Bayesian Convolutional Neural Networks in Medical Imaging Classification: A Promising Solution for Deep Learning Limits in Data Scarcity Scenarios.

Deep neural networks (DNNs) have already impacted the field of medicine in data analysis, classification, and image processing. Unfortunately, their performance is drastically reduced when datasets are scarce in nature (e.g., rare diseases or early-research data). In such scenarios, DNNs display poor capacity for generalization and often lead to highly biased estimates and silent failures. Moreover, deterministic systems cannot provide epistemic uncertainty, a key component to asserting the model's reliability. In this work, we developed a probabilistic system for classification as a framework for addressing the aforementioned criticalities. Specifically, we implemented a Bayesian convolutional neural network (BCNN) for the classification of cardiac amyloidosis (CA) subtypes. We prepared four different CNNs: base-deterministic, dropout-deterministic, dropout-Bayesian, and Bayesian. We then trained them on a dataset of 1107 PET images from 47 CA and control patients (data scarcity scenario). The Bayesian model achieved performances (78.28 (1.99) % test accuracy) comparable to the base-deterministic, dropout-deterministic, and dropout-Bayesian ones, while showing strongly increased "Out of Distribution" input detection (validation-test accuracy mismatch reduction). Additionally, both the dropout-Bayesian and the Bayesian models enriched the classification through confidence estimates, while reducing the criticalities of the dropout-deterministic and base-deterministic approaches. This in turn increased the model's reliability, also providing much needed insights into the network's estimates. The obtained results suggest that a Bayesian CNN can be a promising solution for addressing the challenges posed by data scarcity in medical imaging classification tasks.

Deep-learning-based cardiac amyloidosis classification from early acquired pet images.

The objective of the present work was to evaluate the potential of deep learning tools for characterizing the presence of cardiac amyloidosis from early acquired PET images, i.e. 15 min after [18F]-Florbetaben tracer injection. 47 subjects were included in the study: 13 patients with transthyretin-related amyloidosis cardiac amyloidosis (ATTR-CA), 15 patients with immunoglobulin light-chain amyloidosis (AL-CA), and 19 control-patients (CTRL). [18F]-Florbetaben PET/CT images were acquired in list mode and data was sorted into a sinogram, covering a time interval of 5 min starting 15 min after the injection. The resulting sinogram was reconstructed using OSEM iterative algorithm. A deep convolutional neural network (CAclassNet) was designed and implemented, consisting of five 2D convolutional layers, three fully connected layers and a final classifier returning AL, ATTR and CTRL scores. A total of 1107 2D images (375 from AL-subtype patients, 312 from ATTR-subtype, and 420 from Controls) have been considered in the study and used to train, validate and test the proposed network. CAclassNet cross-validation resulted with train error mean ± sd of 2.001% ± 0.96%, validation error of 4.5% ± 2.26%, and net accuracy of 95.49% ± 2.26%. Network test error resulted in a mean ± sd values of 10.73% ± 0.76%. Sensitivity, specificity, and accuracy evaluated on the test dataset were respectively for AL-CA sub-type: 1, 0.912, 0.936; for ATTR-CA: 0.935, 0.897, 0.972; for control subjects: 0.809, 0.971, 0.909. In conclusion, the proposed CAclassNet model seems very promising as an aid for the clinician in the diagnosis of CA from cardiac [18F]-Florbetaben PET images acquired a few minutes after the injection.

Appropriate choice of stress modality in patients undergoing myocardial perfusion scintigraphy with a cardiac camera equipped with solid-state detectors: the role of diabetes mellitus.

