World Health AIRegister

All insights

Healthcare

CT-based AI flags gastric cancer patients likely to relapse or respond to immunotherapy

In two Chinese immunotherapy cohorts, 54·5% and 53·1% of high-scoring patients responded, against 5·6% and 9·7% of low scorers.

By World Health AI Editorial2 min read1 source

Chinese researchers have built a deep learning model that reads routine CT imaging to predict whether gastric cancer patients will see the disease return in the lining of the abdomen, and which of them are likely to respond to chemotherapy or immunotherapy. The study, a multicentre development, validation and testing exercise, was published in The Lancet Digital Health on 19 November 20261.

The model, called CTransNet, combines a convolutional neural network (ResNet50) with a Transformer and is trained on two tasks at once1. The first is to estimate the presence and maturity of tertiary lymphoid structures, clusters of immune cells in and around a tumour whose maturation is linked to better prognosis. The second is to predict peritoneal recurrence-free survival after curative surgery1. Today those immune structures are assessed from tissue, which the authors describe as invasive and prone to bias1.

What the study found

The team used retrospective data from eight Chinese cohorts treated at Nanfang Hospital, Sun Yat-sen University Cancer Center, Yunnan Cancer Hospital and Guangdong Provincial Hospital of Chinese Medicine, recruited between 1 January 2005 and 28 February 20221. These comprised one training cohort, two internal and three external validation cohorts, and two separate cohorts of patients who had received at least two cycles of anti-PD-1 or anti-PD-L1 drugs1.

  • Classifying immune-structure status in the training cohort, the area under the curve was 0·910 for absent structures, 0·925 for immature ones and 0·953 for mature ones1.
  • For peritoneal recurrence-free survival the concordance index was 0·654 in the training cohort1.
  • Patients with low model scores had a 5-year peritoneal recurrence rate of 66·69% in the training cohort against 26·28% for those with high scores (p<0·0001)1.

In multivariate analysis the model's risk score was an independent predictor of peritoneal recurrence-free, disease-free and overall survival1. Among stage II and III patients, adjuvant chemotherapy was associated with better disease-free and overall survival in the low-score group but not in the high-score group1. In the immunotherapy cohorts, objective response rates were 54·5% versus 5·6% for high against low scores in the first cohort and 53·1% versus 9·7% in the second1.

What it means for the market

The commercial interest lies in getting an immune biomarker out of imaging that hospitals already acquire, rather than from extra tissue work. If it held up prospectively, a tool of this kind could help oncology teams decide who receives adjuvant chemotherapy and who is a better candidate for checkpoint inhibitors, two expensive decisions with very different response rates in this study1. For imaging AI vendors and drug developers running immunotherapy trials, the response-rate gap points to a possible use in patient selection, although the paper reports no regulatory, procurement or commercial plans1. The study received no funding1.

Caveats

All the data are retrospective and every cohort comes from hospitals in China, so performance in other populations and health systems is untested1. Recruitment runs from 2005 to 2022, a long span over which imaging and treatment practice changed1. The recurrence model's concordance index of 0·654 is modest even though its risk groups separate sharply1. The authors describe the tool as showing potential for personalised care, and a prospective trial would be the next test before any clinical use1.

Sources

  1. Multimodal multitask deep learning for predicting tertiary lymphoid structures and peritoneal recurrence in gastric cancer: a multicentre development, validation, and testing study. (opens in a new tab)The Lancet. Digital health · doi.org

This briefing summarises publicly available research and reporting for information only. It is not medical, investment or legal advice. Follow the sources listed for the primary research and reporting.