Bone Scintigraphy in Cardiac Transthyretin-Related Amyloidosis: A Novel Time-Saving Tool for Semiquantitative Analysis, with Good Potential for Predicting Different Etiologies

Abstract

(1) Background: The visual and semiquantitative analysis of Technetium-99metastable-3,3-diphospono-1,2-propanodicarboxylic acid (^99mTc-DPD) bone scintigraphy is promising for diagnosing cardiac amyloidosis but time-consuming. We validated a faster method, the geometric mean (GM) method with a semi-automated workflow, for heart–whole body (WB) ratio (H/WBr), heart retention (Hr), and WB retention (WBr) calculations compared to the classic method (CM) established in the literature. The capability of semiquantitative scintigraphy indexes to differentiate the etiology in transthyretin-related cardiac amyloidosis (cATTR) patients was investigated. (2) Methods: H/WBr, Hr, and WBr were calculated by extracting counts for WB, kidneys, bladder, and heart on early and late planar image scans and applying background, scan-time, and decay corrections, using CM and GM both on a referring workstation and on a semi-automated workflow in external software. The comparison between CM and GM was assessed with Pearson’s correlation, Lin’s Concordance Correlation Coefficient (CCC), and Bland–Altman analysis. H/WBr, Hr, and WBr and several clinical variables were used to implement LASSO, Random Forest (RF), and Neural Network (NN) models to predict mutated and wild-type ATTR etiologies. ROC curves and AUC were calculated. (3) Results: Hr, WBr, and H/WBr using CM and GM were highly correlated. Bland–Altman analysis between CM and GM showed biases of 0.12% [CI:0.04%;0.19%] for H/WBr, 0.07% [CI: 0.01%; 0.13%] for Hr, and -0.50% [CI: −1.22%; 0.22%] for WBr. LASSO and NN models had good performance in predicting etiologies with AUC values of 87.3% and 73.6%, respectively. The RF model showed a poorer AUC of 55.8%. (4) Conclusions: The GM in the assisted workflow was validated against the CM. LASSO and NN approaches allowed a good prediction performance to be obtained for patient etiology.

1. Introduction

Amyloidosis is a systemic infiltrative disease caused by the formation and deposition of misfolded protein fibrils involving several organs, including the heart [1,2,3]. One of the most common forms of cardiac amyloidosis is transthyretin-related amyloidosis (ATTR), a tetrameric structured transport protein synthesized mainly by the liver. There are two types of amyloidosis resulting from the deposition of this protein [4]: a non-hereditary form known as “wild type” TTR amyloidosis (ATTRwt) that primarily involves the heart, particularly in male subjects older than 65 years, and a genetic form known as “hereditary TTR amyloidosis” (ATTRm), an autosomal dominant disease characterized by a particular tropism for nervous and/or cardiac tissue [2,5]. The frequency of these two types of amyloidosis in the population is related to several factors, including specific TTR mutations (there are more than 100 genetic mutations of the TTR gene), sex, and geographic distribution [6]. The cardiac phenotypic expressions are related to Thr60Ala, Leu111Met, Ile68Leu, and Val122Ile gene mutations [2]. Another form of systemic amyloidosis that frequently affects the heart is light chain amyloidosis (AL); this is secondary to the presence of clones of plasma cells in the bone marrow that produce free light chains of immunoglobulins in the circulation, which are responsible for fibrillar deposits. The extreme clinical heterogeneity of ATTR forms makes the diagnosis and treatment a real challenge; in fact, these forms are often underdiagnosed or diagnosed late, given the non-specificity of symptoms [7], and that might lead to a poor prognosis. The need for an early diagnosis of cardiac involvement in individuals with ATTR has been highlighted in the scientific literature, as it can significantly influence patient treatment [8].

The suspicion of amyloidotic cardiomyopathy demands several diagnostic tests, including an electrocardiogram (ECG), which may detect low peripheral voltages, pseudo-necrosis patterns, conduction delays, and repolarization changes, and echocardiography, which may identify left ventricular hypertrophy, valve and interatrial septum thickening, pericardial effusion, granular sparkling appearance, diastolic dysfunction, a reduction of global longitudinal strain, and apical sparing [9]. Unfortunately, none of these provides information on the etiologic type of amyloidosis [5]; thus, the “gold standard” examination for a definitive diagnosis of amyloid cardiomyopathy is an endomyocardial biopsy (EMB). This investigation has good accuracy in characterizing the proteins contained in the deposits (through immunohistochemical and mass spectrometry analysis) and in confirming the presence of amyloid; nevertheless, it is an invasive procedure that is not free from complications [10].