Aims: To evaluate the impact of diabetes mellitus (DM) on the accuracy of myocardial perfusion scintigraphy (MPS) in detecting coronary artery disease (CAD). Methods and results: Two hundred and sixteen patients with DM and 432 matched controls were submitted to MPS on a dedicated cardiac camera equipped with cadmium-zinc-telluride (CZT) detectors and coronary angiography. Exercise stress was performed in 442 (68%) patients, while the remainders underwent vasodilator stress. Exercise level was determined as the percentage of the predicted maximal workload that was attained (%Wattmax). The summed difference score was derived from CZT images. A coronary stenosis >70% was considered obstructive. The prevalence of obstructive CAD was 59.7% in patients with DM and 56.2% in controls (P = NS). The accuracy of MPS in detecting CAD was similar in patients with and without DM [area under the ROC curve (AUC) 0.77 vs. 0.78, P = NS]. An interaction between the accuracy of MPS and cardiac stress-protocol was revealed. In fact, in patients with DM exercise stress CZT had a lower accuracy than vasodilator-stress (AUC 0.70 vs. 0.89, P < 0.001), because of a lower specificity (45% vs. 69%), while in the control group the accuracy of MPS was similar regardless of the stress-protocol adopted. Patients with DM attained a significantly lower %Wattmax during exercise than controls (76 ± 27% vs. 82 ± 26%, P = 0.038), which resulted an independent predictor of reduced specificity (P = 0.026). Conclusion: The accuracy of CZT imaging in patients with DM is elevated, and it is quite comparable to the one obtained in patients without DM. However, a reduced specificity can be expected in the case of exercise stress CZT, because of an impaired exercise capacity.

Determinants of left ventricular mechanical dyssynchrony in patients submitted to myocardial perfusion imaging: A cardiac CZT study.

BACKGROUND: An interaction between coronary anatomy, myocardial perfusion, and left ventricular (LV) functional parameters in the development of mechanical LV dyssynchrony (LVD) has been suggested. This study examined the correlates of LVD in a large sample size of patients with known or suspected coronary artery disease (CAD) using cadmium-zinc-telluride camera. METHODS: Six-hundred and fifty-seven consecutive patients who underwent myocardial perfusion imaging (MPI) and coronary angiography were included. Coronary stenosis >70% was considered significant. LV perfusion and functional parameters were computed from MPI images. The presence of significant LVD was evaluated by phase standard deviation and histogram bandwidth. RESULTS: 415/657 (63%) patients had significant CAD. LVD was present in 247 (38%) patients and was associated with the presence of a higher CAD burden (P < .001), more impaired measures of LV perfusion (P < .001), contractile function (P < .001), and larger LV volumes (P < .001). By multivariate analysis, the LV end-systolic volume index (P < .001) and ischemic burden (P < .001) were the strongest predictors of LVD independent of CAD extent and LV systolic dysfunction. CONCLUSIONS: LVD is frequent in patients undergoing MPI for suspected or known CAD. Its presence is independent of CAD burden and LV systolic dysfunction, but is dependent on the presence of myocardial perfusion abnormalities and LV end-systolic volume.

Evaluation of left ventricular diastolic function with a dedicated cadmium-zinc-telluride cardiac camera: comparison with Doppler echocardiography.

AIMS: To evaluate the relationships between measures of left ventricular (LV) filling dynamics at cadmium-zinc-telluride (CZT) imaging and indexes of LV diastolic function at transthoracic echocardiography. METHODS AND RESULTS: Two hundred and forty-seven patients underwent myocardial perfusion imaging at rest and after stress with a low-dose CZT protocol and a baseline transthoracic echocardiography study. All patients were submitted to invasive or computed coronary angiography. The peak filling rate (PFR) and the time to PFR (TPFR) were derived from gated CZT images as measures of LV filling dynamics. LV diastolic function was also evaluated at echocardiography and the presence of significantly increased LV filling pressures determined. Increased LV filling pressures at transthoracic echocardiography were evident in 103 (42%) patients. Interestingly, independently from the presence of coronary artery disease, there was a strict correlation between the presence and severity of LV diastolic dysfunction at echocardiography and CZT-derived measures of filling dynamics, i.e., PFR (P = 0.001) and TPFR (P = 0.001). At receiving operating characteristic analysis, a composite index of reduced PFR (≤2.11 end-diastolic volume s(-1)) and increased TPFR (>234 ms) showed a sensitivity of 84% and a specificity of 67% in unmasking the presence of elevated LV filling pressures at echocardiography. CONCLUSIONS: CZT-derived measures of LV filling dynamics correlate with echocardiographic parameters of diastolic function and may identify the presence of increased LV filling pressures.

