Echo2Pheno saves time and resources equivalent to previously reported AI-based echocardiography tools in anesthetized rodents [8–10], but unsurpassed in studies with non-anesthetized rodents. It can annotate a mouse’s entire dataset and extract the features within a minute, whereas manual annotation, based on only two consecutive heartbeats, takes 10-times longer on average. This efficiency is in conscious mouse models, to our knowledge, unique and cannot be approximated manually.
Deep learning models are popular in medical and preclinical echocardiography image analysis [6–10] but all with the challenge that the algorithms are very specifically trained and not generally applicable. In mouse models, echocardiograms from anaesthetized mice show excess domain shift compared to those acquired from non-anesthetized ones, making the algorithms not transferable to all echocardiography datasets. There is a meaningful proportion of scientists including the GMC who prefer conscious echocardiographic diagnostics over anesthesia to avoid known anesthetic-induced impairment of the heart [13]. So far, there is no capability for this set of conscious data than the Echo2Pheno tool newly developed here, which is precisely as such, why this study is so important to fill the gap in modeling deep learning in non-anesthetized mouse models.
The reported Pearson correlation between the manual annotations and mean automatic LVIDd/LVIDs measurements per mouse indicate a good correlation, with acceptable variations. LVIDd and LVIDs showed a small but systematic shift from zero, indicating that Echo2Pheno generally extracts smaller LVIDx than manual annotations, an underestimation that could occur because 1) human bias in selecting two heartbeats tends to focus on “broader” regions in the echocardiogram, which results in larger values in manual LV cavity measurements than the average captured by Echo2Pheno, and 2) the quality assessment has a false-positive rate of 0.118, which could contribute to this small underestimation. Minimal deviation in the standard transducer position (i.e., low quality image) during acquisition typically causes the appearance of superimposed structures, such as the papillary muscles. In this case, the segmentation network tended to generate smaller traces since it was not trained to distinguish between superimposed structures and the true ventricular wall.
However, comparisons were performed for the mean measurements of each mouse derived from two heartbeat cycles in the manual annotations, potentially leading to a sampling bias, and a much larger number of heartbeats for Echo2Pheno. This much larger data pool could lead to questioning the current “ground truth,” suggesting that two heart cycles might be insufficient for obtaining instructive mouse features leading to phenotypic discovery. Nevertheless, the high correlation between manual and automatic measurements is an important indication as it validates the accuracy of our proposed framework with respect to existing pipelines.
Natural physiological fluctuations with highs and lows in the data could not be captured manually until now. However, this phenomenon is present during conscious echocardiographic data acquisition (Fig. 6). LVIDs is more affected by fluctuations due to active myocardial contraction than LVIDd, which denotes cardiac relaxation. Consistent LVID values during echocardiography are therefore not guaranteed, further indicating that the current selection of two cardiac cycles is insufficient for robust diagnoses. Using all echocardiographic data can compensate for extreme values within the echocardiogram recordings due to the large amount of output measurements, consequently generating more reliable cardiac phenotypes. Indeed, of the 16 analyzed studies, Echo2Pheno matched the histopathology findings in 87.5% (14/16) cases (vs. 10/16 for manual), confirming the high accuracy of our method, which exceeds the limits of manual annotations.
For the six studies that were insignificant in both the automatic and manual evaluations (Acnat2, Cmas, Dnajc14, Echs1, Ergic2, Gstm1), the histopathological evaluation reported no peculiarities in heart morphology and suggest that these genes most likely do not affect the analyzed cardiac parameters. In four studies (Cisd1, Dmd, Fabp2, Zfp280d), the manual assessment already noted altered cardiac parameters, consolidated in diagnostics, and further substantiated by the concordant automated analysis and heart diseases for Dmd, Zfp280d, and Cisd1 but not completely for Fabp2, wherein the hearts still appeared to be morphologically normal in the histopathological evaluation.
CISD1, a member of the CDGSH domain-containing family, encodes a mitochondrial outer membrane iron–sulfur protein, mitoNEET [14, 15], which is an important compound in mitochondrial function/metabolism. Cardiac mitochondria isolated from mitoNEET (Cisd1)-null mice exhibit reduced oxidative capacity, suggesting that mitoNEET is an important iron-containing protein involved in controlling the maximal mitochondrial respiration rate [16]. mitoNEET overexpression in the adipose tissue of obese/ob mice significantly reduced inflammation and oxidative stress compared with control mice [17]. When mitoNEET was knocked out, the resulting phenotype was characterized by dopamine neurotransmitter loss in the striatum and Parkinson’s disease-type motor deficits [18]. However, the detailed mechanistic aspects underlying these physiological functions remain unclear. The Cisd1-knockout mouse model described here showed significant LV changes, both manually and automatically, in the male mutants compared with controls. The two parameters of myocardial function (EF and FS) reached significance in male mutants only when automated by Echo2Pheno, while females showed echocardiography data comparable with the controls. For the first time, we describe an association between Cisd1 and in vivo LV phenotypes and confirmed these results via independent cardiac histology. Male Cisd1 hearts showed LV dilatation, although the reasons for the phenotype’s incomplete penetrance (1/2) and sexual dimorphism cannot be explained.