Currently, the use of EMB has been curtailed by the increasing diagnostic accuracy offered by other non-invasive imaging methods with technetium-99m (^99mTc)–diphosphonate (DPD),^99mTc–hydroxyl-methylene-diphosphonate (HMDP), or ^99mTc-pyrophosphate (PYP) bone scintigraphy [5]. In fact, according to the study by Gillmore et al. [11] bone scintigraphy can confidently define cardiac amyloidosis due to TTR in the presence of a moderate/intense cardiac uptake when the AL form can be excluded with dedicated serum–urinary investigations. On the contrary, cardiac uptake expressed by Scintigraphic Score can vary from zero to three in AL [11,12] and other extremely rare forms of cardiac amyloidosis. Furthermore, this diagnostic method can identify TTR deposits at a preclinical stage of the disease, when echocardiogram and biomarkers are still normal. In addition to cardiac visual score [13], several studies have evaluated the application of a semiquantitative analysis of cardiac uptake [9,14].

Unfortunately, the reported semiquantitative methodology is time-consuming, as it requires drawing several organs of interest (i.e., the heart, the kidneys, and the bladder) to be copied and mirrored from anterior to posterior views and computing the geometric mean (GM) among their average values [9], hereafter referred to as the classic method (CM). According to the first aim of this study, a new workflow designed to retrospectively evaluate semiquantitative cardiac uptake metrics such as heart retention (Hr), whole-body retention (WBr), and heart–whole body ratio (H/WBr) on the GM image was reported, and its agreement with the CM was assessed. As a second aim, we evaluated whether, within a population with a certain diagnosis of ATTR amyloidosis, the semiquantitative indexes obtained with the GM method can be used to predict the etiology of the two forms of ATTR (wild-type and hereditary).

2. Materials and Methods

2.1. Patient Cohort, Image Acquisition Protocol, and Visual Assessment

All patients with amyloidotic TTR cardiomyopathy that underwent planar scintigraphy with ^99mTc-DPD at the IRCCS Azienda Ospedaliero-Universitaria di Bologna, Italy, between January 2004 and September 2019 and had images stored on the local Picture Archiving and Communication System (PACS) were retrospectively enrolled.

Whole body (WB) planar scintigraphy was obtained with a dual-head camera (Discovery NM/CT 670, GE, Healthcare) equipped with low-energy, high-resolution (LEHR) collimators. WB scans were obtained 5 min (early image) and three hours (late image) after an injection of ^99mTc-DPD with a standard dose of 740 MBq in all patients. Anterior and posterior views were simultaneously acquired. Nuclear medicine (NM) physicians visually assessed myocardial uptake on ^99mTc-DPD scintigraphy on late images, according to Perugini et al. [13], assigning a visual cardiac score ranging from score 0 to 3: score 0, no cardiac uptake with normal bone uptake; score 1, mild cardiac uptake, inferior to bone uptake; score 2, moderate cardiac uptake accompanied by attenuated bone uptake; and score 3, strong cardiac uptake with mild/absent bone uptake.

2.2. ATTR Diagnosis and Classification

Amyloidotic cardiomyopathy was defined as left ventricular end-diastolic interventricular septum thickness ≥12 mm on echocardiography, in the absence of any other cause of hypertrophy. ATTR-CA was diagnosed by the demonstration of transthyretin amyloid deposits on EMB, typical echocardiographic findings associated with the demonstration of transthyretin amyloid deposits on extracardiac biopsy, or typical echocardiographic findings associated with cardiac uptake grade 2 or 3 on ^99mTc-DPD, and clonal plasma cells dyscrasia excluded by a serum free light chain assay and serum and urine protein electrophoresis with immunofixation. In all cases, genetic testing confirmed the absence or presence of mutations. For DNA analysis, peripheral blood was taken from the patients. Total genomic DNA was extracted, and the codifying region and the intron flanking regions of the TTR gene (the number of access on GenBank is M11844), found on chromosome 18 of the DNA sequence, were amplified by the PCR technique with the ABI PRISM 3130 Genetic Analyzer. The analysis was performed by the University of Ferrara, Italy. In ATTRm, there are single amino acid mutations in the 127 amino acid sequence that destabilize the heterotetramer, making aggregation more efficient. Although the nomenclature for ATTRm specifies the normal amino acid, its position, and the substituted amino acid (e.g., Val30Met signifies substitution for valine at position 30 by methionine) [15] and each mutation can be characterized by different phenotypes, for the aim of this work only ATTRwt versus ATTRm will be considered, without the additional differentiation of ATTRm subtypes.