Single region of interest versus multislice T2* MRI approach for the quantification of hepatic iron overload.

PURPOSE: To evaluate the effectiveness of the single ROI approach for the detection of hepatic iron burden in thalassemia major (TM) patients in respect to a whole liver measurement. MATERIALS AND METHODS: Five transverse hepatic slices were acquired by a T2* gradient-echo sequence in 101 TM patients and 20 healthy subjects. The T2* value was calculated in a single region of interest (ROI) defined in the medium-hepatic slice. Moreover, the T2* value was extracted on each of the eight ROIs defined in the functionally independent segments. The mean hepatic T2* value was calculated. RESULTS: For patients, the mean T2* values over segments VII and VIII were significantly lower. This pattern was substantially preserved in the two groups identified considering the T2* normal cutoff. All segmental T2* values were correlated with the single ROI T2* value. After the application of a correction map based on T2* fluctuations in the healthy subjects, no significant differences were found in the segmental T2* values. CONCLUSION: Hepatic T2* variations are low and due to artifacts and measurement variability. The single ROI approach can be adopted in the clinical arena, taking care to avoid the susceptibility artifacts, occurring mainly in segments VII and VIII.

Influence of myocardial fibrosis and blood oxygenation on heart T2* values in thalassemia patients.

PURPOSE: To determine whether T2* measurements quantifying myocardial iron overload in thalassemia patients are influenced by myocardial fibrosis and blood oxygenation. MATERIALS AND METHODS: Multislice multiecho T2* was performed in 94 thalassemia patients in order to quantify myocardial iron overload. The left ventricle was automatically segmented into a 16-segment standardized heart model, and the T2* value on each segment as well as the global T2* were calculated. Delayed enhanced cardiovascular magnetic resonance (DE-CMR) images were obtained to detect myocardial fibrosis. The blood oxygenation was assessed by the noninvasive measurement of partial pressure of oxygen (pO2). RESULTS: Myocardial fibrosis was detected in 31 patients (33%). The global T2* value in patients with fibrosis was comparable with that of patients without fibrosis (P = 0.88) and T2* values in segments with fibrosis were comparable with those in segments without fibrosis (P = 0.83). The global T2* value was not correlated with the pO2 (Spearman's coefficient of correlation = 0.99). CONCLUSION: Myocardial fibrosis and blood oxygenation did not significantly affect the T2* values. These data further support the use of heart T2* as equivalent of heart iron in the clinical arena.

Multislice multiecho T2* cardiac magnetic resonance for the detection of heterogeneous myocardial iron distribution in thalassaemia patients.

The present study investigated myocardial T2* heterogeneity in thalassaemia major (TM) patients by cardiac magnetic resonance (CMR), to determine whether is related to inhomogeneous iron overload distribution. A total of 230 TM patients consecutively referred to our laboratory were studied retrospectively. Three short-axis views (basal, medium and apical) of the left ventricle (LV) were obtained by multislice multiecho T2* CMR. T2* segmental distribution was mapped on a 16-segment LV model. The level of heterogeneity of the T2* segmental distribution, evaluated by the coefficient of variation (CoV), was compared with that of a surrogate data set, to determine whether the inhomogeneous segmental distribution of T2* could be generated by susceptibility artefacts. Susceptibility artefacts offer an explanation for the T2* heterogeneity observed in patients without iron overload. In subjects with global T2* below the lower limit of the normal, T2* heterogeneity increased abruptly which could not be explained by artefactual effects. Some segmental T2* values were below and others above the limit of normal threshold (20 ms) in 104 (45%) TM patients. Among these patients, 74% showed a normal T2* global value. In conclusion, a true heterogeneity in the iron overload distribution may be present in TM patients. Heterogeneity seemingly appears in the borderline myocardial iron and stabilizes at moderate to severe iron burden.

