1421P - AI-powered intracranial tumor response predicts systemic progression with high concordance in endpoint evaluation in the phase III CROWN study

Background

Imaging reading for endpoint evaluation in clinical trials presents challenges in reproducibility and accuracy, limiting the use of novel endpoints beyond human capabilities. This analysis explored the feasibility of AI-powered intracranial response assessment and uncovered novel imaging prognosticators in the phase III CROWN study.

Methods

72 patients (40 patients in the crizotinib arm, 32 patients in the lorlatinib arm) with CNS metastases at baseline in the CROWN study (NCT03052608) underwent intracranial response assessment using MRI at screening and then every 8 weeks until disease progression. VBrain, an FDA-approved brain tumor AI, was used to generate response assessments that were compared with those performed by 2 neuro-radiologists from the blinded independent central review group for the study using modified RECIST (mRECIST). The sum of tumor diameter changes at first on-treatment tumor scan from baseline, denoted as early tumor shrinkage (ETS), was assessed with Cox's proportional hazard model to estimate drug efficacy.

Results

For target lesion measurements, AI showed a high correlation with the independent readers (Pearson correlation coefficient ρ=0.94). The intracranial objective response rate assessed by AI was 27.5% and 71.9% in the crizotinib and lorlatinib arms, respectively. Among 666 time-point assessments, discordance rates were 35.1% between AI and reader 1, 33.5% between AI and reader 2, and 35.7% between the 2 readers. The time to intracranial progression (iTTP) derived from AI and human readers showed no significant difference by log-rank test. AI tracked 394 tumors beyond mRECIST, with a median ETS of -41.8% for all these lesions vs -24.4% for target lesions by mRECIST. Brain-wide ETS better predicted systemic TTP (HR 0.40, [95% CI 0.21–0.76], p<0.01), while target-lesion ETS showed no significant difference (HR 0.59, [95% CI 0.32–1.09], p=0.09) by ETS categories > or <= to median ETS.

Conclusions

There was high agreement between AI and readers in lesion measurement, response assessment, and endpoint evaluations in the CROWN study. ETS leveraging all brain lesions may uncover a potential prognostic indicator of response in ALK-positive advanced NSCLC patients.

Clinical trial identification

NCT03052608. Last update posted: May 31, 2022.

Editorial acknowledgement

Legal entity responsible for the study

Vysioneer Inc.

Funding

Vysioneer Inc.

Disclosure

C. Duan, A. Schayowitz, P. Selaru, K.D. Wilner: Financial Interests, Personal and Institutional, Full or part-time Employment: Pfizer Inc. Y. Chang, C. Liang, P. Chiang, V. Lin, Y. Yang, J. Lu: Financial Interests, Personal, Full or part-time Employment: Vysioneer Inc. All other authors have declared no conflicts of interest.