2.3. Semiquantitative NM Indexes Calculation with the CM

On a random subset of 30 patients, a semiquantitative analysis of Hr, WBr, and H/WBr was calculated by the drawing region of interest (ROI) on target organs, following the CM described by Rapezzi et al. [9]. In brief, the heart, kidneys, and bladder were manually contoured by an NM physician on the anterior images of the early acquisition on a referring workstation (Xeleris Functional Imaging Workstation 4.0; GE Healthcare, Milwaukee, Brookfield, WI, USA). These ROIs were copied, mirrored, and manually positioned to match the organs on the posterior images. The same procedure was applied to late-acquisition images. The background correction (evaluated by placing a rectangular background ROI near the tibia) was applied on all ROIs to subtract the noise component spread over the entire image. Total counts of the images were taken as WB counts. For each ROI, the GM of the counts of the two projections was obtained. Early and late images were corrected for background signal, scan duration, and the physical decay between early and late acquisitions. Data were extracted in .csv format for the computation of Hr, WBr, and H/WBr.

In particular, the scan duration correction was applied considering the actual frame duration of the early (AFD_E) and late (AFD_L) images, as reported in the DICOM header. The decay correction accounted for the ^99mTc physical decay between the early and late image scan acquisition times ($\mathsf{\Delta }t$). This correction was applied by considering the radioactive exponential decay law, using the decay constant λ of ^99mTc ($\lambda =3.209\times {10}^{-5} {\mathrm{s}}^{-1}$), with half-life T_1/2 = 6.01 h.

The WBr was evaluated considering the total counts in WB and subtracting the total counts in the bladder and kidneys, by comparing the decay and scan speed-corrected counts in the late images with early WB counts as follows:

$$W{B}_r={\displaystyle \frac{\left[W{B}_L-\left({Bladder}_L+L\_{Kidney}_L+R\_{Kidney}_L\right)\right]\cdot {\mathrm{e}}^{\lambda \cdot \mathsf{\Delta }t}}{{WB}_E}}\cdot {\displaystyle \frac{{AFD}_E}{{AFD}_L}}$$

(1)

where WB_L, Bladder_L, L_Kidney_L, and R_Kidney_L represent the background-corrected total counts in the WB, bladder, left kidney, and right kidney ROIs, respectively, determined using the late image scans, while WB_E indicates the total counts in the whole body ROI in the early image scan.

The Hr was evaluated by comparing decay and scan speed-corrected counts of the heart in late images with counts in early WB images. The Hr was computed as follows:

$$H_r={\displaystyle \frac{{Heart}_L\cdot {\mathrm{e}}^{\lambda \cdot \mathsf{\Delta }t}}{{WB}_E}}\cdot {\displaystyle \frac{{AFD}_E}{{AFD}_L}}$$

(2)

where Heart_L represents the total counts in the heart ROI in the late image scan, and WB_E indicates the total counts in the WB ROI in the early image scan.

The H/WBr was therefore obtained by dividing the counts in the heart by the counts in late WB images, as follows:

$${H/WB}_r={\displaystyle \frac{{Heart}_L}{\left[{WB}_L-\left({Bladder}_L+L\_{Kidney}_L+R\_{Kidney}_L\right)\right]}}$$

(3)

where Heart_L, WB_L, Bladder_L, L_Kidney_L, and R_Kidney_L represent the total CNTS in the heart, WB, bladder, left kidney, and right kidney in the late image scan, respectively.

2.4. Semiquantitative NM Indexes Calculation with the GM Method

As a second step, on the same subset of 30 patients, the semiquantitative NM indexes Hr, WBr, and H/WBr were calculated using ROIs total counts obtained directly from the geometric mean image calculated in Xeleris Functional Imaging Workstation 4.0: GE Healthcare, Milwaukee, Brookfield, WI, USA (hereafter, Xeleris) from the anterior and posterior views. This approach is here referred to as the GM method. The GM method aims to reduce the time required for ROIs contouring, as the copy-and-mirror operation is not needed. Nevertheless, the ROIs were manually contoured over again as Xeleris did not allow the same ROIs to be copied from the anterior and posterior views to the geometric mean image or between early and late images.

