Scientific publications

The article reviews the variability in radiologic oncology assessments, particularly focusing on RECIST 1.1 criteria, which standardize evaluations to improve consistency and accuracy. It discusses how variability arises from factors like radiologist expertise, image quality, and lesion selection, and emphasizes the importance of standardized protocols and training to mitigate these issues. By addressing the root causes of variability, the article aims to enhance the precision of response assessments, ultimately leading to better patient care and clinical outcomes.

Hepatocellular carcinoma (HCC) constitutes a prominent global health challenge. According to the American Association for the Study of Liver Diseases (AASLD) guideline HCC can be diagnosed by imaging examination. However, it shows that contrast-enhanced CT has limited accuracy in the diagnosis of HCC, particularly, small-size HCC lesions (≤ 20 mm) are the most difficult to identify with CT demonstrating sensitivity = 64% [1]. An Artificial Intelligence (AI) algorithm that can analyze liver CT images and localize HCC lesions would be valuable for patients at risk of HCC. Recent studies have presented promising outcomes regarding the application of AI algorithms in HCC diagnosis; However, the efficacy in localizing small-sized HCC lesions in CT images remains uncertain, and there is a need for improvement in reducing the false-positive rate [2-4].

The FDA recommends Blinded Independent Central Review (BICR) with double reads for imaging in clinical trials, but inter-reader variability raises concerns. Our study examined this variability in lung clinical trials using RECIST. We analyzed 5 phase III trials with 7 readers forming 11 teams, covering 1,017 patients. The study focused on Discrepancy Rate (DR), bias, endorsement rate, and the impact of adjudication on Progression-Free Survival (PFS). Results showed significant bias among readers, affecting double readings but no correlation between bias and DR. Additionally, adjudication significantly affects PFS. These outcomes highlight the need to improve monitoring in clinical trials.

The abstract presents the performances and subpopulation analysis of a computer-aided detection and characterization (CADe/CADx) AI model developed to aid lung cancer screening standard of care. AI 3D-CNN models were used on a test set containing 136 cancer and 2,027 benign patients. Performance at nodule level in terms of AUC-ROC is 0.987, significantly outperforming the NLST Brock model (AUC-ROC 0.971). This result is also consistent across nodules’ key characteristics (size, attenuation, margin) and shows it could aid clinicians in optimizing their clinical routine and the clinical management of patients.