Improved T2* assessment in liver iron overload by magnetic resonance imaging.

In the clinical MRI practice, it is common to assess liver iron overload by T2* multi-echo gradient-echo images. However, there is no full consensus about the best image analysis approach for the T2* measurements. The currently used methods involve manual drawing of a region of interest (ROI) within MR images of the liver. Evaluation of a representative liver T2* value is done by fitting an appropriate model to the signal decay within the ROIs vs. the echo time. The resulting T2* value may depend on both ROI placement and choice of the signal decay model. The aim of this study was to understand how the choice of the analysis methodology may affect the accuracy of T2* measurements. A software model of the iron overloaded liver was inferred from MR images acquired from 40 thalassemia major patients. Different image analysis methods were compared exploiting the developed software model. Moreover, a method for global semiautomatic T2* measurement involving the whole liver was developed. The global method included automatic segmentation of parenchyma by an adaptive fuzzy-clustering algorithm able to compensate for signal inhomogeneities. Global liver T2* value was evaluated using a pixel-wise technique and an optimized signal decay model. The global approach was compared with the ROI-based approach used in the clinical practice. For the ROI-based approach, the intra-observer and inter-observer coefficients of variation (CoVs) were 3.7% and 5.6%, respectively. For the global analysis, the CoVs for intra-observers and inter-observers reproducibility were 0.85% and 2.87%, respectively. The variability shown by the ROI-based approach was acceptable for use in the clinical practice; however, the developed global method increased the accuracy in T2* assessment and significantly reduced the operator dependence and sampling errors. This global approach could be useful in the clinical arena for patients with borderline liver iron overload and/or requiring follow-up studies.

Right heart overload contributes to cardiac natriuretic hormone elevation in patients with heart failure.

BACKGROUND: Atrial and brain natriuretic peptides (ANP and BNP) plasma concentration increases and holds a prognostic significance in patients with left ventricular dysfunction. We assessed the hypothesis that right ventricular (RV) overload might significantly contribute to plasma elevation of cardiac natriuretic hormones in patients with heart failure. METHODS: Forty-one patients with cardiomyopathy and depressed left ventricular (LV) function (ejection fraction, EF, <40%), underwent cardiac magnetic resonance imaging (MRI) and resting plasma determination of ANP and BNP. Nineteen healthy subjects were also studied as control group. Ventricular volumes and function were assessed by MRI. RESULTS: In the group of patients, LVEF was 22.6+/-1.2% (controls: 61.2+/-1.3%, P<0.001, mean+/-S.E.M.), while RVEF was 48.2+/-2.5% (controls: 66.7+/-1.6%, P<0.001); LV and RV end diastolic/systolic volumes, corrected by body surface area, were 143+/-7/114+/-7 ml/m2 (controls 70+/-3/27+/-2 ml/m2, both P<0.001) and 66+/-3/37+/-4 ml/m2 (controls: 63+/-4/21+/-2 ml/m2, P<0.01 only for end-systolic volume). BNP plasma value was on average 324+/-39 pg/ml (range: 23-1280, controls 10+/-2 pg/ml), ANP value was 144+/-17 pg/ml (range: 26-534, controls 15+/-1 pg/ml). BNP positively correlated with either end-diastolic or end-systolic RV volume in patients, less with LV systolic, and not with LV diastolic volume. Moreover, a significant negative correlation was observed between BNP and either LVEF or RVEF. Conversely, ANP showed a significant correlation only with end-systolic RV volume and with both RVEF and LVEF. When multivariate stepwise linear regression analysis was applied LVEF resulted the only independent predictor for ANP plasma values (R=0.591, P<0.001), while LVEF and RV end-diastolic volume for BNP (R=0.881, P<0.001, and R=0.881, P=0.035, respectively). CONCLUSIONS: Right heart overload contributes independently to plasma elevation of natriuretic peptides. RV involvement, which is known to independently worsen prognosis in patients with cardiomyopathy, might contribute to their established prognostic power, inducing compensatory secretion of plasma cardiac natriuretic hormones.

