InfoFusion

Machine Learning for Automated Mitral Regurgitation Detection from Cardiac Imaging

Ke Xiao, Erik Learned-Miller, Evangelos Kalogerakis, James Priest & Madalina Fiterau

MICCAI 2023 Paper, MR vs non-MR video, Patch Selection video, Github repo

Mitral regurgitation (MR) is a heart valve disease with potentially fatal consequences that can only be forestalled through timely diagnosis and treatment. Traditional diagnosis methods are expensive, labor-intensive and require clinical expertise, posing a barrier to screening for MR. To overcome this impediment, we propose an automated five stage pipeline named Cardio-vascular magnetic resonance U-Net localized Self-Supervised Predictor (CUSSP). CUSSP operates on cardiac magnetic resonance (CMR) imaging slices of the 4-chamber view of the heart. It uses standard computer vision techniques and contrastive models to learn from large amounts of unlabeled data, in conjunction with specialized classifiers to establish the first ever automated MR classification system using CMR imaging sequences. Evaluated on a test set of 179 labeled — 154 non-MR and 25 MR — sequences, CUSSP attains an F1 score of 0.69 and a ROC-AUC score of 0.88, setting the first benchmark result for detecting MR from CMR imaging sequences.

Illustration of the mitral valve and mitral regurgitation.
Figure 1. Three CMR images showing the long-axis four-chamber view of the heart. Left: a heart with normal mitral valve. Middle: a heart with normal mitral valve when the valve leaflets are open. Right: a heart with mitral regurgitation. The red dotted line indicates the mitral valve

Challenges:

  • Requires multiple video frames, and tracking blood flow.
  • The regurgitant jet can be faint in the video, and thus difficult to detect.
  • Blood flow has high turbulence

Segmentation

  • Segmentation for CMR views including long-axis 2-chamber (LAX-2CH) view, 4-chamber (LAX-4CH) view and short-axis (SA) view.
  • For patch selection:
    Semantic segmentation of 4CH CMR imaging sequences.
  • For tabular data:
    Semantic segmentation of LAX-2CH, LAX-4CH and SA CMR imaging
    sequences (to estimate the cardiac structure and functions in Table 1).
Example of the segmentation outputs
Figure 2. Example of the segmentation outputs of the long axis 4CH (left), 2CH (middle) CMR
view imaging data and the short axis (right) CMR imaging data

Left AtriumRight AtriumLeft VentricleRight Ventricle
Vol Max (mL)
Vol Min (mL)
Stroke Vol (mL)
Ejection Fraction (%)		Vol Max (mL)
Vol Min (mL)
Stroke Vol (mL)
Ejection Fraction (%)		End Systolic Vol (mL)
End Diastolic Vol (mL)
Stroke Vol (mL)
Ejection Fraction (%)
Cardiac Output (L/min)
Mass (g)		End Systolic Vol (mL)
End Diastolic Vol (mL)
Stroke Vol (mL)
Ejection Fraction (%)
Table 1. Cardiac measurements derived from the semantic segmentation of the CMR.

Our Prior Attempts: Random Forests and CNN-LSTM

  • Random forest (RF) classifier trained on the tabular heart measurements shown in Table 1.
  • A deep learning model using a CNN-LSTM pipeline, shown in Figure 3.
Figure 3. Overview of the CNN-LSTM method pipeline for MR classification. The pipeline includes segmentation, masking and cropping for pre-processing, and DenseNet as the frame encoder, bi-directional LSTM as the sequence encoder, with an injected attention layer after the first convolutional layer of the DenseNet block.

Patch Selection

  • Localization: localizing the region in the heart where mitral regurgitation occurs.
  • Equalization: increasing contrast for the regurgitant jet.
  • Selection: cropping the patch for the ML system.
Figure 4.Top: the 4CH CMR image in the original contrast (left), after histogram equalization based on the whole image (middle), and after histogram equalization based on the patch (right). Blue contours outline the left atrium, and the red square boxes outline the selected patch. Bottom: Example of the cropped mitral valve patch as outlined in the red square boxes above.
Patch selection process for CUSSP pipeline.

CUSSP Pipeline Stages

  • CUSSP pipeline: (1) segmentation, (2) localization, (3) cropping, (4) equalization, and (5) prediction
  • Prediction steps: (i) Barlow-Twins network, (ii) Siamese network, (iii) Multilayer perceptron (MLP).
Figure 5. Overview of the CUSSP pipeline for MR classification