Dmd encodes a large, rod-like cytoskeletal protein found at the inner surface of muscle fibers in skeletal and cardiac muscles. The most prominent mouse model to study DMD is the mdx mouse [19, 20], which shows only mild skeletal muscle histopathology, moderately affected myocardium with mild fibrosis and inflammatory cell infiltration despite the absence of dystrophin expression. Histopathological cardiac function in mdx mice, in contrast to DMD patients, is rarely observed (e.g., when very old or stressed), probably due to the large differences in muscle size, strain, regenerative capacity, and growth phases between humans and mice. Echocardiography data from our 12-week-old DMD single-gene knockout mutants (manually and automatically validated) showed no changes in LV dimensions relative to controls. However, there were altered myocardial capacities represented by EF and FS in the DMD mutants. Independent histopathological examinations of the hearts confirmed our echocardiography results, given the lack of changes in LV morphology; the mutants and controls were comparable. Nevertheless, inflammatory infiltrates, fibrotic lesions, and necrotic foci were observed in the myocardial tissue of all examined DMD mice. These degenerative processes are consistent with other DMD mouse models [19, 20] and demonstrate our phenotype detection’s accuracy. Hence, the results of our Dmd knockout model agree with existing DMD mouse models and could be considered a valid proof-of-concept model for the automatic approach.
The fatty acid binding protein 2 (Fabp2) was the only mouse study wherein cardiac histology revealed no changes, which is unsurprising given that the in vivo phenotype of this study was based on only one parameter (LVIDd) and sex (female) relative to control mice. A unique phenotype need not be a false positive; rather, such a moderate phenotype is difficult to detect in histology because the morphological changes are (still) mild during testing. Fabp2 plays a key role in the absorption and intracellular transport of dietary long-chain fatty acids. Numerous studies have examined the association between Fabp2 gene polymorphisms and Type 2 diabetes mellitus [21]. Several mouse models [22] provide evidence that liver Fabp2 plays an important role in intestinal lipid metabolism; however, there remains a lack of its association with cardiac function and morphology in the mouse model except for our data.
In contrast, the echocardiography data of Zfp280d mutants showed severely dilated and bilaterally confirmed (manual and automatic) LV, and thus severe myocardial function impairment, particularly EF and FS, relative to control mice. The cardiac histology of the Zfp280d mutants showed enlarged LVs in the males, independently confirming our in vivo data. The zinc finger protein 280d (Zfp280d in mice and ZNF280D in humans) has not been previously associated with heart disease; we have therefore made a confirmatory discovery (manually and automatically). In general, ZNF280D is sparsely described and has no disease associations in the literature barring dyslexia [23]. Hence, we have undoubtedly found and described a new candidate gene with a strong link to congenital dilated cardiomyopathy.
In two studies (Kansl1l, Uggt2), the automatic Echo2Pheno evaluation invalidated the significant differences based on the manual evaluation. Whether they are “false positive” phenotype annotations depends on the analysis (i.e., gold standard). Interestingly, the hearts’ histopathological examination revealed no differences between the mutants and controls. Hence, the automatically generated in vivo data and morphology yielded the same results, outweighing the manual annotations. With this new type of Echo2Pheno evaluation, retrospective genotype–phenotype associations can be checked, and if necessary, corrected with reasonable effort.
The automated Echo2Pheno analysis revealed novel phenotypes in LVIDs, LVIDd, EF, and/or FS in four studies (Cnot6l, Gatb, Slc6a15, Sytl4). When manually annotated, one study (Cnot6l) had shown slight changes in EF and FS in males when manually annotated, while the remaining three were insignificant; thus, no genotype–phenotype association had previously been postulated. Echo2Pheno detected a phenodeviation in these studies. Whether they are “false negative” phenotype annotations depends on the analysis (i.e., gold standard).
