Predictive Model For Preliminary Assessment of Repeated Radioiodine Therapy Course Based On 131-I Suv

Purposes: Current research design is dedicated to 131-I SUV evaluation on post therapeutic scintigrams and setting up predictive model for radioiodine therapy repeated course prescription in differentiated thyroid cancer (DTC). Methods: Study includes 148 patients (f-105, m-43) with DTC treated with 131-iodine. Administered treatment activities of 131I calculated according to clinical features and tumor recurrence risk group. Patients were divided into three groups using ATA 2018 recommendations. Absolute risk groups: low risk (L), medium risk (M), high risk (H). Administered 131-I activity [MBq]: <AL> = 3223±729, <AM> = 3696±456, <AH> = 4589±1078; Tg [ng/ml]: <TgL> = 7,4±1.7 ng/ml, <TgM>= 14,8 ±5,9, <TgH> = 68,3 ±18,5; TgAb [IU/ml]: <TgAbL> = 124,3 ± 81.7, <TgAbM> = 29,2 ±15,9, <TgAbH> = 85,7 ±28,9. All null hypothesis were checked using paired Mann-Whitney U-test. Calibration of system SPECT/CT and evaluation SUVs completed according to protocols designed using Jaszczak phantom Deluxe. Model development based on logistic regression with ROC-analysis, regularization and cross-validation. Results: Reference intervals of SUVpeak and SUVmax calculated for all groups of risk. SUVpeak: low risk >155, medium risk 105-155, high risk 0-105 pL-M=0.069 pL-H=0.0037 pM-H=0.7514; SUVmax: low risk >38, medium risk 29-38, high risk 0-29 pL-M=0.052 pL-H=0.0033 pM-H=0.949. Logit model based on SUVpeak without regularization has: AUC = 0.67 (95%CI 0.33-1.00); Accuracy = 0.82; SE = 0.89; SP = 0.4; PPV = 0.41; NPV = 0.89, cross-validated AUC =0.67 (0.4-0.88), regression coefficients: B0=0.037, B1=0.001. Regularization (SUVpeak<100) lead to AUC = 0.75, (95%CI 0.44-1.00); accuracy = 0.89; SE =0.98; SP = 0.4; PPV = 0.76; NPV = 0.87, cross-validated AUC = 0.513 (0.36-0.71) regression coefficients: B0=0.473, B1=0.003. Conclusion: Study shows that SUV has wide range of values and can be matched with existed model of risk assessment of DTC. Algorithm of image segmentation and evaluation of SUVpeak and SUVmax for SPECT/CT systems was developed. According to the ROC analysis, developed predictive model shows an acceptable performance for further clinical investigation and advancement focused on refining model parameters and introducing additional predictors Uptake Value (SUV), SPECT/CT, radioiodine therapy, differentiated thyroid cancer, image


Introduction
People have faced different diseases throughout human history and seek to find effective methods for solving the problems. Currently, medicine is quickly developing. It combines technologies and achievements of fundamental scientific disciplines such as medicine, physics, chemistry and biology. Interdisciplinary approach to problem increases the diagnostic efficiency for revealing diseases at an early stage significantly.
Standardized Uptake Value (SUV) of radiopharmaceuticals (RPs) is one of the most perspective group of parameters in nuclear medicine [1]. The SUVs are feature of quantitative analysis with application of hybrid SPECT/CT systems, that have proved its usefulness in PET/CT but absence of it's application in SPECT/CT set an ambitious task for evidence-based instrumental diagnostics [2,3]. The departments change their own approach to treatment patients from classical such as evaluation of current uptake value (counts in region of interest (ROI) divided by counts in whole body mode on scintigraphy image) and description of ROI, to quantitativeradiomics.
Radiomics is based on creating a complex of digital patient parameters on image different modality [4].
SUV is a part of radiomics and describes tissue pharmacokinetics in current state. For assessing the value investigators conduct phantom studies and design a sequence of measurements [5]. One of the main advantages of SPECT/CT visualization is a lot of medical radionuclide compared with PET/CT. It gives an opportunity to expand the scope of quantitative assessment metabolic activity in addition to the PET facilities [6,7].
Quantitative description of metabolic activity tissue is an important part of treatment. It allows to define therapeutic activity for achievement of treatment effect from radionuclide therapy. In addition, the approach to processing images gives the opportunity to clarify risk of absolute relapse. Consequently, doctors can make predictions about the number of repeated radionuclide therapy courses [8].
Annual increase of new endocrine disease cases equals 4.5 % [9]. Among endocrine system diseases, the leading position has pathologies of thyroid gland (malignant and benign neoplasms) [10]. Therefore, there is a need to improve the quality of diagnostics. Processing images transform from description level to quantitative, which includes development and implementation method for measurement of metabolic activity thyroid tissue and metastasis process using SPECT/CT system. There are countries such as South Korea, the USA, Germany and Japan that actively implement SUVs in medical protocols [11][12][13].
Current research design is dedicated to 131-I SUV evaluation on post therapeutic scintigrams, receiving references range of SUVpeak and SUVmax depending on absolute risk of relapse in differentiated thyroid cancer (DTC). Predictive model for appointment the second course of radioiodine therapy (RIT) in DTC was created to improve effective outcome and acceptable safety.

