AI in oncology roadmap: pattern readers → foundation models → agents in the workflow

Overview

1. 1998-2016historic

   Computer-aided detection

   The first cleared cancer AI was computer-aided detection for mammography in 1998, which marked suspicious regions for the radiologist and, in large observational studies, did not improve accuracy. Rule-based decision support for treatment recommendations was tried and mostly abandoned. The lesson that survived: an algorithm has to be evaluated on the decision it changes, not on the pattern it finds.

   Hide era· 2 records

   AI in radiologyMammography & tomosynthesis
2. 2017-2022historic

   Deep learning reaches cleared devices

   Convolutional networks trained on labelled images matched specialists on narrow tasks. Paige Prostate (2021) became the first FDA-authorised AI for reading pathology slides; radiology triage tools for haemorrhage and embolism were cleared by the dozen; Sybil and Mirai predicted future lung and breast cancer from today's scan. Whole-slide scanning became routine in large centres, which made slide-level AI possible at all.

   Show era· 7 records
3. 2020-2026current

   Structure prediction and generative design

   AlphaFold made protein structure a lookup rather than a two-year experiment; AlphaFold 3 (2024), Boltz and Chai extended it to drug-protein and antibody complexes, and RFdiffusion and ESM3 design proteins that never existed. Insilico's generative chemistry produced the first AI-discovered drug to reach phase 2, Isomorphic's first oncology candidate was cleared for trials, and Recursion and Xaira are betting that image and perturbation data can find targets no hypothesis would. None has yet produced an approved cancer drug, which is the honest benchmark.

   Show era· 13 records
4. 2023-2026current

   Foundation models and the first predictive tests

   Pathology models pretrained on millions of slides (Virchow, Prov-GigaPath, UNI and CONCH, H-optimus, TITAN) predict mutations, biomarkers and outcomes from a routine stain; MUSK adds clinical text. ArteraAI Prostate (2025) was the first AI test cleared to predict benefit from a treatment, and ArteraAI Breast followed in 2026. MASAI gave the first randomised evidence that AI-supported screening finds more cancers with less workload; Aidoc CARE (January 2026) was the first foundation-model triage platform cleared. Single-cell models (Geneformer, scGPT, State) and the Tahoe-100M dataset began the same arc for biology.

   Show era· 20 records
5. 2025-2028emerging

   Language models enter the workflow

   The first widely deployed AI in cancer care is not a diagnosis but a time-saver: auto-contouring of organs and tumours for radiotherapy planning now runs in hundreds of centres. Language models are being tested to match patients to trials from the record at the moment a treatment is chosen, to draft tumour-board summaries and pathology reports, and to answer patient questions under supervision. Federated learning lets models train across hospitals without moving data. The evidence standard for each is still being written.

   Show era· 13 records
6. 2027-2032emerging

   From prediction to prospective proof

   The field has thousands of retrospective models and a handful of prospective trials. The infrastructure being proposed: a registry of external validation datasets with mandatory reporting, AI-first reading for high-volume common diagnoses with pathologists handling exceptions, every routine CT checked opportunistically for early cancer signs with a tracked pathway, AI central reads to cut trial endpoint cost, and digital twins as virtual control arms where a randomised control is unethical. Regulators are building predetermined change control plans so that models can update without re-clearance.

   Show era· 8 records
7. 2030+speculative

   Patient-level models and the virtual cell

   The two long-range bets are a multimodal model that reads slides, scans, genomics and the record to recommend and monitor treatment, and a virtual cell accurate enough to run a drug experiment in silico before it is run in a dish. Both depend on data at a scale no single institution holds, on validation standards that do not yet exist, and on liability and consent questions that are open today. The companion roadmaps on the AI clinic and the virtual cell follow each in detail.

   Show era· 8 records
8. What sets the pacecurrent

   Validation, data and compute

   The bottleneck is not model quality but the path from a published model to a deployed one: prospective evidence, external validation, regulatory status for updating models, payment, and data that can be shared. Records, scans and genomes sit in silos; real-world outcomes are weakly recorded, so there is little to learn from; and the workforce that would supervise AI is already short. Compute and model platforms are the one input that is not scarce.

   Hide era· 8 records

   AI that is built but not validated or deployedData silosWeak real-world evidence and registriesNot enough oncologists, nurses, pathologists, physicistsPreclinical results do not reproduceAI compute and model platforms for oncologyNVIDIANVIDIA BioNeMo

Every era's records, trial outcomes and papers, and every watch item, as JSON.