Combining blood tumour markers with chest CT — read either by a radiologist or by a deep-learning model — identified a larger share of lung-cancer cases than blood tests alone in a 346-patient study from Catharina Hospital and Eindhoven University of Technology. Catharina Ziekenhuis described the work on 4 November 2026; TU/e followed on 5 November. The results were shown as a late-October poster at ESMO in Madrid, where researcher Hao Cao received an ESMO Merit Award. The team says it is preparing a journal manuscript; there is not yet a peer-reviewed journal article for these figures.
Blood markers alone identified about 62% of lung-cancer cases, the institutions say. Adding radiologist CT assessment raised identification to about 78% in the first study group. Blood markers plus deep-learning CT analysis reached about 76% in that first group and about 73% in an independent second group — in the same range as the radiologist path. The AI route did not outperform radiologists.
What the percentages are — and are not
Those shares are the study’s reported identification metrics as Catharina and TU/e state them. Neither release TSN used for this draft labels the headline percentages as “accuracy,” “sensitivity” or “specificity” in the sense of a fully specified diagnostic trial readout, so this post keeps their wording: how many cases the combined pathways identified in the cohorts described. Over-reading them as an approved diagnostic performance claim would go beyond the sources.
Cao’s quoted line on the blood-only arm is plain: “Using blood tests alone, we were able to identify around 62 per cent of patients with lung cancer.” On imaging expertise, the TU/e note has Cao saying radiologists already extract “a tremendous amount of valuable information from a CT scan,” and that “AI can help, but our research also demonstrates just how valuable a radiologist’s expertise is.”
The next step the hospitals name is further validation in practice without using the outputs for clinical decisions yet. Separate Catharina Health Foundation funding of €45,000 for a DNA-blood project is a different line of work and is unpublished in these releases; it is not folded into the 346-patient metrics above.
Limits the teams themselves put on the tool
Several hard boundaries sit in the institutional copy. A negative combined result does not rule out lung cancer. The approach is not a substitute for biopsy. It cannot yet determine cancer type. It is research-stage work presented on a conference poster, not a CE-marked or FDA-cleared diagnostic, and the partners say clinical decision-making should not yet rest on these outputs.
That framing matters for any reader who meets the story through a social headline about “AI detecting lung cancer.” The useful comparison inside the study is multimodal enrichment of blood markers — radiologist CT versus deep-learning CT — with AI landing close to, not above, the human radiology arm in the reported identification shares.
Where this sits next to other TSN health-AI posts
TSN’s recent medicine file has several research-stage screening and imaging tools that stop short of bedside use. RPI’s Quest2Dx Parkinson’s questionnaire model is a screening aid on research data, not a diagnosis. Google’s smartwatch hypertension signal in Nature remains a research result, not a diagnosis product. UKRI’s five healthcare AI hubs are funding infrastructure for later evidence. On the vaccine side, Cambridge and DIOSynVax’s AI-designed bird-flu candidate still awaits human trials after animal challenge data.
Catharina/TU/e belongs with that research cluster: multimodal signals that improve case identification in a defined cohort, with an explicit bar on clinical use until further validation. It is Dutch hospital–university work with an ESMO poster pedigree, not a UK NHS procurement story and not a replacement for the biopsy pathway Dutch or UK guidelines already require when imaging and markers raise concern.
Why “comparable to radiologists” is the careful headline
The temptation in multimodal AI stories is to crown the model. The numbers here cut the other way. In the first group, blood plus radiologist CT sat at about 78% identification; blood plus AI CT at about 76%. The second-group AI path landed at about 73%. The institutions’ own emphasis is that AI helped relative to blood alone and stayed in the radiologist’s band — evidence for augmentation of an existing CT read, not for sidelining it.
For compliance and clinical-governance readers, the poster-versus-journal distinction is as important as the percentages. Conference abstracts and posters are how early oncology methods travel; they are not the same evidence grade as a completed peer-reviewed methods paper with full inclusion criteria, preprocessing, model cards and locked analysis plans. Until that paper appears, outside teams cannot audit the deep-learning CT pipeline from these newsroom pages alone.
UK readers should treat the piece as a Netherlands research update. Neither Catharina nor TU/e announces NICE evaluation, MHRA device registration or an NHS pilot tied to this poster. Any future UK relevance would need a separate validation cohort and a regulatory path these releases never announce.
What to watch next
Three public breadcrumbs would move the file: a peer-reviewed journal article that locks the 346-patient methods and metrics; a prospective validation protocol that still withholds outputs from clinical decisions, as the teams currently require; or a clear statement that a named metric (sensitivity, specificity, AUC) has been pre-specified and met. Until then, the dual institutional record is blood markers plus CT — human or AI — lifting identification above blood alone in two study groups, with AI comparable rather than superior, and with clinical use still off the table.
Sources
Primary:
- Catharina Ziekenhuis: Bloedonderzoek en CT-scan samen maken longkanker beter herkenbaar (4 Nov 2026)
- TU/e: Combining blood tests and CT scans improves lung cancer detection; research wins international award (5 Nov 2026)
Related TSN:
- RPI’s Quest2Dx AI Screens for Parkinson’s Using Health Questionnaires
- UKRI Funds Five Healthcare AI Research Hubs Over £52m
- Google Smartwatch AI Flags Hypertension in Nature Study
- Cambridge AI Bird Flu Vaccine Protects Animals in Trials
Related stories
- RPI’s Quest2Dx AI Screens for Parkinson’s Using Health Questionnaires
- UKRI Funds Five Healthcare AI Research Hubs Over £52m
- Google Smartwatch AI Flags Hypertension in Nature Study
- Cambridge AI Bird Flu Vaccine Protects Animals in Trials
Follow @tsnmediahub on X for breaking AI news as it happens.
Follow on X