2.5. A Novel Assisted Workflow for the Assisted Contouring on GM Images

To assist the NM physician and reduce the time needed for the computation of semiquantitative indexes, a homemade assisted workflow was developed with the software MIM Maestro v7.3.5(MIM Software Inc., Cleveland, OH, USA; hereafter, MIM). In this workflow, the anterior and posterior planar ^99mTc-DPD images of early and late scans are imported into the software. The GM image is automatically obtained by mirroring the posterior view and applying the GM operation pixel by pixel. A predefined set of ROIs, including heart, left kidney, right kidney, bladder, background, and WB, was automatically loaded, and the NM physician adjusted their positioning and/or shape on the images (Figure 1).

The anterior view, posterior view, and GM image are automatically co-registered so that any change to an ROI is simultaneously applied on all images. This allows a better definition of the ROI by considering the ROI visualization on all images at the same time. When the early image set of ROIs is completed, the same ROIs are automatically transposed on the late images, with simultaneous visualization on the anterior view, posterior view, and GM image. After the automated ROI transfer, the ROI can still be edited for better patient positioning. Finally, total counts on the selected ROI are extracted to .csv file format for the computation of Hr, WBr, and H/WBr. The assisted workflow is described in Figure 2.

Hr, WBr, and H/WBr were obtained with the assisted workflow and GM approach for all the remaining patients in the dataset.

2.6. Assessment of the Agreement Between GM and CM

To validate the GM with respect to the CM, the agreement between the GM and CM for Hr, WBr, and H/BWr was evaluated in terms of Lin’s concordance correlation coefficient (CCC), Pearson’s correlation, considering the CM as the gold standard, and Bland–Altman analysis, in which the differences between the H/WBr calculated with the two methods were plotted against their mean. The 95% confidence interval (CI) was obtained as the mean difference ±1.96 times the standard deviation of the differences σ. The GM and CM agreement was separately assessed when the analyses were performed with the referring workstation (i.e., by manually drawing and positioning the ROIs for each image) and with MIM (where the same ROIs were automatically placed on the various images).

2.7. Comparison Between ATTR Etiology Populations

ATTRwt and ATTRm patient populations were compared in terms of Hr, WBr, and H/WBr. Wilcoxon’s test was used to assess the statistical significance (p-value < 0.05).

2.8. Development of Computational Models for Genomic TTR Prediction

Hr, WBr, and H/WBr data computed with the GM method were inserted in a clinical database, including the following clinical features: age, sex, left ventricular mass (LV-mass), calculated with Devereux formula [16], divided on body-surface area (LV-mass/BSA), and echocardiographic variables (interventricular septum, IVS, left-ventricle ejection fraction, LVEF, left-atrial anterior–posterior diameter, d_LA, and posterior wall, PW).

Data were used as an input of a least absolute shrinkage and selection operator (LASSO) regression model to predict the patient etiology (i.e., ATTRwt versus ATTRm). The LASSO model was implemented with alpha equal to 1, and elastic net (elnet) regularization with the R library glmnet [17]. Variable normalization was performed before the application of the regression model. The entire patient population was split 80–20%. The first dataset was further subdivided into a proportion of 70–30% training–test set, respectively, and the second dataset was used as a validation set. A LASSO with cross-validation (LASSO-CV) was implemented to find the best lambda value that minimizes the mean squared error and then a LASSO was computed to select the variables with non-zero coefficients in correspondence with each previously estimated best lambda found by LASSO-CV. The LASSO regularization method works as a feature selection throughout a progressive shrinkage of variable coefficients. LASSO-CV and the subsequent LASSO were iterated 100 times to verify which variables are selected more frequently and, therefore, are considered more reliable for predicting the outcome of the two different etiologies. This process of 100 iterations permits the assessment of a frequency score by evaluating the number of times that each feature appears in the build models as proof of the robustness of the feature selection operation. The most frequently selected features (recurrence > 30) were used to train a LASSO regression model based on a training dataset (i.e., 70% randomly selected from the 80% of the entire dataset) and tested on the remaining 30% subset. Finally, the model was validated on the validation dataset (20% of the entire database).

Furthermore, the most important and recurrent features selected by the LASSO iterated model were inserted to implement the other two models, Random Forest (RF) and Neural Network (NN).

