Pathology AI: two tracks

Cancer pathology is splitting into two AI tracks: one aims to read many tumors from slide images alone, while the other aims to combine slides with lab, imaging and record data. (hkust.edu.hk, nature.com) Pathology is the practice of diagnosing disease by examining tissue and blood under a microscope. Digital pathology turns those glass slides into images that software can scan for patterns, counts and features that a human specialist would otherwise review manually. (news-medical.net, medprimetech.com) On April 21, 2026, the Hong Kong University of Science and Technology said its researchers built an AI pathology system that can recognize multiple cancer types from only a small number of samples and without additional model training. The university said the system is designed to avoid retraining each time it is adapted to a new cancer task. (hkust.edu.hk, eurekalert.org) That claim points at a long-running bottleneck in medical AI: models trained in one hospital, one scanner setup or one cancer type often need tuning before they work elsewhere. HKUST said its approach uses very few slides and still generalizes across cancers, a result the team framed as a way to make deployment more flexible. (hkust.edu.hk, eurekalert.org) A second track was laid out the same day in Nature Reviews Cancer, where authors argued that cancer AI needs more than one data stream. They said combining genomics, proteomics, imaging and clinical records gives a fuller picture of how tumors start, change and respond to treatment. (nature.com) In plain terms, that means moving from one photograph of a tumor to a dossier: what the cells look like, which genes are active, what proteins are present, what scans show and what happened to similar patients. The review said artificial intelligence is useful here because these datasets are high-dimensional and hard to interpret with standard tools. (nature.com) Recent reporting suggests the commercial market is advancing first on narrower tasks. On April 21, Medprime Technologies launched Cilika AI with an initial module for reticulocyte counting, automating the identification of immature red blood cells from digital blood-smear images rather than attempting broad cancer diagnosis. (biospectrumindia.com, moneycontrol.com) Zifo, a scientific informatics company, argued in an April 21 note that biopharma’s AI problem is often not the model but the missing scientific context around data. The company said fragmented, inconsistently managed information limits how well AI can support discovery, development and manufacturing. (prnewswire.co.uk, zifornd.com) Another April 20 report in breast cancer showed the middle ground between those two visions: AI systems that assist imaging, pathology and treatment decisions when paired with clinician review. That is closer to today’s deployable model than a fully general cancer system that works across institutions and data types out of the box. (news-medical.net, nature.com) The result is a field moving on two timelines at once. Research groups are pushing toward general, multimodal cancer AI, while labs and vendors are shipping narrower tools that save time on specific parts of the pathology workflow. (hkust.edu.hk, biospectrumindia.com, nature.com)