How many digits after decimal point are needed to use SUVmax as an identifier of tumors on FDG PET-CT? CURRENT

Purpose The maximum standardized uptake value (SUVmax) is often described in daily clinical reports of FDG PET-CT. We investigated whether it would be possible to localize the voxel automatically based on the SUVmax. Methods The institutional review board approved this retrospective study. We investigated a total of 112 lesions from 30 FDG PET-CT images acquired with 3 different scanners. The number of voxels showing the given SUVmax was counted, where SUVmax was provided with various degrees of precision, such as 3, 3.1, 3.14, 3.142. The effects of local maximum restriction, where only local maximum voxels were chosen, were evaluated. Results SUVmax ranged from 1.3 to 49.1 (median = 5.6, IQR = 5.2). Generally, when larger and more precise SUVmax values were given, fewer voxels were included in the range. The local maximum restriction was very effective. When SUVmax was determined to 4 decimal places (e.g., 3.1416) and the local maximum restriction was applied, 33.3% (lesions with SUVmax<2), 79.5% (2≤SUVmax<5), and 97.8% of lesions (5≤SUVmax) were successfully identified (i.e., only a single voxel satisfied the criteria).


Introduction
The clinical usefulness of positron emission tomography (PET) using 18F-fluorodeoxyglucose (FDG) has been well established in oncology (1). In addition to visual assessment (qualitative analysis), several quantitative measurements have been used to express the degree of FDG uptake. Among them, the standardized uptake value (SUV) has long been used as the de facto standard. To our knowledge, SUV was first extensively used around 1991 (2). In the initial years of its use, SUV was also known as the differential uptake ratio (3) or dose uptake ratio (4). The in-lesion maximum of SUV, or SUVmax, has frequently been used since 1999. By 2009 SUVmax had become the most frequently used measurement by far, with 6-fold more frequent use compared to the next most-often used measurement, according to a comprehensive review (5). Although SUV is most commonly calculated as the radioactivity concentration normalized to injection dosage and body weight, other definitions include the radioactivity concentration normalized to the body surface area (6), to lean body mass (5), and to blood glucose (7). While metabolic tumor volume and total lesion glycolysis have been extensively investigated in recent studies (8,9), SUVmax is still superior to them in terms of its extremely high inter-operator reproducibility. Many lines of evidence have demonstrated the usefulness of SUVmax for differential diagnosis, treatment response prediction, and prognosis (10).
In 2015, 229.2 CT and 113.4 MRI examinations per 1000 population were performed (11). 42.1 radiologists per million population were employed, and thus 8137 CT/MRI examinations were performed per radiologist (11). As for nuclear medicine, 0.6 million FDG PET studies were performed in Japan in 2017 (12). A significant number of these studies require radiologist(s) to interpret the images and write the reports. Describing intensity of FDG uptake either using SUVmax or qualitatively has been recommended (13). Radiology reports not only contribute greatly to helping the attending physician interpret the image and diagnose the disease, but also prevent important findings from being overlooked. More recently, in the era of artificial intelligence (AI), the importance of training data is increasing. Collectively, radiology reports form a highly useful and efficient training database (14)(15)(16)(17)(18).
We hypothesized that if the SUVmax described in radiology reports was sufficiently precise, it might contribute to localization of the lesion, because there should be a limited number of voxels showing the same SUVmax in the entire image. In other words, we thought that SUVmax could be used as an identifier of the lesion. Thus, in this study, we aimed to clarify whether it would be possible to identify the lesion location using the SUVmax under various conditions by varying the degree of SUVmax precision and applying local maximum restriction. Such a technique could also be used to realize an automated system to generate a visual summary of the radiology report (Fig. 1).

