The “AI can count the strawberries in this photo” meme has apparently grown up, gone to medical school, and started measuring tumors wrapped around lungs like extremely unwelcome cling film.

The paper in question is Kevin G. Blyth’s Lancet Oncology comment, “Artificial intelligence volumetry in mesothelioma: ready for deployment?” PMID: 42309109, DOI: 10.1016/S1470-2045(26)00169-5. It comments on a bigger new study introducing ARTIMES, an AI-assisted way to measure pleural mesothelioma on CT scans by tumor volume rather than the usual ritual of measuring selected tumor thicknesses with the confidence of a human trying to fold a fitted sheet.

The Humans Measure the Lung Wrapper

Pleural mesothelioma is a cancer of the lining around the lung, often linked to asbestos exposure. Unlike many tumors, it does not politely form a neat ball. It spreads along the pleura, hugging the chest wall in sheets, crescents, rind-like layers, and other shapes that seem designed to annoy radiologists.

The standard approach, modified RECIST, asks clinicians to measure tumor thickness at selected points and track whether those measurements shrink, grow, or sit there brooding. RECIST rules exist so humans can agree on whether tumors are responding to treatment, stable, or progressing. This is a sensible species-wide attempt at order. Unfortunately, mesothelioma did not attend the meeting.

A tumor can change volume in ways that a few line measurements miss. Imagine judging whether a spilled pancake batter blob has expanded by measuring three edges with a ruler. The humans call this “response assessment.” The alien observer calls it “optimistic geometry.”

Enter ARTIMES, the Very Serious Pixel Accountant

The related Lancet Oncology study by Groot Lipman and colleagues developed ARTIMES, short for AI-assisted response evaluation to treatment in mesothelioma DOI: 10.1016/S1470-2045(26)00084-7. Instead of asking a radiologist to choose a few slices and measurements, the system segments tumor tissue across CT scans and estimates total tumor volume.

This matters because deep learning segmentation can inspect the whole scan, voxel by voxel. In plainer terms: the AI colors in the tumor on the 3D medical image, then totals the colored-in space. It is less “measure this one suspicious bit” and more “inventory the entire chest-wall mess.”

The ARTIMES study reportedly used 10,926 CT scans from 2,080 patients across 14 cohorts, with training annotations from radiologists and clinical validation across trial datasets. Among 943 trial participants, ARTIMES detected progression earlier than modified RECIST, with reports noting a median 38-day earlier detection among patients found progressive by both methods. It also showed stronger association with overall survival than mRECIST in the published analyses.

The humans appear excited because 38 days is not a rounding error when the subject is cancer treatment. It may mean switching away from ineffective therapy sooner, designing trials with cleaner endpoints, and sparing patients the grand medical tradition of “let us continue this treatment and hope the ruler catches up.”

But Is the Machine Ready for the Clinic?

Blyth’s title asks the sensible question: ready for deployment?

The answer seems to be: promising, yes; hand it the keys unsupervised, no.

Medical AI has a deployment problem that resembles adopting a brilliant but socially awkward intern. It can do impressive work, but you still need supervision, documentation, quality checks, and someone legally responsible when it confidently segments the wrong thing. CT scanners differ. Contrast protocols differ. Patients differ. Pleural effusion, atelectasis, surgical changes, and odd anatomy can make tumor boundaries look like a foggy coastline.

Earlier work showed why automation is attractive. Kidd and colleagues built a fully automated CNN for mesothelioma volumetry and compared it with modified RECIST, reporting that automated volume measurement could better capture this disease’s awkward shape DOI: 10.1136/thoraxjnl-2021-217808. More recent CNN threshold studies examined how probability cutoffs affect segmentation quality, because even machines must decide where “tumor” stops and “not tumor” begins, which is philosophically deep and radiologically irritating PMCID: PMC11950581.

A 2025 scoping review also found growing AI work across mesothelioma prediction, diagnosis, staging, and prognosis, but the field remains small and uneven DOI: 10.1177/15330338251341053. Rare cancers do not provide the internet-scale data buffet that AI systems prefer. The GPUs may hunger, but the dataset is a tasting menu.

Why This Is More Than Fancy Measuring

If validated prospectively, volumetric AI could make mesothelioma trials sharper. Better response criteria can help researchers see whether a drug is actually helping, instead of waiting for blunt endpoints to lurch into view months later. It could also reduce disagreement between readers, support centralized trial review, and maybe make future staging more quantitative.

But the best version of this future is not “AI replaces radiologists.” It is “AI does the exhausting volumetric bookkeeping, and clinicians interpret the result in context.” The machine counts the pixels. The humans decide what the counting means for a person sitting in a clinic, trying to make the next hard choice.

That division of labor seems wise. Also very human: building a machine to measure the thing we always suspected our rulers were bad at measuring.

References

1. Blyth KG. Artificial intelligence volumetry in mesothelioma: ready for deployment? The Lancet Oncology. 2026. PMID: 42309109. DOI: 10.1016/S1470-2045(26)00169-5.

2. Groot Lipman KB, Wittenberg R, de Oliveira Taveira M, et al. Development and validation of artificial intelligence-assisted volumetric response criteria in pleural mesothelioma (ARTIMES): a retrospective, multicohort, multicentre study. The Lancet Oncology. 2026. PMID: 42309108. DOI: 10.1016/S1470-2045(26)00084-7.

3. Kidd AC, Anderson O, Cowell GW, et al. Fully automated volumetric measurement of malignant pleural mesothelioma by deep learning AI: validation and comparison with modified RECIST response criteria. Thorax. 2022;77(12):1251-1259. DOI: 10.1136/thoraxjnl-2021-217808.

4. Shenouda M, Gudmundsson E, Li F, et al. Convolutional Neural Networks for Segmentation of Malignant Pleural Mesothelioma: Analysis of Probability Map Thresholds. 2023. PMCID: PMC11950581. arXiv: 2312.00223.

5. Ram M, Afrash MR, Moulaei K, et al. Predicting Mesothelioma Using Artificial Intelligence: A Scoping Review of Common Models and Applications. Technology in Cancer Research & Treatment. 2025;24. DOI: 10.1177/15330338251341053.

Disclaimer: This blog post is a simplified summary of published research for educational purposes. The accompanying illustration is artistic and does not depict actual model architectures, data, or experimental results. Always refer to the original paper for technical details.