Cnot6l had previously shown modified EF and FS in male mutants compared with controls. Presently, all (males and females) Cnot6l mutants had significantly altered LVIDs, EF, and FS. Histopathological examinations of the hearts confirmed our new phenotypes by altered LV morphology in Cnot6l mutants. To date, no human diseases have been associated with Cnot6l nor is its genetic influence on cardiovascular disease known. Cnot6l is conserved from yeast to humans [24] and maintaining its structural integrity and enzymatic activity is important for controlling cell viability [24, 25]. Our mouse data describe, for the first time, a causal association between Cnot6l and the heart and dilated LV.
Analysis of Gatb by Echo2Pheno showed new differences in EF and FS in the female mutants. Both parameters describe myocardial performance rather than the LV cavity. Accordingly, no morphological LV changes were detected in the cardiac histology; however, pericarditis was detected in 25% of the examined female hearts, which could explain the impaired pumping capacity in Gatb mice. Observations regarding Gatb mice are nonetheless interesting because of an already established link with congenital severe heart disease. Genetic defects in a subunit of the GatCAB complex, including Gatb, have recently been described in patients with fatal metabolic cardiomyopathy syndrome [26]. The Gatb knockout mouse model could contribute to the modeling and research of congenital metabolic cardiomyopathy syndrome.
Slc6a15, in contrast, showed significant differences in LVIDs, EF, and FS in male mutants; females remained comparable in all parameters. Histology of the heart, however, showed no LV changes in the two male hearts studied and could not support this new phenotype. The small number of hearts examined in a high throughput screening at 16 weeks of age, the four weeks age difference between echocardiogram and pathology screens and their arbitrary selection for histology (two out of five mutants per sex) could explain the different results. Nonetheless, this genotype–phenotype association is valuable because this gene encodes for protein family solute carrier family 6, which participates in neuronal amino acid transport, potentially associated with major depression [27, 28]. Before our report, no potential link of Gatb to heart disease had been described in mice.
The Sytl4 gene is relevant to neuronal system development and is implicated in neurological and psychological diseases. Sytl4 is most likely a candidate gene for autism [29] but thus far, has no known association with cardiac impairment. Using Echo2Pheno, we found an association between Sytl4 and cardiac impairment in male knockout mice, particularly in LVIDs, EF, and FS. Histopathologically, a dilated LV was confirmed and myocardial fibrosis was detected. These observations provide a new causal association between Sytl4 and the heart, possibly even myocardial remodeling.
The detection of the cardiac phenotypes of these four exemplary genes (Cnot6l, Gatb, Slc6a15, Sytl4) was possible only by an automated and fully comprehensive analysis of all data. The established manual method would not have attributed any significance to these genes, thereby overlooking important findings, given that three of those associations were already confirmed by histopathological findings. Although one (Gatb) has already been linked to heart disease in the literature, the other three phenotypes (Cnot6l, Slc6a15, Sytl4) describe novel findings worth further exploration.
By design of high-throughput, such an automated approach does not come without limitations. Because of a particular type of echocardiography (restricted to SAX m-mode) from conscious (i.e., without sedation) 12-week-old mice with normal body weight, the Echo2Pheno algorithm is quite constrained and fails for anesthetized mice and overly dramatic phenotypes of known severe heart diseases, such as hypertrophic and dilated cardiomyopathies. Greater generalization needs to be achieved by feeding the models with larger and more diversified training data. This can occur by the incorporation of multi-centric large-scale data from the International Mouse Phenotyping Consortium (IMPC) with conscious and anesthetized echocardiography data enabling Echo2Pheno-based evaluation of all mouse echocardiographic data types. During conscious echocardiography, sudden movements and incorrect positioning of the transducer can degrade image quality, and image features may occur that do not accurately reflect the cardiac status of the mouse. On average, 49.6% of the echocardiogram data in this study did not meet our quality standards, were classified as low quality and excluded from further analysis (Fig. 7). Improving ad hoc data collection could minimize data loss. As this is a new approach designed for short-axis m-mode echocardiography of conscious mice, there is no publicly available benchmark dataset that we can use for comparison yet.
We believe that automated evaluation of all technical replicas (e.g., DICOM files) provides a new look at conscious echocardiography data and enables the extensive evaluation in all areas of the echocardiogram. The discrimination of high- from low-quality recordings provides previously impossible but important insight into the data quality and operator effect in echocardiographic data acquisition.
Echo2Pheno, a deep-machine-learning approach, makes high-throughput conscious m-mode SAX echocardiography data analysis fully automatic, simple, and fast. Thus, we propose Echo2Pheno-based evaluations as a state-of-the-art method for future echocardiography analysis to achieve standardization and reproducibility in conscious mouse models. This lays the foundation for overcoming the sampling bias and exploring previously unused data, while providing novel insights into gene function linked to the heart.