Artificial Intelligence for RECIST-Based Radiologic Treatment Response Assessment in Solid Tumors: A Systematic
Agnieszka Leszczyńska1,2, Michał Seweryn1,2, Rafał Obuchowicz3,4
1EconMed Europe, Młyńska 9/4, 31-469 Krakow, Poland.
Background/Objectives:
To systematically review and critically appraise AI methods for RECIST-based radiologic treatment response assessment in solid tumors, comparing image-derived and report-derived approaches and summarizing their performance, agreement with reference standards, and validation quality.
Methods:
This systematic review followed PRISMA guidelines. We searched Embase, MEDLINE, Web of Science, Scopus, and the Cochrane Library on 6 December 2025. We included English-language original studies (2015-2025) in solid tumors where AI directly assigned RECIST response categories and was validated against a reference standard; studies without RECIST-based response endpoints or non-solid tumor populations were excluded. We distinguished image-based techniques that assign RECIST categories from direct analysis of imaging data from report-based techniques that infer RECIST categories from radiology reports using natural language processing.
Results:
Evidence remains sparse; we identified four eligible studies (two image-based and two report-based). DeepSeek-V3-0324 and GatorTron, both report-based approaches, achieved high accuracy (96.5% and 89%, respectively) in treatment response evaluation, with DeepSeek demonstrating higher expert agreement (κ 0.85-0.90). The nnU-Net and 3D U-Net pipelines, both image-based, showed high segmentation performance (DSC 0.85, VS 0.89) and treatment response classification accuracy of 0.77 for R1, with moderate agreement with the manual reference (κ = 0.60); nnU-Net also achieved moderate to almost perfect agreement (Cohen's κ 0.67-0.81) in RECIST 1.1 measurements.
Conclusions:
AI-based RECIST-oriented response assessment is feasible and potentially beneficial for standardization, efficiency, and scalability, but current evidence is limited and heterogeneous, requiring larger multi-center studies with rigorous external validation before clinical adoption. Key limitations include data source variability, reference standard inconsistencies, and lack of robust external validation.
