a transformation in medical image analysis that would have seemed improbable in 2012, when AlexNet's ImageNet performance first suggested that deep convolutional networks could match or exceed human-level perception on structured visual tasks. This review traces the arc of that transformation across four imaging modalities — radiology, pathology, ophthalmology, and dermatology — and examines how the field has matured from proof-of-concept demonstrations to clinically deployed systems. We identify three overlapping phases: a detection-dominated phase (2012–2017) focused on binary classification and lesion detection; a segmentation and quantification phase (2017–2021) that moved toward dense prediction and anatomical measurement; and an emerging integration phase (2021–present) characterised by multi-modal fusion, foundation models, and uncertainty-aware inference. For each phase, we assess the gap between benchmark performance and clinical utility, and examine the recurring obstacles — data scarcity, distributional shift, explainability demands — that have slowed translation. We conclude with a frank assessment of where the evidence genuinely supports clinical deployment and where enthusiasm has run ahead of rigour.

Article details

How to Cite

Mensah, A., Brückner, T., & Nguyen, L. (2023). From Pixels to Predictions: A Decade of Deep Learning in Medical Image Analysis. DATAMIND, 1(2), 1-4. https://doi.org/10.63646/