Objective:

To evaluate the impact of artificial intelligence (AI) assistance on lesion measurement time and classification agreement in follow-up CT examinations of cancer patients.

Approach:

• Study Design: A retrospective study involving follow-up chest, abdomen, and pelvis CT examinations from 212 patients with 539 target lesions, evaluated by 15 radiologists and 8 radiology residents under different conditions.
• Measurement Conditions: Readers assessed lesions under unassisted, AI-assisted, and expert-assisted conditions, remeasuring predefined target lesions according to RECIST 1.1.
• Outcomes Measured: Primary outcomes included reading time and interobserver measurement variability; secondary outcomes included proposal acceptance, patient-level change in sum of longest diameters (SLD), and RECIST response classification agreement.

Key Findings:

• AI assistance reduced mean reading time by 34% (from 105 seconds to 71 seconds per patient).
• Expert-assisted reading was faster at 56 seconds but required an initial unassisted measurement, resulting in a cumulative reading time of 151 seconds.
• AI assistance increased interreader agreement on RECIST response classification by approximately 8 percentage points compared to unassisted assessment.
• AI-assisted measurements showed greater deviation from expert-derived reference measurements (mean absolute error of 4.35 mm with AI vs. 3.08 mm without assistance vs. 1.51 mm with expert assistance).
• Radiologists accepted 61% of AI proposals compared to 77% of expert proposals, with substantial modifications occurring in 23% of AI-assisted measurements.

Interpretation:

AI assistance can accelerate RECIST assessment and improve consistency, but expert assistance yields greater agreement and efficiency.

Limitations:

• Baseline target lesions were predefined, limiting assessment of interobserver differences in baseline lesion selection.
• The study did not include new or nontarget lesion assessments, which could affect RECIST outcomes.
• Measurement variability analysis was biased in favor of expert assistance due to the reference standard used.
• Different readers evaluated measurements for a given patient across the three conditions, introducing potential confounding.

Conclusion:

AI-assisted RECIST assessment may enhance workflow and consistency in response classification while providing a benchmark for future AI development.

Sources:

• Radiology Advances