The RF model was built with 500 decision trees that aim to solve a regression problem, implemented by the R library randomForest [18]. The importance of each variable inside the model was evaluated through two parameters: the percentage of the Increment of Mean Decrease Accuracy (%IncMSE) and the Increment of Node Purity (IncNodePurity). The %IncMSE represents the percentage of the incidence of each variable in evaluating the accuracy of the model, i.e., it specifies quantitatively how much the accuracy of our model decreases if that variable was left out. The IncNodePurity measures the importance of the variable based on the Gini impurity index used for calculating the divisions in the trees. The Gini index, also known as a Gini impurity, calculates the amount of probability of a specific feature being misclassified when randomly selected. If all of the elements are linked with a single class, then it can be called pure. The Gini index varies between values 0 and 1, where 0 expresses the purity of classification, i.e., all of the elements belong to a specified class or only one class exists there; instead, the value 1 indicates the random distribution of elements across various classes.

In addition, NN was developed with the same aim of predicting the patient etiology with the R library NeuralNet [19]. The same entire dataset was divided into training, test, and validation sets, as in the LASSO, with the same proportion. The NN was fed with the eight variables that were most frequently selected by the LASSO regularization method after 100 iterations. The NN was built with one layer and three hidden neurons, and the logistic function was used as an activation function for smoothing. The relative importance of clinical and NM variables was calculated with Garson’s method, which identifies all weighted connections between the nodes of interest, with the R library NeuralNetTools.

For the LASSO regression, RF, and NN models, a receiver operating characteristic (ROC) curve was obtained, and the area under the curve (AUC) was computed to quantify the performance of the models in the training, test, and validation phase. Regarding the validation set, a confusion table for each model was generated.

3. Results

3.1. Patient Cohort

Images of 198 patients were retrieved. In 32 cases, the early WB scan was not available, so it was not possible to compute Hr and WBr, while in 12 cases, other clinical variables were missing. These 44 patients were excluded from the dataset for the development of the LASSO regression model and the NN model, but when Hr and WBr were available, they were included in the comparison of populations with different ATTR etiology. Among the 154 patients included in the development of the computational models, there were 100 cases of ATTRwt and 54 cases of ATTRm with specified TTR mutated genes (Supplementary Material Table S1).

The characteristics of the 154 patients included in the dataset for LASSO regression, RF, and NN are reported in

Table 1. Main characteristics of patients included in the development of the LASSO regression model and NN model. Reported NM scintigraphy parameters were obtained with the GM method. Abbreviations: IVS = interventricular septum, LVEF = left-ventricle ejection fraction, d_LA = left-atrial anterior–posterior diameter, PW = posterior wall, ATTRwt = ATTR wild type, ATTRm = ATTR mutated. The bold was used to highlight sections among variables.

Variable	Occurrences (Percentage) or Median [Range]
Patient characteristics
Age (years)	76 [25–90]
Sex [Male/Female]	135 (88%)/19 (12%)
BSA (m2)	1.8 [1.4–2.4]
Echocardiographic Variables
d_LA (mm)	48 [4–67]
LVEF (%)	57 [24–81]
IVS (mm)	18 [12–29]
PW (mm)	16 [9–27]
LV-mass (kg)	345 [172–690]
LV-mass/BSA (kg/m2)	186 [91–381]
NM scintigraphy parameters
WBr (%)	82.4 [31.5–99.5]
Hr (%)	4.3 [1.6–9.7]
H/WBr (%)	5.4 [2.0–10.7]
Visual score	0: 2 (1%) 1: 3 (2%) 2: 120 (61%) 3: 73 (37%)
Time between early and late acquisitions (h)	2.8 [1.8–4.7]
Etiology
ATTRwt	100 (65%)
ATTRm	54 (35%)

. The reported Hr, WBr, and H/WBr are the ones obtained with the GM method and the homemade assisted workflow.

The patient cohort was composed of 154 patients that had a median age of 76 years and contained a total of 135/154 (88%) males with a median BSA of 1.8 m². Echocardiographic variables included d_LA, LVEF, IVS, and PW having median values of 48 mm, 57%, 18 mm, and 16 mm, respectively.

3.2. Agreement between GM and CM

The Hr, WBr, and H/WBr obtained using the GM method showed a high correlation with ones derived using the CM. When evaluated in Xeleris, Lin’s CCC was 0.93 [0.87–0.96], 0.94 [0.88–0.97], and 0.99 [0.98–0.99] for H/WBr, Hr, and WBr, respectively, while Pearson’s linear correlation coefficient was always higher than 0.87, with R² > 0.94. For H/WBr, the Bland–Altman analysis resulted in a bias of 0.46%, and the CI ranged between −0.35% and 1.27%. Similarly, Hr showed a bias of 0.33% and a CI between −0.31% and 0.97%, and the WBr showed a bias of −0.91% and a CI between −2.20% and 0.38%. When the comparison between GM and CM was performed within MIM (that is, using exactly the same ROIs set on anterior and posterior views and on the geometric mean image), the H/WBr bias decreased to 0.12% [CI: 0.04%; 0.19%], the Hr bias decreased to 0.07% [CI: 0.01%; 0.13%], and the WBr bias decreased to −0.50% [CI: −1.22%; 0.22%]. Therefore, the assisted workflow developed in MIM guarantees a better agreement between the GM and CM and justifies the use of the GM in the following evaluation on the larger dataset of patients. The Bland–Altman plot of H/WBr with the GM versus the CM calculated in the MIM-assisted workflow is reported in Figure 3, while all data are summarized in