Baseline/postnitrate tetrofosmin SPECT for myocardial viability assessment in patients with postischemic severe left ventricular dysfunction: new evidence from MRI.

UNLABELLED: The aim of this study was to compare (99m)Tc-tetrofosmin SPECT with contrast-enhanced MRI (ceMRI), a new direct sign of myocardial fibrosis. METHODS: We studied 21 patients (age, 60 +/- 11 y; 19 male) with previous myocardial infarction and severe left ventricular dysfunction (ceMRI EF, 29% +/- 6%). All patients underwent resting and postnitrate (intravenous isosorbide dinitrate) (99m)Tc-tetrofosmin gated SPECT as well as ceMRI. Scintigraphic analysis was performed using quantitative perfusion SPECT (QPS), providing the percentage radiotracer uptake and defect severity in a 20-segment model. Hyperenhancement was defined by the increase of ceMRI signal intensity 20 min after intravenous injection of gadolinium-diethylenetriaminepentaacetic acid and its regional extension as the percentage of the total segment area. RESULTS: In ceMRI dysfunctional segments, the correlation between the extent of hyperenhancement and both (99m)Tc-tetrofosmin uptake and defect severity was significantly better after nitrate administration rather than at rest (P < 0.0001). Using a ceMRI cutoff below 40%, 102 of 196 (52%) segments were viable, whereas 94 (48%) segments were not viable. According to receiver-operating characteristic curve analysis, diagnostic accuracies were higher for postnitrate (99m)Tc-tetrofosmin uptake as well as perfusion defect severity than using resting data (0.84 vs. 0.71, P < 0.001; 0.89 vs. 0.76, P < 0.001, respectively). CONCLUSION: After nitrate administration, (99m)Tc-tetrofosmin uptake and perfusion defect severity were closely related to ceMRI, demonstrating, in vivo, the existence of an inverse correlation between the transmural distribution of fibrosis and tracer delivery to the myocardium.

Contrast media in cardiovascular magnetic resonance.

Among the available imaging techniques, Magnetic Resonance Imaging (MRI) is gaining an increasing role in the cardiologic setting because its specific properties such as the use of non ionising energies, the natural strong contrast between different tissues, the absence of spatial limitations, the good spatial and temporal resolution, the reduced operator dependency. To further improve the images quality and the histopathologic characterisation of tissues the use of contrast media (molecules containing gadolinium, manganese, iron, dysprosium ions) has been proposed both in the experimental and in the clinical settings. Among these ions gadolinium, which having 7 odd electrons in the external orbit has a strong magnetic momentum, is the most used. Gadolinium by itself is extremely toxic but once it is linked with a chelanting agent such as DTPA (Dietilen-Triamin-Penta-Acetic acid) the resulting complex shows a very low toxicity. The number of Gadolinium based compound is growing together with the use of contrast agents in MRI. These contrast agents are routinely used to perform Magnetic Resonance Angiography (MRA) and to a better definition of several cardiac diseases such as the presence of a intra- or paracardiac mass, the evaluation of myocardial perfusion and the evaluation of viability. Both the latter applications have relevant clinical implications. In fact the assessment of myocardial perfusion is one of the most used approach for detecting inducible myocardial ischemia due to major coronary artery disease or to assess the presence of a microvascular disease. The presence and the extent of viable myocardium is deeply modifying the clinical decision making as this viable tissue can recruit a normal function spontaneously or after revascularisation. Furthermore, the extent of viable myocardium has a strong correlation with negative prognosis. Clinical events are also time related to the detection of viable tissue. These evidences imply that the diagnostic procedure needs the highest level of accuracy. Either in the case of myocardial perfusion and in that of myocardial viability the advantages of MRI with respect to the others techniques are the use of non ionising radiations, the superior spatial resolution, an overall cost/benefit favourable ratio which explains the growing interest among cardiologists toward this new diagnostic tool.