Study subjects
This retrospective observation study was approved by the institutional review board (approval no. 017-0454). The requirement of written informed consent from each patient was waived because of the study's retrospective nature. We confirmed that all methods were carried out in line with the relevant guidelines and regulations. A total of 30 PET-CT scans (sequential examinations for each scanner) were investigated in this study. No more than one scan was included for each patient. All the images were acquired between April 2019 and November 2019. Images were evaluated visually, and included to the study population if there were any pathological FDG uptakes in visual analysis until the number of scans reached 10 for each scanner. When all the FDG accumulation mass were considered physiological, the case was excluded. Note that not only uptake due to pathological malignancy but also malignancy-suspected and inflammatory uptakes were included in the analysis.
In cases in which more than 5 uptakes were found, a maximum of 5 uptakes were recorded for a patient. An experienced nuclear medicine physician visually evaluated all the images.

PET-CT image acquisition and reconstruction
In this study, we investigated images acquired with 3 different PET-CT scanners made by 2 different manufacturers. Scanner 3 was a Vereos PET-CT (Philips Japan, Tokyo), which was the newest scanner of the three and equipped with digital photon counting detectors (19). The transaxial and axial fields of view were 67.6 cm and 16.4 cm, respectively (19). Images were reconstructed using the OSEM algorithm. Both The number of voxels in the z-direction (i.e., cranio-caudal direction) ranged from 234 to 553, resulting in the final number of voxels ranging from 4.85 × 10 6 to 4.41 × 10 7 . CT images were used for attenuation correction for all the scanners and for visual assessment, but were not analyzed quantitatively in the current study. All patients fasted for ≥ 6 hours before the injection of FDG (approx. 4 MBq/kg), and the emission scanning was initiated basically around 60 min post-injection.
One scan was acquired 95 min post-injection due to mechanical troubles. Fasting blood sugar was confirmed to be smaller than 200 mg/dl in each study.

SUVmax calculation
Commercially available DICOM viewers / PET viewers do not display SUVmax to 4 decimal places (DP) or higher. In order to obtain the ground truth of SUVmax, we modified Metavol software package, which we previously developed for PET-CT volumetric analysis (20). We used Windows 10, Microsoft Visual Studio Community 2019 Version 16.4.0, C# 8.0 language, .NET Core 3.1, and fo-dicom 4.0.3 for modifying Metavol. For instance, in the case that the true SUVmax is 3.14159, the modified Metavol will display it as it is, whereas XTREK VIEW software (J-MAC SYSTEM, Sapporo, Japan) will display it as 3.142. A nuclear medicine physician measured SUVmax by placing a spherical volume of interest (VOI) whose diameter can be changed by the operator.
After the VOI definition, SUVmax was calculated based on the newest QIBA publication (21). Briefly, in Biograph64 and Vereos, the radioactivity concentration c (Bq/ml) was calculated as follows: Here, ρ represents the raw pixel value that was stored with DICOM tag of (7FE0,0010) with each voxel expressed in a 16-bit integer. s represents the rescale slope, which is stored as a float value at (0028,1053). i represents the rescale intercept, which is stored as a float value at (0028,1052). Next, decay-corrected injection dosage D c was calculated as follows: Here, D 0 represents the radionuclide total dose (i.e., injected dosage of FDG) (Bq) stored as a float value at (0018,1074). T a represents acquisition time stored at (0008,0032). T i represents the radiopharmaceutical start time (i.e., injection time) stored at (0018,1072). Both times are stored in a "hhmmss" form string, and thus conversion to second is needed. h represents the radionuclide halflife (second) stored as a float value at (0018,1075).
Finally, SUV was calculated as follows: Here, w represents the patient's weight (g), which is stored in kilograms at (0010,1030) and thus must be multiplied by 1000.
The SUV calculation was much simpler in GEMINI TF64, as follows: Here, p represents the Philips Factor (float) stored as a float value at (7053, 1000). The values of s and i were 1 and 0, respectively, for all the GEMINI TF64 examinations investigated in the current study.

