Abstract B010: From RECIST to reality: A foundation model pipeline for scalable therapy response evaluation
Bibliographic record
Abstract
Abstract Motivation: Quantifying tumor burden is central to response assessment in oncology clinical trials and treatment monitoring. Despite widespread use, the clinical standard Response Evaluation Criteria in Solid Tumors (RECIST) has several limitations: it relies on unidimensional measurements of a limited set of lesions, ignores volumetric change, and exhibits high interobserver variability. These constraints limit its sensitivity in complex disease presentations, ultimately affecting treatment decisions and trial outcomes. Volumetric tumor assessment is more robust and sensitive, but the process of manual segmentation is impractical at scale. Advances in deep learning and foundation models provide an opportunity to transform response evaluation by automating volume estimation with minimal human input. Methods: We present AI-Augmented Response Assessment (AAuRA), a novel computational workflow that leverages existing RECIST-like measurements to seed MedSAM—a foundation model for semi-automated medical image segmentation. As a proof-of-concept, we reverse-engineer RECIST annotations by decomposing 3D tumor segmentations to extract the longest axial diameter and its perpendicular, to generate bounding boxes that simulate RECIST-based input. We evaluate AAuRA on two public computed tomography imaging datasets with tumor segmentations: LIDC-IDRI and NSCLC-Radiogenomics. Performance is assessed using Volumetric Dice Similarity Coefficient (DSC), 95th percentile Hausdorff Distance (95HD), Added Path Length (APL), and Pearson correlation between predicted and ground truth volume. To account for size-related bias in spatial metrics, APL and 95HD are normalized by tumor volume. Results: In the NSCLC-Radiogenomics dataset, AAuRA achieved a mean DSC of 0.72, APL of 0.20, and 95HD of 0.01. In the LIDC-IDRI dataset, the mean DSC was 0.50, with an APL of 0.42 and 95HD of 0.11. Volume predictions from AAuRA demonstrated strong correlation with ground truth segmentations: the Pearson correlation coefficient was 0.86 (p < 0.001) and 0.90 (p < 0.001) for the LIDC-IDRI and NSCLC-Radiogenomics datasets, respectively. Segmentation performance was lower for smaller, ill-defined lesions—highlighting known limitations of current foundation models and motivating the need for domain-specific adaptation and fine-tuning. Outlook: AAuRA serves as a clinically viable bridge between current standards and AI-augmented tools for therapy response evaluation. Using familiar annotations as input, AAuRA can avoid disruptive changes to clinical workflows while enhancing accuracy and scalability of tumor burden assessment. This work highlights how clinically meaningful priors—like RECIST—can be repurposed to guide foundation models, offering a blueprint for real-world deployment of AI in medicine. Future work will extend the pipeline to real RECIST annotations across diverse disease sites and explore integration into prospective imaging workflows. AAuRA’s modular design also enables adaptability to new foundation models and evolving segmentation standards. Citation Format: Katarina Vucic, Caryn Geady, Katy L. Scott, Joshua Siraj, Andrew J. Hope, Benjamin Haibe-Kains. From RECIST to reality: A foundation model pipeline for scalable therapy response evaluation [abstract]. In: Proceedings of the AACR Special Conference in Cancer Research: Artificial Intelligence and Machine Learning; 2025 Jul 10-12; Montreal, QC, Canada. Philadelphia (PA): AACR; Clin Cancer Res 2025;31(13_Suppl):Abstract nr B010.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.002 | 0.006 |
| Meta-epidemiology (narrow) | 0.001 | 0.001 |
| Meta-epidemiology (broad) | 0.001 | 0.002 |
| Bibliometrics | 0.001 | 0.000 |
| Science and technology studies | 0.001 | 0.001 |
| Scholarly communication | 0.002 | 0.001 |
| Open science | 0.002 | 0.002 |
| Research integrity | 0.001 | 0.002 |
| Insufficient payload (model declined to judge) | 0.013 | 0.005 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".