<\/p>\n
Shoppers of tech and clinicians alike are eyeing PRET, a new AI that recognises 18 cancer types and completes fresh pathology tasks using just a handful of annotated slides , no extra training required , which could matter in clinics with tight resources and heavy caseloads.<\/strong><\/p>\n Essential Takeaways<\/p>\n PRET borrows a trick from natural language processing called in\u2011context learning, and it uses small, annotated image patches as the \u201cexamples\u201d the model references when making decisions. That gives it a quiet elegance , instead of months of retraining, you show it a few labelled slides and it adapts on the fly, which is a breath of fresh air if you\u2019ve ever waited for a model to fine\u2011tune on a new dataset. According to the Hong Kong University of Science and Technology team, this method exploits fine\u2011grained local visual cues so the model can shift its answers without changing its parameters.<\/p>\n HKUST and partners tested PRET on 23 benchmark datasets from China, the US and the Netherlands, covering screening, tumour subtyping and segmentation tasks. The results were eye\u2011catching: perfect AUC for some screening tasks and near\u2011perfect segmentation scores for others. The team also noted robust performance when faced with data from different regions and resource settings, which matters if you’re thinking about deployment beyond major academic centres. The researchers say most validation data were freshly scanned and unavailable before the study, which reduces the risk of data leakage.<\/p>\n PRET is pitched as a plug\u2011and\u2011play diagnostic aid that could reduce the compute and manpower needed for routine AI deployment. For hospitals with limited AI teams, that\u2019s a meaningful shortcut: less infrastructure and fewer rounds of labelling for every small change in workflow. Still, it\u2019s not a magic wand. The team flagged limitations in telling apart tumours with very similar morphology, so pathologists would still need to review difficult or ambiguous cases. And crucially, PRET hasn’t been through clinical pilots or hospital rollouts yet, so real\u2011world integration questions remain.<\/p>\n PRET isn\u2019t HKUST\u2019s only play. The university has also developed mSTAR, a large language model assistant for pathology tasks, and SmartPath, which automates parts of the pathology workflow using extensive whole\u2011slide image training. Elsewhere in the region, institutions such as SingHealth are preparing to fold more AI tools into their services as digital pathology takes hold. Taken together, these efforts suggest a shift from experimental prototypes to toolkits hospitals can actually test in day\u2011to\u2011day practice.<\/p>\n If you run a pathology lab and want to explore PRET, start small: trial it on a single task such as screening for colorectal lesions where the model scored highly, and compare its outputs against your usual diagnostics. Use freshly scanned slides for validation to avoid hidden overlap with pre\u2011training data. Finally, involve practising pathologists early so they can flag morphologically tricky cases and help set sensible thresholds for when the AI\u2019s opinion should trigger human review.<\/p>\n It’s a small change that could speed up diagnoses and ease resource pressure , but expect careful pilots before full hospital adoption.<\/p>\n\n
What exactly is PRET and why it feels different<\/h2>\n
How the developers proved the point<\/h2>\n
Why clinics might actually use it , and what to watch for<\/h2>\n
How PRET fits into a bigger AI pathology picture<\/h2>\n
Practical tips for labs and pathologists curious about PRET<\/h2>\n
Source Reference Map<\/h3>\n