Background
Amyloidosis is a systemic disorder characterized by the extracellular deposition of circulating proteins into amyloid fibers. These proteins are monoclonal light chains or transthyretin in the two most common forms, namely AL and ATTR amyloidosis, respectively [
1]. Cardiac involvement is common in both AL and ATTR amyloidosis, predicts a worse outcome, and has important implications for treatment [
1].
The diagnostic work-up for cardiac amyloidosis (CA) begins with the identification of clinical features, electrocardiographic (ECG) and imaging findings suggestive or compatible with CA, and often requires the histological demonstration of amyloid deposition, except when diphosphonate scintigraphy shows an intense myocardial uptake (Perugini scores 2-3) in the absence of a monoclonal gammopathy [
2]. Despite its unique capability of allowing myocardial tissue characterization, the role of cardiovascular magnetic resonance (CMR) in this diagnostic flowchart is not well defined [
2]. Several CMR findings are quite specific for CA, most notably a pattern of variable biventricular pseudohypertrophy with diffuse subendocardial-to-transmural late gadolinium enhancement (LGE). The degree of wall thickness increase and LGE extent are correlated to the degree of myocardial infiltration by amyloid fibers. Indeed, amyloid deposition is confined to the subendocardium of few myocardial segments in early disease stages and becomes more diffuse in patients with more advanced disease [
3]. In the later stages, amyloid infiltration is so extensive that myocardial and blood-pool gadolinium kinetics are completely deranged, with diffuse gadolinium retention in the myocardium and an accelerated gadolinium washout from the bloodpool. This may cause problems in identifying the best inversion time (TI) of the myocardium in post-contrast images to obtain good LGE images. In these cases, TI-scout sequences [
4], early-to-late enhancement acquisitions [
5], phase sensitive inversion recovery (PSIR) LGE sequences [
6], as well as native T1 mapping [
7] and extracellular volume fraction (ECV) [
8] may be helpful to establish the diagnosis and define the disease stage. Conversely, LGE areas might be very limited in the earlier disease stages, and lead to erroneous diagnoses of other ischemic or nonischemic cardiac disorders. We may add that several other cardiac and extracardiac findings (such as pericardial and pleural effusion) can be found in CA, but are not specific for this condition. Overall, the ability of human readers to diagnose CA is limited by the highly variable appearance of the disease across different stages, the technical difficulties related to the peculiar gadolinium kinetics in more advanced stages and is highly dependent on operator experience. These possible limitations of human reading might be overcome by using the tools of artificial intelligence (AI).
Machine learning (ML) algorithms build a mathematical model based on sample data, known as "training data", in order to make predictions or decisions without being explicitly programmed to perform the task. Automated ML analysis is faster with similar precision to the most precise human techniques [
9]. Deep learning (DL) is a subset of ML using "raw data" to automatically identify salient features by means of a series of hierarchical representation levels that are not directly designed by humans, as in the case of ML. By avoiding the need for pre-processing techniques based on a priori knowledge of the human operator, DL allows the automatic extraction of salient information from "raw data" by using a series of levels of representation. In the field of medical imaging, integration of DL-based predictive analytics within clinical imaging is a natural order of progression wherein developments in cardiovascular imaging now provide high-fidelity datasets that possess more data than those acquired from prior generation scanners [
10,
11]. The integration of DL-based algorithms with clinical imaging holds the promise to automate redundant tasks and improve disease diagnoses and prognostication, as well as to provide new insights into novel biomarkers associated with specific disease processes [
12].
In the present study we tested the diagnostic performance of CMR-based ML and DL strategies in CA, focusing on conventional LGE images acquired using standardized parameters in a specialized CMR center.
Discussion
We report for the first time that an automated interpretation of CMR exams by a DL-based approach allows to reliably identify patients with CA with a high degree of accuracy. We evaluated a population of patients referred to CMR because of clinically suspected CA; among these patients, 52% were adjudicated as having CA by the standard diagnostic algorithm. A CNN establishing the likelihood of CA based on 2C, 4C and SAx LGE acquisitions was developed, and tested in a population subgroup, where it displayed an AUC of 0. 982. This approach had a similar diagnostic performance than a combination of manually extracted CMR features (p = 0.39 for the comparison of the AUC values) by a ML-based approach, which recapitulates CMR reading by experienced operators.
In the near future, AI applications are expected to “transform cardiac imaging” by “covering a range of applications from image classification, image reconstruction, automation in segmentation and quantification and guiding diagnosis and prognosis” [
19]. In the diagnostic setting, AI can allow accurate image segmentation and automated measurements, can define the likelihood of a specific condition (such as obstructive coronary artery disease based on perfusion single-photon emission computed tomography [
20,
21]), and may assist human readers in diagnosing cardiac disorders, from heart failure [
22] to rare disorders such as CA, which may be misdiagnosed outside of referral centers. An AI-based diagnostic tool would quantify the likelihood of CA based on automated image analysis, and could simplify human interpretation of CMR examinations. To create such a tool, we trained a neural network to establish the likelihood of CA based on the most relevant features from LGE images, which we considered the most relevant acquisitions for diagnostic purposes; this assumption was confirmed by a dedicated analysis showing that a circumferential subendocardial or diffuse LGE pattern is the strongest predictor of CA, followed by LGE presence (Fig.