Table 2. Bland–Altman analysis of Hr, WBr, and H/WBr calculated with the GM vs. the CM. The agreement is expressed in terms of bias and 95%CI.

	Referring Workstation		Assisted Workflow
	Bias	95% CI	Bias	95% CI
Hr	0.33%	−0.31%; 0.97%	0.07%	0.01%; 0.13%
WBr	−0.91%	−2.20%; 0.38%	−0.50%	−1.22%; 0.22%
H/WBr	0.46%	−0.35%; 1.27%	0.12%	0.04%; 0.19%

.

3.3. Comparison between NM Semiquantitative Indexes in Different ATTR Etiology Populations

The boxplots of H/WBr, Hr, and WBr versus the etiology (i.e., ATTRwt and ATTRm) are reported in Figure 4. ATTRwt showed significantly higher Hr and H/WBr (p < 0.001) than ATTRm, while a trend was observed for WBr (p = 0.11).

3.4. Evaluation of the Performances of the Computational Models

The LASSO regularization, iterated 100 times, identified the following as the most frequently selected features: age_FE, d_LA, Hr, LV-mass/BSA, PW, and WBr. H/WBr was excluded by the LASSO regularization process as it was highly correlated (R² > 0.9) with Hr. The best lambda value identified in the cross-validation step, after the 100 iterations, was 6.34 × 10⁻² (Figure 5).

The variables that were characterized by a non-zero coefficient after the LASSO regularization at the best lambda value were the patient age at first examination and the Hr. The model developed a high AUC of 88.0% [CI 80.6–95.5%], 89.6% [CI 77.0–100%], and 87.3% [CI 74.3–100%] for the training, test, and validation sets, respectively. The best cut-off obtained a specificity of 90% and a sensitivity of 84% on the test dataset and 66.7% and 95.7% on the validation dataset.

Similarly, a specificity of 88% and a sensitivity of 60% on the test dataset and 78.3% and 33.3% on the validation dataset were found for the RF approach. The AUC of the RF approach was significantly lower compared to LASSO, ascribed to the triangular shape of the ROC curve, which was obtained from a single point in addition to the 0–100% extremities.

The NN model (reported in Supplementary Material Figure S1) had good performance with high AUC values of 92.3% [CI 84.7–99.9%], 84.4% [CI 70.4–98.4%], and 78.3% [CI 61.1–95.4%] for the ROC curve on the training, test, and validation sets. All ROC curves are reported in Figure 6.

The %IncMSE and the IncNodePurity for each variable inserted in the RF classifier are shown in Figure 7. The variables age_FE and Hr had a greater impact with values equal to 25.4 and 14.7, respectively, and therefore contribute to making the model more accurate. The IncNodePurity also showed higher values for the clinical variable, age_FE, and the MN variable, Hr, with values equal to 6.45 and 4.45, respectively.

Table 3. Relative importance of variables expressed by non-zero coefficients for the LASSO, by %IncMSE and IncNodePurity for the RF model, and by Garson’s algorithm for the NN model. The bold was used to highlight sections among variables.

Model	Variable Importance
LASSO	Intercept		Non-zero coefficients
	1.619		age_FE	0.203
			Hr	0.082
RF	%IncMSE		IncNodePurity
	age_FE	25.365	age_FE	6.453
	d_LA	3.028	d_LA	1.729
	Hr	14.711	Hr	4.447
	LV-mass/BSA	0.129	LV-mass/BSA	2.138
	PW	−0.015	PW	1.514
	WBr	−2.058	WBr	1.592
NN	Relative Importance by Garson’s method
	age_FE	0.288
	d_LA	0.164
	Hr	0.242
	LV-mass/BSA	0.110
	PW	0.126
	WBr	0.0698

reported the values of the intercept and non-zero coefficients associated with the variables in correspondence with the best lambda for the LASSO model, while the importance of each variable within the RF model was indicated by the parameters %IncMSE and IncNodePurity, and the relative importance of clinical and NM variables in the Neural Network model was estimated by Garson’s algorithm.

