Abstract
3206
Introduction: FDG PET-CT visualizes glucose metabolism of tumor and thus is useful in many clinical settings from initial staging to re-staging and treatment monitoring. To describe the intensity of FDG uptake semi-quantitatively, maximum standardized uptake value (SUVmax) is most frequently described in diagnostic reports of FDG PET-CT. SUVmax is superior to the other indicators in that it is highly reproducible between operators, because SUVmax is not affected by slight difference of ROI definition (size or position). Considering such a high reproducibility, we have investigated the use of SUVmax string appearing in the diagnostic report as an identifier of the voxel in the tumor. We demonstrated that, in an experimental study, precisely described SUVmax (e.g., 2.718 rather than 2.72) was useful to locate the lesion (Hirata et al. Front. Med. 2021). Subsequently, we showed that the method could work efficiently for real-world reports (Hirata et al. SNMMI 2021). However, we also observed that voxel identification sometimes failed, especially when the number of SUVmax digits after the decimal point was low (e.g., 2.7). We thought that anatomical words near the SUVmax string can be used to solve the problem. Thus, the purpose of this study is to generate probability maps of representative anatomical terms to help identify the voxel.
Methods: The institutional review board approved the retrospective study. First, from all the images of FDG-PET/CT in 2019, we used an in-house deep-learning model to extract 1847 FDG-PET/CT images where the field-of-view was brain to thigh, which was confirmed by human eye. For each image, the diagnostic report written in Japanese was available. Regular expression and additional algorithms were used to find SUVmax strings. Our program searched voxels satisfying the given SUV range in the whole-body image. The SUV range was determined as follows. When ‘2.72' was described in the report, the program automatically extracted the local maximum voxels satisfying 2.715≤SUV<2.725. When only 1 voxel was chosen by the program, we defined the SUVmax string as ‘identifier SUVmax'. If the sentence coincidentally included identifier SUVmax and a given anatomical term (e.g., lung), the corresponding location was painted in the probability map after body size normalization of the whole-body image.
Results: A total of 636 sentences that coincidentally included the identifier SUVmax and ‘lung' were extracted automatically, whereas 101 sentences had non-identifier SUVmax and ‘lung'. The probability map displayed in maximum intensity projection was visually consistent to the location of both lungs, although some parts other than lung (e.g., brain, thigh) was slightly highlighted (Fig). Similarly, neck (185 sentences), mediastinum (216), liver (169), pancreas (39), adrenal grand (41) were successfully localized in the probability map. Each portion of the colon (i.e., ascending, transverse, descending, sigmoid colon, and rectum) was also successfully visualized (Fig).
Conclusions: Using the combination of existing image and its report of FDG-PET/CT, we successfully constructed probability maps for several anatomical terms, which will be useful for less precise (non-identifier) SUVmax string to identify the single voxel. This method will help prepare training dataset retrospectively from a huge number of existing clinical data.
In this issue
Jump to section
Related Articles
Cited By...
- No citing articles found.