6a). In the testing subset, the CNN displayed a good diagnostic performance, with an AUC value approaching 1, and a satisfactory accuracy (88%). Importantly, among the 5 patients incorrectly classified by the CNN, there was only 1 false negative: should we have decided whether or not to perform further diagnostic investigations based on automated CMR interpretation, we would have missed only one CA case of 42 patients evaluated (2%). A lower threshold for the likelihood of CA (0.4) had 100% negative predictive value, thus being an ideal threshold to exclude CA.
The fact that only LGE sequences were evaluated through the CNN may be questioned. Nonetheless, it is important to consider that all imaging data from LGE acquisitions were considered, and not LGE patterns alone. While a CNN can be assimilated to a black box, some hints of its functioning are provided by attention maps, which suggested the evaluation of myocardial walls, as well as of the blood-pool, and even of extracardiac findings such as pericardial and pleural effusion (Fig.
5b). According to the training curves (Fig.
2), the LGE acquisition in four-chamber view seemed to be the most informative, possibly because it explores not only the whole heart (including both atria and the right ventricle), but also the presence of pericardial and pleural effusions. We also compared the diagnostic performance of our CNN with image analysis by experienced CMR readers, which was simulated by including manually delineated LV, RV and atrial contours (from which several parameters associated with chamber volumes, mass and function could be calculated), and several categorical variables (LGE presence and pattern, presence of early blood-pool darkening, etc.; Fig.
6a). When assessing the likelihood of CA based on the combination of these findings, which recapitulates the process of CMR interpretation by human readers, this process displayed a similar diagnostic yield than the CNN, with no significant differences between AUC values at discrimination analysis.
These findings corroborate the conclusion that the DL algorithm (which could be easily implemented as a software for automated image analysis) may provide a valuable support to CMR reading when patients are referred for suspected CA. The two main advantages of DL are speed and accuracy. Accuracy appears similar to ML (and likely an experienced observer), but the speed advantage is unquestionable.
In our analysis, we included 19 patients with prior myocardial infarction, all presenting typical regional wall motion abnormalities and subendocardial-to-transmural LGE in the infarcted areas. The DL and ML approaches were not affected by the presence of an ischemic scar, and all 4 patients with prior infarction in the test dataset were correctly classified.
Several limitations of this hypothesis-generating study must be acknowledged. First, sample size was small, although the loss and accuracy curves still displayed a good diagnostic accuracy with no evidence of overfitting (Fig.
2). Second, the prevalence of CA was very high (107 out of 206), and this diagnostic algorithm should be validated in non-specialized centers with a lower prevalence of CA. Third, the study would have also benefited from an external validation cohort with a good representation of patients with hypertensive heart disease, hypertrophic cardiomyopathy, cardiac sarcoidosis and other pathologies that could be mistaken for CA. Fourth, ML on imaging features was considered as a surrogate of expert reading blinded to the clinical data given the retrospective study design. Fifth, PSIR and parametric mapping (native T1 mapping and ECV quantification) were not implemented, because these techniques were not available for the earlier exams, and are still not available at all CMR centers. Sixth, the analysis focused only on CMR findings, and particularly on LGE images, while human interpretation of CMR examinations takes into also clinical data, ECG and echocardiographic findings, etc. On the other hand, the CNN could be easily implemented to consider additional variables for the purpose of diagnosing CA. Seventh, we considered AL and ATTR cardiomyopathies as a single diagnostic entity (CA), given the relatively small patient number. Eighth, the functioning of our DL-based system for image interpretation cannot be explained, by its very nature, unless partially (and in a patient-based fashion) by attention maps, which show which elements of the image are particularly important. Finally, our diagnostic system implies LGE acquisitions obtained at a single high-volume CMR lab using a conventional gradient-echo inversion-recovery sequence and carefully setting the TI to null the normal myocardium and to highlight (as bright) the affected myocardium. This approach is based on the acquisition of 5–10 early enhancement images every minute after contrast injection with a fixed TI as previously validated by our group [
5], and on the acquisition of TI-scout sequences to choose the appropriate inversion time of the normal myocardium and to check the presence of paradoxical blood/myocardium TI; further studies are needed to test this algorithm across different LGE sequences (including PSIR LGE), different contrast doses and different TIs, causing a highly variable signal intensity in cases of CA if acquisition parameters are not standardized. Non-contrast CMR is attracting attention as a potential novel perspective for the diagnosis of CA [
7,
23,
24], and should be considered in future studies.
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.