Robust cardiac segmentation corrected with heuristics

Abstract

Cardiovascular diseases related to the right side of the heart, such as Pulmonary Hypertension, are some of the leading causes of death among the Mexican (and worldwide) population. To avoid invasive techniques such as catheterizing the heart, improving the segmenting performance of medical echocardiographic systems can be an option to early detect diseases related to the right-side of the heart. While current medical imaging systems perform well segmenting automatically the left side of the heart, they typically struggle segmenting the right-side cavities. This paper presents a robust cardiac segmentation algorithm based on the popular U-NET architecture capable of accurately segmenting the four cavities with a reduced training dataset. Moreover, we propose two additional steps to improve the quality of the results in our machine learning model, 1) a segmentation algorithm capable of accurately detecting cone shapes (as it has been trained and refined with multiple data sources) and 2) a post-processing step which refines the shape and contours of the segmentation based on heuristics provided by the clinicians. Our results demonstrate that the proposed techniques achieve segmentation accuracy comparable to state-of-the-art methods in datasets commonly used for this practice, as well as in datasets compiled by our medical team. Furthermore, we tested the validity of the post-processing correction step within the same sequence of images and demonstrated its consistency with manual segmentations performed by clinicians.

Fig 1. An example of a low resolution, annotated echocardiographic image.

Fig 2. An example of the cone segmentation task executed in a renal liver ultrasound.

a, the original image features all the technical elements of the ultrasound scan. b, the predicted mask from the cone segmentation superimposed on the ultrasound scan, c, the mask smoothed and applied over the image. d, finally a dilation step is applied to improve the cone segmentation.

Fig 3. U-Net architecture implemented.

Fig 4. An example of a mask from the right ventricle (top-left in pink labeled VD) surrounding the right atrium (bottom-left in violet labeled AD).

Fig 5. Mask VD after the correction.

Fig 6. The final result improved for visualization purposes.

Fig 7. Some instances from the custom dataset with their respective ground truth masks.

Fig 8

Left (GT): Mask generated by the clinician’s labeling. Center (PR): Mask predicted by the model. Right (CR): Predicted mask after heuristic correction.

Fig 9. Visible differences between the PR mask and the CR mask.

Fig 10. Loss history for the training (red) and validation (blue) phases.

Fig 11. Dice Coefficient (dice_coeff), Intersection over Union (IoU) and mean pixel accuracy achieved in the validation phase.

Fig 12. Two examples from the first case where the segmentations were very close to the ground truth.

It can be proved the good quality of the images for this case (especially on the contrast between the cavities chambers and the rest of the heart structure).

Fig 13

There are larger left ventricles (b) and variable shapes for the rest of the cavities and less defined borders (a) for this two examples. The model achieved good segmentations for both.

Fig 14

In this third case, the quality drop is notorious for this two examples: (a) and (b). This is reflected in how segmentation quality is lower compared to the two previous cases but the model does not lose shape sense and keeps them regular given the cavity.

Fig 15. Two examples of the fourth case, where annotations are incorrect and additional areas have been marked.