AI-Driven Tumor Response Evaluation for Solid Tumors

Background: Current tumor response evaluation relies primarily on RECIST 1.1 and its variants, which measure changes in the longest diameter of a limited number of target lesions. While standardized and widely used, these criteria have limitations: they may not fully reflect total tumor burden changes, fail to capture spatial and temporal heterogeneity across lesions, and are subject to inter-observer variability. Advances in artificial intelligence, particularly in medical image analysis, now enable automated tumor segmentation and volumetric quantification, offering a more comprehensive assessment of tumor burden dynamics. However, systematic validation of AI-driven volume-based response criteria against traditional methods remains limited.

Study Design: This is a multi-center, retrospective-prospective cohort study designed to develop and validate an AI-driven prognostic model based on total tumor volume changes and multi-dimensional features. The study is being conducted across approximately 40 participating sites in China.

Data Sources:

Training Set (Retrospective): Approximately 6,000 patients with solid tumors who received non-surgical treatment between January 2015 and December 2025, including the Hepatorch cohort (~300 patients). Data include imaging (CT/MRI), clinical characteristics, laboratory tests, and molecular markers.

Validation Set: Approximately 20% of the training set data randomly extracted for hyperparameter tuning and internal validation.

External Test Set (Prospective): Approximately 120 patients consecutively enrolled from mid-2026 through 2027, independent of the training/validation sets, to assess model generalizability in real-world clinical settings.

AI Model Development: The prediction model integrates imaging biomarkers (total tumor volume, single-lesion volume, enhanced volume, lesion count), clinical variables (age, sex, tumor type/stage, liver function, treatment modality), and laboratory/molecular markers (AFP, CA19-9, immune markers, genetic sequencing). Modeling approaches include joint models for longitudinal volume changes and survival outcomes, random survival forests, gradient boosting survival models, and deep learning survival models. Model performance is evaluated using C-index, time-dependent ROC curves, calibration curves, and decision curve analysis.

Prospective Sub-studies:

Lesion Tracking and Biological Characterization (~20 patients): Serial tracking of individual lesions with volumetric measurement, plus collection of leftover tumor tissue from clinically indicated biopsies for molecular and immune microenvironment analysis.

Patient Experience and Communication Value Assessment (~100 patients): Participants are randomized 1:1 to receive either a standard RECIST report or an AI-assisted dynamic tumor response report. Standardized questionnaires assess report comprehension, cognitive burden, anxiety, trust, and treatment decision confidence.

Human-Machine Collaboration: In both retrospective and prospective components, the study evaluates AI-assisted tumor response assessment by comparing independent clinician reading, AI-assisted reading, and expert adjudication. Consistency metrics include Kappa statistics, intraclass correlation coefficients, and measurement error. Efficiency metrics include reading time and report generation time.

Follow-up Schedule: For the prospective cohort, follow-up visits occur every 3 months (up to 6 months) aligned with routine clinical care, collecting imaging, laboratory results, treatment changes, disease progression, survival status, and subsequent therapy information.

Statistical Considerations: Model development employs LASSO regularization, stepwise regression, and machine learning-based feature selection. Model validation uses K-fold cross-validation and independent external validation. Predictive performance is assessed using time-dependent ROC, C-index, calibration plots, and decision curve analysis. Missing data are handled using multiple imputation.

Ethical Considerations: The retrospective component uses de-identified data from patients who previously consented to biobank participation, with a waiver of informed consent requested. The prospective component requires written informed consent from all participants. The study protocol has been approved by the Institutional Review Board of Zhongshan Hospital, Fudan University.