Lesion localization
We implemented a function that searches voxels satisfying the given SUV range and illustrate the locations in the whole body image (Figs. 2-4). The SUV range was determined as follows. When "3" was provided by the operator, the range was considered to be 2.5 ≤ SUV < 3.5. When "3.1" was provided, the voxels satisfying 3.05 ≤ SUV < 3.15 were picked out, and so forth. Thus, the more precise the provided value of SUVmax (i.e., more digits) was, the narrower the range of SUV applied to extract voxels was. We compared the results from integer precision to 4th DP precision. Note that we do not show the results of 5th DP precision because there were no cases in which 5th DP precision improved the identification rate compared to 4th DP precision.
We performed experiments in different settings. First, the voxels within the range were extracted simply. Then, local maximum restriction was added to discard the voxel that was adjacent to the higher-value voxel, because such a voxel cannot have SUVmax. For local maximum restriction, milder restriction and stricter restriction were tested. Milder restriction was a condition under which the voxel must be highest in the 3 × 3 × 3 cube. Stricter restriction was a condition under which the voxel must be highest in the 5 × 5 × 5 cube.
Here, we defined that "identical detection" was achieved when only 1 voxel satisfied the criterion.

Statistical analysis
The relationship between SUVmax vs. the number of voxels detected (N) was estimated using Pearson's correlation coefficient of the log of SUVmax vs. the log of N. The effects of the precision of SUVmax, i.e., the number of digits after the decimal point, and local maximum restriction on the rate of identical detection were tested using a chi-square test. P values less than 0.05 were considered statistically significant.

Results
Patient characteristics are summarized in Table 1. Diagnosis and lesion locations are summarized in Table 2. In this study population, head-and-neck cancer was the most common diagnosis, and the mediastinal and hilar lymph nodes were the most frequent locations. In the 112 lesions investigated, SUVmax ranged from 1.3 to 49.1, with median and IQR values of 5.6 and 5.2, respectively. SUVmax was significantly higher for Vereos than for Biograph64 and TF64 (P < 0.01 and P < 0.05, respectively; Supplementary Fig. 1). The numbers of lesions for Biograph64, GEMINI TF64, and Vereos were 37, 37, and 38, respectively.  First, local maximum restriction was not applied. A number of voxels were identified corresponding to the given SUVmax (Fig. 5). Generally, when a larger SUVmax was given, a smaller number of voxels was detected (0.83<|r|<0.84, P < 10 − 28 ). Because the SUVmax was given with 10-fold greater precision, an approximately 0.1-fold number of voxels were extracted, as expected theoretically.
Next, local maximum restriction was applied. Both 3 × 3 × 3 and 5 × 5 × 5 restriction reduced the number of extracted voxels up to 1/1000 (Fig. 6). More specifically, the rate of identical detection increased because the given SUVmax was more precise and local maximum restriction was stricter ( Fig. 7). For instance, while identical detection was successful only in 2.7% of patients when integer precision and no restriction were used, the success rate was elevated to 86.6% when 4th DP precision and 5 × 5 × 5 restriction were used. The effects of 5 × 5 × 5 restriction over 3 × 3 × 3 restriction are shown in Fig. 7 except for integer precision, although none of these effects reached the level of statistical significance (P > 0.05).

