3D ultrasound system to investigate intraventricular hemorrhage in preterm neonates

Semi-supervised learning framework with shape encoding for neonatal ventricular segmentation from 3D ultrasound

BACKGROUND

Three-dimensional (3D) ultrasound (US) imaging has shown promise in non-invasive monitoring of changes in the lateral brain ventricles of neonates suffering from intraventricular hemorrhaging. Due to the poorly defined anatomical boundaries and low signal-to-noise ratio, fully supervised methods for segmentation of the lateral ventricles in 3D US images require a large dataset of annotated images by trained physicians, which is tedious, time-consuming, and expensive. Training fully supervised segmentation methods on a small dataset may lead to overfitting and hence reduce its generalizability. Semi-supervised learning (SSL) methods for 3D US segmentation may be able to address these challenges but most existing SSL methods have been developed for magnetic resonance or computed tomography (CT) images.

PURPOSE

To develop a fast, lightweight, and accurate SSL method, specifically for 3D US images, that will use unlabeled data towards improving segmentation performance.

METHODS

We propose an SSL framework that leverages the shape-encoding ability of an autoencoder network to enforce complex shape and size constraints on a 3D U-Net segmentation model. The autoencoder created pseudo-labels, based on the 3D U-Net predicted segmentations, that enforces shape constraints. An adversarial discriminator network then determined whether images came from the labeled or unlabeled data distributions. We used 887 3D US images, of which 87 had manually annotated labels and 800 images were unlabeled. Training/validation/testing sets of 25/12/50, 25/12/25 and 50/12/25 images were used for model experimentation. The Dice similarity coefficient (DSC), mean absolute surface distance (MAD), and absolute volumetric difference (VD) were used as metrics for comparing to other benchmarks. The baseline benchmark was the fully supervised vanilla 3D U-Net while dual task consistency, shape-aware semi-supervised network, correlation-aware mutual learning, and 3D U-Net Ensemble models were used as state-of-the-art benchmarks with DSC, MAD, and VD as comparison metrics. The Wilcoxon signed-rank test was used to test statistical significance between algorithms for DSC and VD with the threshold being p < 0.05 and corrected to p < 0.01 using the Bonferroni correction. The random-access memory (RAM) trace and number of trainable parameters were used to compare the computing efficiency between models.

RESULTS

Relative to the baseline 3D U-Net model, our shape-encoding SSL method reported a mean DSC improvement of 6.5%, 7.7%, and 4.1% with a 95% confidence interval of 4.2%, 5.7%, and 2.1% using image data splits of 25/12/50, 25/12/25, and 50/12/25, respectively. Our method only used a 1GB increase in RAM compared to the baseline 3D U-Net and required less than half the RAM and trainable parameters compared to the 3D U-Net ensemble method.

CONCLUSIONS

Based on our extensive literature survey, this is one of the first reported works to propose an SSL method designed for segmenting organs in 3D US images and specifically one that incorporates unlabeled data for segmenting neonatal cerebral lateral ventricles. When compared to the state-of-the-art SSL and fully supervised learning methods, our method yielded the highest DSC and lowest VD while being computationally efficient.

A framework for quantification and visualization of segmentation accuracy and variability in 3D lateral ventricle ultrasound images of preterm neonates

PURPOSE

Intraventricular hemorrhage (IVH) is a major cause of brain injury in preterm neonates. Three dimensional ultrasound (US) imaging systems have been developed to visualize 3D anatomical structure of preterm neonatal intracranial ventricular system with IVH and ventricular dilation. To allow quantitative analysis, the ventricle system is required to be segmented accurately and efficiently from 3D US images. Although semiautomatic segmentation algorithms have been developed, local segmentation accuracy and variability associated with these algorithms should be evaluated statistically before they can be applied in clinical settings. This work proposes a statistical framework to quantify the local accuracy and variability and performs statistical tests to identify locations where the semiautomatically segmented surfaces are significantly different from manually segmented surfaces.

METHODS

Three dimensional lateral ventricle US images of preterm neonates were each segmented six times manually and using a semiautomated segmentation algorithm. The local difference between manually and algorithmically segmented surfaces as well as the segmentation variability for each method was computed and superimposed on the ventricular surface of each subject. To summarize the segmentation performance for a whole group of subjects, the subject-specific local difference and standard deviation maps were registered onto a 3D template ventricular surface using a nonrigid registration algorithm. Pointwise, intersubject average accuracy and pooled variability for the whole group of subjects can be computed and visualized on the template surface, providing a summary of performance of the segmentation algorithm for the whole group of ventricles with highly variable geometry. In addition to pointwise statistical analysis performed on the template surface, statistical conclusion regarding the accuracy of the segmentation algorithm was made for subregions and the whole ventricle with the spatial correlation of pointwise accuracy taken into account.

RESULTS

Ten 3D US images were involved in this study. Pointwise local difference, ΔS, its absolute value |ΔS| as well as the standard deviations of the manual and algorithm segmentations were computed and superimposed on the each ventricle surface. Regions with lower segmentation accuracy and higher segmentation variability can be identified from these maps, and the localized information was applied to improve the accuracy of the algorithm. Intersubject average ΔS and |ΔS| as well as pooled standard deviations was computed on the template surface. Intersubject average ΔS and |ΔS| indicated that the algorithm underestimated regions in the neighborhood of the tips of anterior, inferior, and posterior horns. Intersubject pooled standard deviations indicated that manual segmentation had a higher segmentation variability than algorithm segmentation over the whole ventricle. Statistical analysis on the template surface showed that there was significant difference between algorithm and manual methods for segmenting the right lateral ventricle but not for the left lateral ventricle.

CONCLUSIONS

A framework was proposed for evaluating, visualizing, and summarizing the local accuracy and variability of a segmentation algorithm. This framework can be used for improving the accuracy of segmentation algorithms, as well as providing useful feedback to improve the manual segmentation performance. More importantly, this framework can be applied for longitudinal monitoring of local ventricular changes of neonates with IVH.