Patient inclusion
Present study included 250 patients with diagnosed DTC and after radioiodine therapy was done in the nuclear medicine department (Endocrinology Research Center) from January to December in 2018. There were 198 patients with confirmed histology and 148 of them had information about level Tg and AT-Tg. Children were excluded from the research. After thyroidectomy patients had euthyroid status or obtained suppressive therapy with levothyroxine. Prescribed activities of 131I calculation were based on protocol of medical profile. Patient cohort was divided into three groups according to (ATA 2018) TNM risk assessment [14].

Sensitivity test of SPECT/CT system
The camera calibration factor (sensitivity) S is determined according to ̇ is the count rate measured in the VOI, and Avoi is the decay-corrected activity in the VOI. Deluxe Jasjczak (Model ECT/DLX/ P) was completely filled with water (without air bubbles). Furthermore, activity of 131-I was injected in the volume of phantom (activity in syringe A=248 MBq) (Fig 3.). Finally, information about characteristics of radiation source was combined using SPECT/CT system (the first part of scan was emission and the second was CT-part) (  Eventually, the results of each scans were transferred to Xeleris workstation. There was Volumetrix MI (GE Healthcare). The application is created for processing of tomography images where image reconstruction is performed. A reconstruction was done with certain parameters, which were considered at the top of the topic. The reconstructed image was processed using MATLAB.

Study of OSEM algorithm. The dependence of the sensitivity on the number of iterations
The research conducted with 131-I, but from further experiments it can be concluded that the results are identical for any radionuclides. Activity of radionuclide in phantom during the study was 22,3 MBq with correction of radiation decay. Average count with image and slice sensitivity were studied during processing the reconstructed images.

FWHM study depending on the source positioning depth
In order to determine FWHM of SPECT/CT (emission part) insulin syringes with 131-iodine (model of point source) were put in phantom volume in distinct depth in a special cap (Fig 5).  The syringe was as close as possible to the side wall of the phantom. The last time insulin syringe was put at a depth of 9.24 cm (Fig 7). Therapeutic activity of RPs was determined 72 hours after administration with correction for decay. After that, scanning of patient was carried out using SPECT/CT system. Subsequently, information about the research was picked up: total research score, total count in sphere 1 ml or total count in ellipsoid with linear size getting with CT, maximum count in voxel, volume of ellipsoid with linear size getting with CT or volume sphere 1 ml.
Total research score was inserted in dependence of the tomography sensitivity from projection count rate.
The sensitivity value was substituted in formula to find activity in ROI, respectively: A= • Ssensitivity (this value from phantom study. count/MBq*sec), Сount -Total count in ROI after segmentation it, ttime of exposition (sec).
Three type of activity were calculated during the study:

Statistical analysis
In this research, the GraphPad Prism 8.0 package was used for statistical analysis. The instrument was used for processing and interpretation results of the investigation. Also it was a foundation for formation of SUVs reference intervals. At first, for each cohort was done the normal test and for all parameters. Normal test was failed for all cohorts because of it the data were processed like nonparametric methods. Also, according to all available evidence for statistical analysis was chosen non-parametric Mann-Whitney test.

Logistic regression
Logistic regression is used for predicting the likelihood of some event, when there is a binary resultyes/no disease [15]. In contrast to linear regression logistic regression doesn't predict numeric variable values from a sample of initial values. Instead of it, value of function is a probability that initial value belongs to a certain class. Logistic regression equation: (5) p -probability that a patient with the predictor has positive results; B0, ; х1, х2 …. хncorresponding predictor value. Method of maximum likelihood estimation is utilized to get logistic regression coefficients.After potentiating the formula 5 (6) Let stand t, t -function of chances. The ratio of chances is (7) After that the resulting P-value is compared with the selected threshold to conclude on true-positive cases.

Creation of logistic regression model.
At the beginning of creating a logistic regression model all patients were separated into several groups.
Afterwards, training and test cohorts were set up. The model was based on 74 patients with high risk of absolute relapse according to ATA18 from the total cohort (148 patients). Among 74 patients was found 9 with proved the second course of radioiodine therapy. As a training group was distinguished cohort which was made up of patients with repeat (9 patients) and without it (11 patients). Test group contained 63 patients (9 with repeat and 54 without repeat). Logistic regression model was developed using MATLAB.