The confusion matrices, shown in Figure 8, describe the performance of the three models on the external validation dataset, highlighting the true positives and negatives (green boxes) from the misclassifications (red boxes).

Since the number of correct classifications is largely higher than the incorrect ones, the three predictive models can be considered satisfactory for distinguishing the etiology of a new patient who does not belong to the training and test database.

4. Discussion

Nowadays, although ATTRwt is increasingly diagnosed [20], the gold standard for the diagnosis of amyloid cardiomyopathy type still requires an EMB, which remains an invasive and potentially complicated procedure. Nevertheless, Gillmore et al. [11] have reported in a large multicentric cohort of patients the potential of bone scintigraphy for the diagnosis of cardiac ATTR, showing >99% sensitivity and 86% specificity considering a visual score ≥2 according to Perugini et al. [13], with specificity reaching 100% when excluding the AL form [11]. In addition, as shown by Bokhari et al. [21], the use of NM imaging semiquantitative indexes can enable a better differentiation between cardiac amyloidosis subtypes. They reported a significantly higher semiquantitative cardiac uptake in ATTR patients than in the AL cohort using ^99mTc-PYP as a bone-seeking radiotracer. Lately, Gallini et al. [22] used H/WBr and other metrics in ^99mTc-HMDP bone scintigraphy to discriminate ATTR and AL patients in a heterogeneous cohort of 76 patients.

Nevertheless, the use of semiquantitative NM indexes might require a time-consuming procedure, depending on the available referring workstation tools, as the drawing of ROIs is needed on both anterior and posterior views to compute the GM [9].

The GM method could ease the calculation of semiquantitative NM indexes as it requires fewer operations by the NM physician, while its correlation with previously published data that adopted the CM is still strong [23]. In this work, the GM was validated in a cohort of ATTR patients and implemented in a novel assisted workflow developed in MIM that further speeds up and facilitates the process and makes it more reproducible among different users. The validation of the GM with respect to the CM was performed in terms of Bland–Altman analysis, which basically reports the systematic over- or under-estimation (in terms of positive or negative bias) and the possible deviation (in terms of 95%CI) when the same quantity is estimated with two alternative methods. Although GM and CM were in good agreement when Hr, WBr, and H/WBr were evaluated in the referring workstation, both the bias and the 95%CI were further reduced with the assisted workflow. This is mainly because each ROI was automatically mirrored and positioned from the anterior image to the posterior and geometric mean images, thus reducing the uncertainty due to ROI positioning or recontouring. The residual bias showed that GM systematically overestimated Hr and H/WBr and, conversely, underestimated WBr, although this was negligible in terms of clinical content. This can be ascribed to the properties of the geometric mean operation, which assigns zero value to a pixel if its value is null in either the anterior or posterior image, which is more frequent in the low-uptake areas (e.g., soft tissues) than in the heart. The same characteristic is likely to explain why the Bland–Altman Plot shows that the absolute difference between the two methods increases as the average value increases. Although the reported difference was deemed not clinically relevant (the highest reported difference was 0.2% over a mean of 9.5%, that means an H/WBr of 9.6% with the GM versus 9.4% with the CM), the fact that the discrepancy between the GM and the CM was more relevant when the radiopharmaceutical accumulation in the heart was higher should be considered.

The Hr, WBr, and H/WBr data obtained on a large cohort of 154 patients were used to compare ATTRwt and ATTRm populations. ATTRwt showed statistically significantly higher Hr and H/WBr, while a trend of lower WBr in ATTRwt patients was observed. This data is in disagreement with Galat et al. [23], although this might be explained by a slightly different definition of Hr and H/WBr and by the lower number of analyzed patients (i.e., 55 among ATTRwt and ATTRm) [23].

Of note, the fact that no statistically significant difference was observed in terms of WBr suggests that the higher cardiac uptake in ATTRwt patients (proved by a large and statistically significant difference in Hr between the two populations) is balanced by a higher diffuse uptake in soft tissues and bones in ATTRm patients. Furthermore, the fact that the semiquantitative evaluation between the two populations (ATTRwt vs. ATTRm) is different could be explained not only by the etiology but also by the stage of the disease, since ATTRm patients are often identified in an early stage of the disease because they reach specialist cardiologist attention through family screening, in the absence of symptoms or laboratory/radiological findings suspicious for cardiac amyloidosis. Of note, a limitation of these results is that the calculation of the semi-quantitative cardiac indexes was processed on two numerically different ATTR populations, i.e., a lower cohort of ATTRm patients than ATTRwt patients (n = 54 vs. n = 100, respectively).

