Automatic detection of lesions by “artificial intelligence” in [18F]FDG PET/CT in clinical practice: What is the most efficient solution?

Abstract

Introduction

Artificial intelligence (AI) algorithms have recently been commercialized to assist nuclear medicine physicians in lesion detection in clinical practice. The aim of this study was to evaluate the lesion detection performance on [18F]FDG PET/CT by two commercially available AI algorithms using a convolutional neural network (CNN), when used alone or in addition to a human reading.

Materials and methods

151 [18F]FDG PET/CT scans of patients managed for melanoma or lymphoma in 3 French centers were retrospectively analyzed. Lesion detection was performed according to four methods: manually (M1); by using a « PET Assisted Reporting System » (PARS) based on CNN trained on 629 patients with lymphoma and lung cancers, with a detection threshold at SUV 2.5 (M2) or PERCIST-like (M3); and by using a one-step U-net algorithm trained on 4906 patients with multiple neoplasias (M4). All volumes of interest (VOIs) identified by each of the 4 methods were reviewed by an expert consensus. VOIs judged to correspond to neoplastic lesions related to the pathology studied were labeled “true positives” (TP). The sensitivities of detection methods were compared in pairs using Student's paired-sample t-test, as well as sensitivities of automated methods combined with human reading.

Results

A total of 1544 lesions were considered as the reference standard. The respective sensitivities one-step method (M4), PARS with a SUV threshold of 2.5 (M2) and PERCIST-like (M3), and the manual method (M1) were 95.7%, 60.2%, 44.3%, and 76.6%, respectively. When combining M4 with human detection (M1), its sensitivity reached 99.9%, significantly higher than that of any other method alone or combined. M4 reported the highest number (1435) of false-positive VOIs, compared with 837 for M2 and 151 for M3.

Conclusion

The use of some clinically available AI algorithms could provide a robust safety net to physicians for lesion detection on [18F]FDG PET/CT, with more than 99% sensitivity. Its routine use is nevertheless currently limited by its high false-positive rate.