A fast and effective method to assess myocardial necrosis by means of contrast magnetic resonance imaging.

PURPOSE: Contrast magnetic resonance (CMR) can identify myocardial necrosis after gadolinium administration as a hyperenhanced (HE) area. Yet there are no software tools that can effectively quantify such an area. The aim of this study is to develop a robust and effective algorithmic method for defining the extent of myocardial necrosis evidenced through CMR. METHOD: Fifteen patients with previous myocardial infarction underwent nitrate Tetrofosmin G-SPECT and CMR. A software tool was developed, allowing semiautomatic detection of endocardial and epicardial borders and the automatic detection of HE regions. The accuracy of the proposed quantitative method of analysis has been tested with G-SPECT analysis that it is less than an ideal method for assessing myocardial viability, but at present is accepted and widely used in the clinical arena. RESULTS: Segmental (SEHE) and global extension of HE were evaluated. HE was present in 161 of the 255 analyzed segments. Of the 161 HE segments, the mean SEHE was 36 +/- 30%. The operator independence (intraobserver: r = 0.97, p < 0.0001, interobserver: r = 0.95, p < 0.0001) was good and significant, with noticeable time savings with respect to manual analysis. There was strong and inverse correlation between SEHE and scintigraphic regional uptake reduction (r = -0.66, p < 0.0001), and also a positive correlation between SEHE and SPECT defect extension (r = 0.75, p < 0.0001). When assessing the global extent of necrosis, the correlation between the two techniques was strong (r = 0.79, p = 0.0004). CONCLUSIONS: The proposed method of quantifying myocardial necrosis by CMR is highly reliable, reproducible, and operator-independent for quantifying.

Does subclinical hypothyroidism affect cardiac pump performance? Evidence from a magnetic resonance imaging study.

OBJECTIVES: We sought to assess the effects of subclinical hypothyroidism (SHT) on the cardiac volumes and function. BACKGROUND: The cardiovascular system is one of the principal targets of thyroid hormones. Subclinical hypothyroidism is a common disorder that may represent "early" thyroid failure. METHODS: Thyroid profile was evaluated in 30 females with SHT and 20 matched control subjects. Left ventricular end-diastolic volume (EDV) and end-systolic volume (ESV), stroke volume (SV), cardiac index (CI), and systemic vascular resistance (SVR) were calculated by cardiac magnetic resonance (CMR). Regional greatest systolic lengthening (E1) and greatest systolic shortening (E2) were calculated by tagging CMR. RESULTS: EDV was lower in SHT than in controls (64.3 +/- 8.7 ml/m(2) vs. 81.4 +/- 11.3 ml/m(2), p < 0.001), as well as SV [corrected] (38.9 +/- 7.5 ml/m(2) vs. 52.5 +/- 6.1 ml/m(2), p < 0.001) and CI (2.6 +/- 0.5 l/[min.m(2)] vs. 3.7 +/- 0.4 l/[min.m(2)], p < 0.001). SVR [corrected] was higher in SHT (12.5 +/- 2.5 mm Hg.min/[l.m(2)] vs. 8.6 +/- 1.1 mm Hg.min/[l.m(2)], p = 0.003). The E1 was higher in controls than in SHT at the basal (p = 0.007), equatorial (p = 0.05), and apical (p = 0.008) levels, as well as E2 at the equatorial (p = 0.001) and apical (p = 0.001) levels. All parameters normalized after replacement therapy. A negative correlation between TSH and EDV (p < 0.001), SV (p < 0.001), CI (p < 0.001), and E1 at the apical level (p < 0.001) and a positive correlation between TSH and SVR (p < 0.001) and E2 at the apical level (p < 0.001) were found. CONCLUSIONS: Subclinical hypothyroidism significantly decreased cardiac preload, whereas it increased afterload with a consequent reduction in SV and cardiac output. Replacement therapy fully normalized the hemodynamic alterations.