Enabling hybrid AI–radiologist endpoints in clinical trials imaging workflows.

Abstract

2019 Background: Imaging endpoints in oncology clinical trials commonly rely on a 2 + 1 radiologist reading paradigm, in which two independent readers perform primary assessments and a third adjudicates disagreements. This model underpins response evaluation frameworks such as RECIST and RANO, but is resource-intensive and contributes to trial cost, operational complexity, and inter-reader variability. Advances in transformer-based deep learning for 3D medical imaging enable hybrid AI–radiologist workflows that may preserve adjudicated oversight while reducing primary human reads. We evaluated the feasibility of replacing one primary radiologist reader with AI in a hybrid 2+1 imaging workflow, using glioma segmentation as a representative use case and with consideration for near-term integration into clinical trial imaging platforms such as Voiant Hub. Methods: A 3D SwinUNETR open-source deep learning model implemented using the MONAI framework pre-trained to segment glioma subregions from multimodal MRI, including enhancing tumor and non-enhancing/edema components. Model performance was assessed using Dice similarity coefficients across a multi-case dataset, complemented by qualitative review of representative cases spanning high-agreement and high-discordance scenarios. Two reading paradigms were evaluated conceptually: (1) traditional radiologist–radiologist + adjudicator, and (2) AI–radiologist + adjudicator. AI–radiologist segmentation divergence was used as a proxy for adjudication likelihood and workflow impact. Results: AI segmentation demonstrated substantial concordance with reference contours in well-defined enhancing tumor regions, with performance variability reflecting known biological heterogeneity. Dice scores for enhancing tumor reached values as high as 0.80, with a broad distribution across cases. Qualitative review showed that AI–radiologist disagreements were most commonly associated with small-volume lesions, infiltrative margins, or low-contrast regions: patterns similar to known human inter-reader variability. The presence of an adjudicator preserved clinical oversight for discrepant cases while eliminating the need for a second full human primary read. Conclusions: A hybrid AI–radiologist 2+1 imaging workflow is feasible for glioma volumetric assessment and has broader implications for RECIST- and RANO-based endpoints in oncology clinical trials. Substituting AI for one primary radiologist reader preserves adjudicated decision-making while reducing manual read burden. The availability of configurable, regulator-aligned imaging platforms such as Voiant Hub positions this hybrid approach as a near-term, easily adoptable evolution of current reading paradigms. Prospective evaluation is warranted to quantify operational impact, adjudication rates, and endpoint consistency across solid tumor and neuro-oncology trials.