AI-Assisted Volumetric Response Criteria Outperform mRECIST in Pleural Mesothelioma Assessment
核心洞察
The ARTIMES (搜索) AI framework uses automated CT-based 3D tumor volumetry to assess treatment response in pleural mesothelioma (搜索), addressing limitations of diameter-based mRECIST criteria.
In a retrospective study of 943 trial patients, ARTIMES (搜索) detected disease progression a median of 38 days earlier than mRECIST and showed superior prognostic performance (concordance index 0.83 vs 0.73).
ARTIMES (搜索)-based progression-free survival demonstrated a strong trial-level association with overall survival (R²=88%), compared with only 6% for mRECIST-based progression-free survival.
A new artificial intelligence-assisted volumetric response framework called ARTIMES (搜索) has demonstrated superior performance over conventional modified RECIST (mRECIST) criteria for assessing treatment response in pleural mesothelioma (搜索), according to a landmark multicenter study published in The Lancet Oncology.
The framework, developed by researchers at the Netherlands Cancer Institute (搜索), combines automated deep-learning segmentation of pleural tumors on CT imaging with biologically informed volumetric response criteria. In a retrospective analysis spanning 10,926 CT scans from 2,080 patients across 14 cohorts—including ten phase II and phase III clinical trials—ARTIMES (搜索) detected disease progression earlier, showed stronger associations with overall survival, and outperformed both physician assessments and standard RECIST-based measurements.
The Challenge of Mesothelioma Response Assessment
Pleural mesothelioma (搜索) presents unique challenges for response evaluation. Unlike most solid tumors, it does not grow as a rounded mass but instead spreads along the pleura in a crescent-like pattern around the lung. This morphology makes traditional diameter-based measurements—the foundation of RECIST and mRECIST—particularly problematic.
mRECIST estimates tumor burden by measuring up to six representative tumor thicknesses perpendicular to the chest wall or mediastinum, with progressive disease defined as at least a 20% increase and partial response as at least a 30% decrease. However, different radiologists may select different measurement sites, tumor thickness may not fully reflect changes across the entire pleural tumor burden, and small measurement variations can influence response classification.
How ARTIMES (搜索) Works
ARTIMES (搜索)—Artificial Intelligence-Assisted Response Evaluation to Treatment in Mesothelioma—addresses these limitations through two core principles. First, changes within expected inter-reader variability should not be interpreted as true biological progression or response. Second, tumor growth similar to untreated mesothelioma growth may indicate treatment failure.
The AI model identifies pleural and adjacent tumor tissue on CT scans, calculates total three-dimensional tumor volume, and tracks volumetric changes across follow-up scans. A medical professional must still review and approve the AI-generated segmentation before ARTIMES (搜索) response criteria are applied.
Partial response under ARTIMES (搜索) is defined as either a reduction of more than 15% and more than 35 mL compared with the maximum tumor volume during treatment, or a reduction of more than 75% compared with that maximum. Progressive disease is defined as a new lesion outside the pleura, an increase of more than 40% and more than 35 mL compared with the lowest tumor volume since treatment began, or an increase of more than 70 mL compared with that lowest volume.
Strong Segmentation Performance
The deep-learning model was trained on 1,176 CT scans annotated by 12 radiologists and one pulmonologist, supplemented with 100 CT scans from tumor-free controls. In internal testing, the model achieved a median Dice similarity coefficient of 94% and a normalized surface distance of 98%. AI-derived tumor volumes strongly correlated with radiologist-derived volumes, with an R² of 99% on individual scans. The model maintained performance even in challenging imaging settings, including scans with pleural effusion, atelectasis, and thoracic wall invasion.
Earlier Detection of Progression
Among 943 trial participants with 4,674 CT scans and available survival data, mRECIST identified progressive disease in 629 patients, while ARTIMES (搜索) identified progressive disease in 635 patients. Critically, ARTIMES identified progression earlier: median time to progression was 126 days with ARTIMES compared with 161 days with mRECIST. Among patients classified as progressing by both methods, ARTIMES detected progression a median of 38 days earlier.
"ARTIMES (搜索) demonstrated superior prognostic performance (concordance index = 0.83; 95% CI = 0.79–0.87) compared with modified RECIST criteria (concordance index = 0.73; 95% CI = 0.66–0.80; P = .023)," the study reported. ARTIMES achieved a higher concordance index than mRECIST in seven of eight evaluated trials.
Trial-Level Surrogate Endpoint Performance
One of the most striking findings involved surrogate endpoint performance at the trial level. ARTIMES (搜索)-based progression-free survival showed a strong association with overall survival, with an R² of 88% (95% CI = 42%–100%). By comparison, mRECIST-based progression-free survival had an R² of only 6% (95% CI = 0%–97%).
ARTIMES (搜索) also demonstrated a surrogate threshold effect: a progression-free survival hazard ratio below 0.82 predicted a statistically significant overall survival benefit at the trial level. No surrogate threshold effect was observed for mRECIST.
Prognostic Value of Baseline Tumor Volume
The study further evaluated the prognostic value of pretreatment AI-derived tumor volume. Baseline tumor volume independently predicted overall survival. When AI tumor volume was added to a model including age, sex, histology, treatment line, T stage, and WHO performance status, the prognostic model reached a concordance index of 0.65, compared with 0.58 without AI tumor volume. Notably, baseline AI tumor volume outperformed T stage and WHO performance status as an independent prognostic variable.
Implications for Clinical Trials and Practice
Mesothelioma trials have often faced challenges when early response signals fail to translate into survival benefits in later-phase studies. More reproducible and clinically meaningful response criteria may improve the selection of therapies for phase III testing. ARTIMES (搜索) could potentially support blinded independent central review in future trials, reduce measurement variability, and expand eligibility by removing the requirement for traditionally measurable disease.
"We obviously want patients worldwide to benefit from this," said lead study author Kevin B. W. Groot Lipman, PhD, a technical physician in the Department of Thoracic Oncology at the Netherlands Cancer Institute (搜索). "We are in the process of getting the model approved for use in other hospitals."
Dr. Lipman added, "I expect this model to come as a shock to physicians and researchers outside the mesothelioma field. This is going to open up a whole new field of research. We expect that AI will also be able to help with many other types of tumors."
Limitations and Next Steps
The study was retrospective, and treatment discontinuation in the historical datasets was often guided by mRECIST, which may have introduced bias when evaluating progression-free survival. Variability in non-contrast CT scans can also make distinction between tumor, pleural effusion, and atelectasis more difficult.
Prospective validation is essential before ARTIMES (搜索) can be recommended for broad clinical use. The COMET trial, already underway, will assess both ARTIMES and mRECIST criteria in a setting where neither classification alone determines treatment cessation. The study authors have also shared their code so that researchers globally can begin using the new model, with the Netherlands Cancer Institute (搜索) already exploring AI applications for lung cancer and brain metastases.