Identification of anatomical landmarks on 4-chamber views for CMR workflow automation

Background: The four-chamber (4CH) view is routinely acquired in most cardiac MRI protocols and contains several anatomical structures of interest that can be leveraged to prescribe more advanced views. Such is the case of the right ventricle apex, required to prescribe the right ventricular in-/outflow view (RVIO), and the mitral (MV) and tricuspid valve (TV) attachments, which identify the optimal basal slice for stroke volume measurement [1].

We present here the evaluation of a deep learning network that automatically identifies these structures on bSSFP 4CH views. The network has been integrated into a previously presented cardiac workflow framework [2,3], enabling the automated prescription of RVIO, MV and TV views.

Methods: A deep learning model was developed to automatically identify the following landmarks in 4CH views: MV leaflet attachments; TV leaflet attachments; left ventricle apex; and right ventricle apex. The model was trained on a manually annotated dataset of 1,345 bSSFP frames from 615 independent clinical exams.

The model was configured as a 2D U-Net with five levels of feature depth (8 to 128 channels), including max pooling, batch normalization and dropout. Optimization used the Adam optimizer and a Dice metric, with an initial learning rate of 3e-5. Data augmentation included translation, rotation, scaling, and intensity transformations.

The test database consisted of a set of 163 4CH frames from 100 clinical exams acquired at a second site, including structural, ischemic and non-ischemic heart disease cases. These were automatically labelled using the model and then reviewed by an experienced CMR clinical scientist.

Results: The model correctly identified all target landmarks in 80.4% of the test cases. Among the remaining cases, 50% were flagged by the reviewer solely due to a discrepancy in the definition of the right ventricular apex used in the training dataset, as illustrated in Figure 1.

Table 1 details the incidence of model failures for each of the target landmarks. In the case of the valve attachments, in all but one cases the issue was detectable, with the model tagging the landmark as not found. Amongst these were three instances of prosthetic MV, which will require expanding the training database. There were also three cases of suboptimal views (two with insufficient right chamber coverage and one with excessive aortic output), as well as a markedly dilated right ventricle impeding the identification of the TV. Of the two cases of erroneous left apex placement, one was due to ambiguity caused by severe hypertrophy.

Conclusion: We presented an evaluation of a deep learning model for identifying anatomical landmarks in 4CH cardiac MR series. Preliminary results indicate that the current architecture is well-suited for this task.

Overall, model performance could be further improved by adopting a consistent definition of the right ventricular apex and by incorporating a broader range of pathological cases into the training dataset.

Table 1: Number of model failures for each target landmark, including cases where the model either reported being unable to detect the landmark or produced an incorrect prediction.

Landmark	Not detected	Misplaced
Mitral valve (left)	7 (4.3%)	0 (0.0%)
Mitral valve (right)	2 (1.2%)	0 (0.0%)
Tricuspid valve (left)	3 (1.8%)	1 (0.6%)
Tricuspid valve (right)	5 (3.1%)	0 (0.0%)
Apex (left)	0 (0.0%)	2 (1.2%)
Apex (right)	5 (3.1%)	14 (8.6%)

Figure 1: Example images showing the model’s predicted landmarks (red) and the reviewer’s corrections (green): a) Typical case with no corrections required; b) Discrepancy in the definition of the right ventricular apex; c) Model failure to identify landmarks in the presence of a prosthetic valve.
Figure 2: Histograms showing the reviewer’s corrections for each landmark misidentified by the model.