Deep Learning Volumetric Assessment of Liver Metastases in CECT in Comparison to RECIST

Objectives: To compare disease status classification of abdominal CECT by radiologists using RECIST 1.1 to computer-aided assessment using a new deep learning method that automatically quantifies volumetric changes in liver metastasis between the prior and the current scans.

Methods: Computer-aided evaluation was performed by a novel method based on a 3D U-Net classifier that simultaneously analyzes the prior and current CECT scans. Forty-three pairs of studies of patients with metastatic liver disease were analyzed. Two radiologists performed 172 readings to assess disease status classification using RECIST 1.1 (43 pairs each) and two weeks later using computer-aided assessment (43 pairs each). The primary outcome measure was the disease status assessment with the reference standard defined by a third radiologist who manually delineated all lesions. The disease statuses are: Stable Disease (SD), Partial Response (PR), Progressive Disease (PD) and Cmplete Response (CR).

Results: For reader 1, (11 cases, 30%) and for reader 2, (13 cases, 35%) there were differences in disease status classification (p=0.01) between the conventional and the computer-aided reading that were in agreement with the reference standard [reader 1: SD changed to PR in 4 cases and PD in 4 cases, and PD changed to SD in 3 cases; reader 2 PR changed to SD in 1 case and PD in 1 case, and SD changed to PR in 7 cases and PD in 4 cases. The main reason for the difference between the two readings was lesion volume differences (p=0.01). These results indicate an improvement in the specificity of the assessment with the AI-based computer-aided method, as fewer cases are SD.

Conclusions: AI-based computer-aided analysis of changes in the volume of liver metastases may improve the estimation of changes in neoplastic involvement of the liver and the accuracy of the evaluation of disease status.