ROC-analysis
Cross-validation is a method for estimating work the model on a sample of independent data. ROCanalysis is a significant instrument, which allows to find parameters for getting the best and the worst results [16].
During the research was used following parameters: • Sensitivity (Se)percentage of correctly predicted repeated courses of RIT.
• Specificity(Sp)percentage of correctly predicted RIT without relapse • Accuracypercentage of correctly predictions relative to the total number of observations Another indication of quality for the predictive model functioning is ROC (receiver operating characteristic)-curve. This method is based on a graphical representation of the relationship between sensitivity and specificity. It allows to analyze them and to solve an important problem linking with set a threshold for making diagnostic decisions at which the research errors will be minimal. In order to appreciate diagnostic ability of test is put up with a decision threshold. if P-value is higher than threshold than test will be positive and if P-value is lower than threshold than test will be negative. The most essential quantity that characterizes the ROC-curve is area under ROC-curve (AUC). AUC is determined according to (8) where X and Y -specificity and sensitivity, respectively.

Results
Dependence of the tomography sensitivity from count rate is linear (Fig 8). The graph shows that with increasing rate, the sensitivity decreases. Average value of sensitivity is 76 ± 3 cps/MBq/sec. Fig. 8. The dependence of the sensitivity from the count rate in the phantom volume for iodine-131.
All scans were completed using Body Contour (device for po because sensitivity of SPECT is the same without Body Contour (device for positioning the gamma camera over the patient body) Therefore, Body Contour doesn't have affect on research results.
During the study the dependence of the sensitivity on the number of iterations was obtained (Fig 9). Count value with increasing number of iterations was a gateway to the plateau and after 5 iterations was changed slightly. It allows to decrease processing time and minimize the influence of OSEM on reconstruction. The investigation showed that behavior of FWHM changed slightly with depth (Fig 10). The deviation is observed in the area of depths more than 6 cm. It is able to connect with edge effect on reconstruction images.
Radiologists' segmentation manual for fused SPECT/CT was developed during the study. Segmentation was carried out with helping software -GE Medical Systems Xeleris 4.0 (Volumetrix MI) (Fig 11).

Fig 11. Algorithm segmentation
Statistical analysis was carried out for three groups dividing according to absolute risk of relapse.
Descriptive statistics are put forward in table 1.  The second part of the investigation included a predictive model with filtration. According to the learning group was obtained B0= 0.473 B1=0.003 -logistic regression coefficients.The most appreciate parameters of the model were defined using visualization of parameter changes (Fig.13). The optimal threshold and set of model parameters for the logistic regression coefficients are р=0.67 and Se=46%, Sp=98%, Accuracy=0.89, PPV=76% NPV=87%., respectively.  The research showed that SUV will be able to become a part of digital patient profile in future. The main result is obtaining another tool for improving the quality of medical care.

Discussion
The calibration curve of SPECT/CT system was obtained During the first part of the investigation.
However we found some characteristics which were suitable for realization of our main idea (getting an SUV for radionuclide therapy with SPECT/CT image). Majority of SPECT/CT parameters were performed by Jonathan Gear and et al. [17] The curve depends on the mean count within projection. Based on it, we made a decision that sensitivity decreases with increasing activity. The dependence is leaner. This behavior of sensitivity is a consequence of a growth in the number of decays per unit time and an overload of the spectrometric path with a large number of signals. If we didn't take into account the fact, it would not reach the universal method of determining SUV for SPECT/CT system.
The second step included investigation of spatial resolution. Work of the approach depends on spatial resolution. The value is one of the most important for our method of segmentation. The less spatial resolution of SPECT/CT, the more accuracy of SUV. Also there is an effect of partial volume for small structures. The phenomena is negative for any visualization system. There are some methods to reduce the contribution of this effect but it isn't a cheap way. [18].
During the study we distinguished exactly only 2 groups (low and high risk). According to statistical analysis, to define a medium group, it is necessary to increase the sample of patients.
As for predictive model, it is the first quantitative instrument for estimating risk of relapse DTC. Our model has some disadvantages which are a part of some innovation. Nevertheless, we have precise suggestions on how to make the model better. The main offer is to increase the number of patients in studies with a prescribed repeated course of radioiodine therapy. And the second part we are planning to include thyroglobulin (Tg) in the modal. It will give us the opportunity to create the first model embracing clinical and physical parameters.

Conclusion
In conclusion, this study reports the utility of SUV for SPECT/CT system. It was the first investigation which adopted SUV in radionuclide therapy but it is definitely not the least. Now we have some ideas to make our model more simple and universal for realization. There are a lot of directions to apply the new method in nuclear medicine.