Even in our population, nevertheless, the NM semiquantitative indexes are not sufficient alone to discriminate ATTRwt and ATTRm patients with acceptable accuracy; for this reason, different computational models have been developed simultaneously considering both clinical and NM-derived variables.

In the last year, machine-learning approaches have been explored for the detection and classification of cardiac amyloidosis or patient risk stratification [24]; nevertheless, this is, to the best of our knowledge, the first case in which semiquantitative NM indexes derived from bone scintigraphy were used to develop a computational model which is able to discriminate ATTRwt and ATTRm patients, with the exception of the work of Bonnefous et al. that used the cardiac uptake in bone scintigraphy as a binary variable among various clinical variables in an unsupervised machine learning approach to cluster patients with suspected cardiac amyloidosis [25].

In detail, in the LASSO regression model, and similarly (in terms of specificity and sensitivity) in the RF and NN models as well, the most important variables identified were patient age_FE and Hr, which resulted in a higher discrimination ability than other clinical or echocardiography-derived variables. Overall, these data highlight the importance of bone scintigraphy in the diagnostic process.

In any case, the development of machine learning-based computational models requires wide patient populations, for which the calculation of the semiquantitative NM indexes can be eased by the GM, which has been shown to be substantially equivalent to the CM, and by the development and use of assisted or semi-automated tools, as proposed in this work.

Etiology is an important determinant of ^99mTc-DPD scintigraphic variability in cardiac amyloidosis. Nevertheless, a further step would be the use of semiquantitative NM indexes to aid the classification of patients among ATTRm subtypes. Our analysis indicates that ^99mTc-DPD scintigraphy could offer a useful first step in the workup of the differential diagnosis of ATTRm versus ATTRwt etiology in patients with documented cardiac amyloidosis. Unfortunately, this goal was not feasible with the available patient population as subtypes were under-represented. In detail, the largest ATTRm subgroup, Ile68Leu, was reported in 27 out of 54 ATTRm patients, while all other ATTRm subgroups were represented by less than eight patients each. In this sense, only the development of multicentric datasets would enable this goal.

According to our findings, ^99mTc-DPD could offer a preliminary assessment of different etiologies in order to find out in advance if the patient with certain clinical variables and NM parameters could belong to the ATTRm or ATTRwt class. When the patient etiology with TTR genomic mutation is confirmed by the predictive model, this could represent an incentive for the patient to undergo a more invasive examination or a blood test to find out the type of mutation.

In view of the enhancement of semi-quantitative analysis, it is feasible to implement the examination with a heart-centered SPECT/CT acquisition. This methodology could assist in the delineation of cardiac regions affected by the disease, potentially correlating with clinical outcomes. A higher uptake in a specific subcardiac area might show a potential prognostic value to distinguish the etiology, ATTRm or ATTRwt, building a more robust and accurate model, also based on myocardial infiltration. In this context, 3D images might have the potential to identify NM indices capable of predicting, beyond etiology, the low or high risk of heart failure or overall survival; this will be investigated in further studies.

Moreover, the non-invasive and easy calculation of semiquantitative NM indexes derived from the development of the assisted MIM workflow will enable the same methodology to be extended to other outcomes, such as myocardial amyloid infiltration or the prediction of major adverse cardiac events.

5. Conclusions

The present study aims to validate the GM with respect to the CM with the advantage of an assisted workflow that uses an ROI template that allows the transfer of the same ROIs from the early to the late planar scintigraphy, thus avoiding the inter-variability between operators in drawing the ROIs on the geometric mean images. The diagnostic value of planar scintigraphy is underlined by the calculation of the three NM parameters, H/WBr, Hr, and WBr, which are indices of the quantification of cardiac uptake (the accumulation of amyloid) and of the WB uptake. Our results indicate that machine learning approaches can be used as an effective and non-invasive prognostic tool for distinguishing two different types of etiology, ATTRm, and ATTRwt. The pharmacopoeia currently has approved several target therapies for the different etiologies (i.e., ATTRm vs. ATTRwt), thus the possibility of differential diagnosis with non-invasive and inexpensive methods becomes crucial for personalized treatments. This approach represents a standardized support decision system for an early identification of high-risk patients who require further examination to refer them to different target therapies.