Organ-agnostic automated RECIST measurements across time points with foundation models.

Abstract

e13669 Background: RECIST 1.1 remains the gold standard for evaluating tumor response in clinical trials and patient care. However, its reliance on manual measurements introduces subjectivity and inter-observer variability, which can affect the accuracy of therapeutic response assessments. While lesion segmentation methods have improved, automating longitudinal measurements remains a critical challenge. This study focuses on advancing the automation of follow-up tumor measurements. By building on the baseline lesions’ segmentations, our approach aims to streamline the longitudinal evaluation process, paving the way for automated and reproducible tumor tracking. Methods: A multi-center cohort of 36 patients with two time points (22 with stable disease, 7 with progressive disease, and 7 with partial response) and a total of 90 lesions from various organs, including the lung, pancreas, liver, bone, and breast, was used to evaluate the automatic assessment of lesions across time points. Two radiologists performed manual RECIST 1.1 measurements and segmentations of the lesions. Ground-truth diameters were extracted from the lesions’ segmentations. The automatic pipeline relies on the following steps: Registration : The baseline scan was registered to the follow-up scan using affine and elastic registration, leveraging features from a foundation model trained on 281,000 MR/CT scans. [1] Segmentation Propagation : Using Oncopilot, a promptable foundation model for segmentation [2,4], the registered baseline mask was used as a visual prompt to propagate the segmentation on the follow-up scan, yielding a 3D follow-up lesion mask. Measurement Extraction : RECIST measures (small or large axis respectively for lymph nodes or solid tumors) were extracted from the propagated masks. The proportion of lesions accurately identified by the automatic pipeline (Dice > 0.1 [3]) is reported. For these identified lesions, absolute and relative errors in RECIST diameters, relative errors for volumes, and Dice scores are evaluated. Results: 80% of the lesions were properly located in the follow-up exam. These lesions resulted in a median absolute error in RECIST measure of 1.9 mm (relative error 8%), comparable to the 1.9 mm error (relative error 10%) observed for manual radiologist measurements. The median relative error in volume assessment was 20%, with a median Dice of 0.71. Conclusions: Our approach leverages foundation models to automate RECIST 1.1 measurements from baseline to follow-up, enabling accurate tumor assessments over time. Beyond RECIST, it introduces automatic longitudinal tracking of lesion volumes, offering a more comprehensive evaluation method. Defining clear standards for volume-based lesions monitoring is a crucial next step to establishing guidelines and performance benchmarks for automated methods. [1] Song et al. 2024 [2] Machado et al., 2024 [3] McKinney et al., 2020 [4] Hérent et al., 2024.