Improving a Deep Learning Model to Accurately Diagnose LVNC

Abstract

Accurate diagnosis of Left Ventricular Noncompaction Cardiomyopathy (LVNC) is critical for proper patient treatment but remains challenging. This work improves LVNC detection by improving left ventricle segmentation in cardiac MR images. Trabeculated left ventricle indicates LVNC, but automatic segmentation is difficult. We present techniques to improve segmentation and evaluate their impact on LVNC diagnosis. Three main methods are introduced: (1) using full 800 × 800 MR images rather than 512 × 512; (2) a clustering algorithm to eliminate neural network hallucinations; (3) advanced network architectures including Attention U-Net, MSA-UNet, and U-Net++.Experiments utilize cardiac MR datasets from three different hospitals. U-Net++ achieves the best segmentation performance using 800 × 800 images, and it improves the mean segmentation Dice score by 0.02 over the baseline U-Net, the clustering algorithm improves the mean Dice score by 0.06 on the images it affected, and the U-Net++ provides an additional 0.02 mean Dice score over the baseline U-Net. For LVNC diagnosis, U-Net++ achieves 0.896 accuracy, 0.907 precision, and 0.912 F1-score outperforming the baseline U-Net. Proposed techniques enhance LVNC detection, but differences between hospitals reveal problems in improving generalization. This work provides validated methods for precise LVNC diagnosis.

Keywords: MRI Image segmentation; cardiomyopathies; convolutional neural networks; left ventricular non-compaction diagnosis.

Figure 1

Output slices for the patient P241 (from the test set) from the U-Net++. Green indicates the compacted external layer of the left ventricle, yellow the trabecular zone, and light blue the internal cavity.

Figure 2

Sample images of each dataset: (a) P dataset, (b) X dataset, and (c) H dataset.

Figure 3

Diagram of the entire 800 × 800 image neural network.

Figure 4

Example of the clustering algorithm improving an image, showing in red in (b,d) the errors of the corresponding outputs (a,c).

Figure 5

Attention U-Net used in the article [28].

Figure 6

MSA-UNet modified from [29].

Figure 7

Basic architecture of U-Net++ with an additional layer for output fusion called MOST (Multiple Side-Output Fusion) for training modified from [30].

Figure 8

This graph displays the changes that occurred in the U-Net due to the effects of Algorithm 2 in the five folds simultaneously. The same image can appear multiple times. The unaltered values are shown on the “x” axis, and the results after clustering are shown on the “y” axis. Improved results are indicated in green, whereas worsened results are shown in red. The blue line is meant to help distinguish improvements from worsening.

Figure 9

(a) Confusion matrix for LVNC detection from the article [27]. (b) Mean of the confusion matrices obtained from 5 to fold cross-validation using U-Net++ on the entire dataset (P + X + H) with a threshold of 27.4%. The standard deviation across all values is approximately 3 or 4 patients, omitted for simplicity in comparison.

Figure 10

Determine the Dice coefficient between the area of a given section (measured in pixels) and its corresponding image number.

Figure 11

Dice coefficient against the area of that section (measured in number of pixels).