SUVmax described in FDG PET-CT reports can provide information of lesion location: an investigation of real-world data

Purpose: FDG PET-CT is a well-established in-vivo imaging modality in oncology from initial staging to treatment response evaluation. Because the intensity of FDG uptake is significantly associated with malignancy of the lesion, maximum standardized uptake value (SUVmax) is frequently written in diagnostic reports of FDG PET-CT. SUVmax has a great advantage of extremely high inter-operator reproducibility, where slight difference of ROI definition (size or position) does not affect SUVmax. We hypothesized that SUVmax can be used to identify the voxel and to localize the ‘tumor of interest’ appearing in diagnostic reports. We presented the preliminary results at SNMMI 2020 that precisely described SUVmax (e.g., 3.142 rather than 3.1) and local maximum restriction were useful to locate the lesion. However, the previous study was based on a simulation but not on the real-world reports. Thus, the aim of the current study was to investigate the actual reports in our institute to clarify whether SUVmax can be used to localize the tumor.

Methods: The institutional review board approved the retrospective study. Using electronic medical record, we reviewed a total of 230 reports of FDG PET-CT from 30 days (September 11 to 20 in 2011, May 11 to 20 in 2014, and January 11 to 20 in 2017). First, regular expression and additional algorithms were used to extract strings potentially meaning SUVmax. Then, a board-certificated nuclear medicine physician carefully reviewed the reports in relation to the images to determine whether each computer-extracted string truly meant the SUVmax of the lesion. We built an in-house software package that searches voxels satisfying the given SUV range in the whole-body image. The SUV range was determined by rounding; for instance, when ‘3.14’ was provided in the report, the program automatically extracted the voxels satisfying 3.135≤SUV<3.145. The program extracted the local maximum voxels (i.e., the voxel that has higher value than any other neighbor voxels). When only 1 voxel was extracted by the program, we considered the lesion was successfully localized.

Results: Of 230 repots, the algorithm detected 159 reports (69%) that contained a total of 435 SUVmax-like strings (i.e., 2.74 SUVmax-like strings/report in average), among which the physician judged that 333 (77%) were truly SUVmax. The reasons of exclusion included 1) SUVmean, 2) SUVmax of previous study, and 3) date string. Among 333 SUVmax, 246 were written in 1st decimal places (DP) (e.g., 3.1) while 80 were in 3rd DP (e.g., 3.142). When SUVmax was written in 1st DP, the lesion was successfully localized in only 71 of 246 (29%) cases. In contrast, 71 of 80 (91%) lesions were successfully localized when SUVmax was written in 3rd DP. In the further analysis, the successful rate was also affected by SUVmax value itself. More specifically, successful rate was 30% (2≤SUVmax<5), 48% (5≤SUVmax<10), 73% (10≤SUVmax<20), and 91% (20≤SUVmax), respectively (Table).

Conclusions: The current analysis of the real-world data suggested that SUVmax written in diagnostic reports may be useful to localize the lesion, if SUVmax is written precisely and is high. If the SUVmax does not satisfy the condition, further algorithms, such as use of anatomical terms in the same sentence, are needed to localize the lesion. The current algorithm will contribute to constructing training data for AI.

• Download figure
• Open in new tab
• Download powerpoint