Impact of Convolutional Neural Network Deep Learning Algorithms on Clinical F18 FDG PET/CT Reporting Time in Staging Esophageal Malignancies

Introduction: Convolutional neural network (CNN) deep learning algorithms can be used to assist lesion segmentation and classification. The algorithm was previously trained by using numerous annotated patient studies exhibiting various malignancies to differentiate physiological from suspicious activity via a two stage process: activity selection and lesion segmentation followed by lesion characterisation. This aims to assist readers by suggesting inclusion or exclusion of areas of uptake. Specifically, it is not designed to fully automate the task of reading. The aim of our study was to assess the impact of the use of a CNN on reporting times in a clinical setting.

Methods: Thirty staging F18 FDG PET/CT scans for cancer of esophagus (adenocarcinomas and squamous cell carcinomas) were anonymised for the purpose of this study. These were read by three experienced PET/CT reporters separately - two PET/CT radiologists (with18 years of experience each) and a nuclear medicine physician (with 7 years of experience). Reporting times were recorded including interpretation, dictation and checking of reports. The first routine read was performed using manual volume of interest tools available on syngo.via (VB50, Siemens Healthineers). The second read was performed 5 weeks later. Between the 2 reads, reporters had reported 230 cases on average. The second read was performed in random order aided by the CNN (Lesion Scout with Auto ID, Siemens Healthineers) enabled with optimal segmentation criteria. In accordance with the EANM procedure guidelines for tumor imaging: version 2.0, 3D isocontour at 41 % of the maximum pixel value for segmentation of the FDG-avid lesions was used. The total times for case interpretation, dictation and checking reports on each separate method were compared using descriptive statistics and the non- parametric Wilcoxon signed-rank test.

Results: It was found that the reporting time reduced with the aid of Lesion Scout/ Auto ID compared to manual reporting by 32%, corresponding to a reduction of 5 +/- 2 mins per case. The Wilcoxon signed-rank test also shows a significant difference in reporting times when using Auto ID compared to manual reporting at the 95% significance level (p<1x10^-4); Figure 1.

Figure 1. Wilcoxon signed-rank test

**** indicates significance, p < 1x10^-4

Conclusions: In this small cohort of esophageal malignancies, reporting F18 FDG PET/CT scans with the aid of Lesion Scout/ Auto ID significantly reduces reporting time.

• Download figure
• Open in new tab
• Download powerpoint