Discussion
In this retrospective study, we aimed to clarify whether SUVmax can be used as a lesion identifier to localize the voxel in the whole-body image of FDG PET. We observed that SUVmax successfully localized the voxel for > 80% examinations in the case that SUVmax was given to the 3rd or higher DP and local maximum restriction (5 × 5 × 5) was applied. However, the sub-analysis showed that the lesions having SUVmax < 2 were difficult to localize using SUVmax only.
The pixel data was stored in DICOM files in a 16-bit integer form for all 3 scanners investigated. A 16bit integer can express 65,536 different values. Since the number of voxels in the entire image was around 10 7 , theoretically > 100 voxels on average may have exactly the same value. In fact, the distribution of SUV was quite skewed, as shown in Supplementary Fig. 2. It is reasonable that many voxels were detected when a smaller SUVmax was given, whereas only one voxel was detected when a larger SUV (e.g. >5) was given. In Fig. 5, the number of voxels suddenly dropped once SUVmax became larger than 10. This can be explained as follows. In this study, we used DP instead of significant figures. They are slightly but clearly different. DP means the number of digits located to the right of the decimal point. Significant figures refers to the total number of digits irrespective of the decimal point location. For example, 9.8 is 1st DP and 2 significant figures, whereas 12.3 is 1st DP and 3 significant figures. Since 12.3 has more information than 9.8, fewer voxels were included within the range.
The effect of local maximum restriction was significant. The number of voxels that can become local maxima depends on the noise level of the image. When 3 × 3 × 3 restriction was applied, at most 1 of 2 voxels in each axis could become local maxima, indicating that 1/8 or a smaller number of voxels could become local maxima. Similarly, when 5 × 5 × 5 restriction was applied, at most 1 of 3 voxels in each axis and thus 1/27 or a smaller number of voxels could become local maxima. We did not try 7 × 7 × 7 restriction because we were worried that it might prevent identification of the voxel of 7 voxels account for as much as 48.5 mm.
Some may argue that use of the 3rd or higher DP for SUVmax is redundant for daily radiological reports. That is true. SUV calculation uses body weight and the precision of body weight measurement may be 3 significant figures (e.g., 56.7 kg) or less. Radioactivity dosage cannot be measured so precisely. We think this is why SUVmax is often written to the 1st DP (e.g., 3.1).
However, in order to permit the future use of SUVmax as an identifier, we would like to propose that SUVmax be written as precisely as the viewer allows. As mentioned before, this use of SUVmax would allow the radiology report to be summarized as a figure (Fig. 1). In addition, it may also help radiologists to locate a lesion mentioned in a previous report so as to compare between past and present images. Our ultimate goal is to build a massive training dataset based on radiology reports and corresponding images. Writing the coordinate values (x,y,z) in the reports will be the best way to transfer the information to artificial intelligence. Currently, that may not be possible in most viewers and reporting systems. Also, the appearance of such information in the middle of a report may distract readers, and thus an automated system is needed to hide this information when humans read the report.
The use of SUVmax is specific for PET. Although the maximum voxel value may not often be useful for CT or MRI, the idea could be applied to the apparent diffusion coefficient (ADC) images derived from diffusion weighted imaging of MRI, because the minimum of ADC is meaningful for diagnosis.
As limitations of the current study, we did not investigate the SUVmax shown in different image viewers. In some viewers, PET volumes are reconstructed (resliced) in the CT alignment, making slight changes to SUVmax. Second, we did not directly use the radiology reports but reviewed the images to re-measure SUVmax. Thus, we could not estimate the number of actual cases in which the SUVmax written in the reports could successfully locate the lesion. Such a study will need to be carried out.
Finally, radiology reports often provide anatomical terms in the same sentence with SUVmax. This would be great information for selecting the appropriate location when SUVmax suggests several candidates, as in Fig. 3. Such a method will be tested in future studies.

Conclusion
The data suggested that SUVmax with 2nd or higher DP successfully located the lesion with > 80% probability, although in the case of SUVmax < 2, the success rate was low. SUVmax may be useful  Figure 1 A conceptual image of AI generating the visual summary of the report of FDG PET. SUVmax in the sentence appearing in the report text is used for localization. In this case, the primary lesion (right palatine tonsil) and metastatic nodes show high FDG uptakes.

Figure 5
The number of voxels extracted by a given SUVmax with various levels of precision. Local maximum restriction was not applied.

Figure 6
The number of voxels extracted by a given SUVmax with various levels of precision. Local maximum restriction was applied.

Figure 7
The overall rate of identical detection of the lesion. DP = decimal places. Free, 3×3×3, and 5×5×5 express no restriction and each local maximum restriction.

Supplementary Files
This is a list of supplementary files associated with this preprint. Click to download. Supl20190105.pptx