[MICCAI 2019] Comparative Evaluation of Hand-Engineered and Deep-Learned Features for Neonatal Hip B

Comparative Evaluation of Hand-Engineered and Deep-Learned Features for Neonatal Hip Bone Segmentation in Ultrasound

1. Author

Houssam El-Hariri¹, Kishore Mulpuri², Antony Hodgson³, and Rafeef Garbi^1
1 Department of Electrical and Computer Engineering,
University of British Columbia, Vancouver, BC, Canada
² Department of Orthopedic Surgery, BC Children’s Hospital,
Vancouver, BC, Canada
³ Department of Mechanical Engineering, University of British Columbia,
Vancouver, BC, Canada

2. Abstract

Undetected cases may lead to serious consequences including limping, leg length discrepancy, pain, osteoarthritis, disability, and total hip replacement.
Diagnosis typically relies on ultrasound (US) screening of the infant hip between 0–4 months of age.
An inexpensive and safe non-ionizing modality, US imaging enables measurement of DDH metrics based on hip bone features such as the $\alpha$ angle.

To improve bone localization, from which the metrics are calculated, we first build upon recent phase-based feature extraction by applying spatial anatomical priors to eliminate false positives and accurately segment the ilium and acetabulum contour.
Second, we propose the use of deep-learned features, using the popular UNet with single and multi-channel inputs.

3. Introduction

Developmental dysplasia of the hip (DDH) is one of the most common disorders seen in newborns, with prevalence up to about $3\%$, and incidence up to about $7.5\%$ in some populations, ranging widely between populations.

Traditionally, the standard metric for clinical diagnosis is the $\alpha$ angle, defined as the angle between the vertical cortex of the ilium and acetabular roof, measured from a coronal ultrasound image of the hip.

More recently, three-dimensional (3D) US has been proposed as an alternative to two-dimensional (2D) US for more reproducible DDH measurement.

Typically, a prerequisite for extracting DDH metrics such as the α angle is accurate segmentation of the ilium and acetabulum bone surfaces.

Manual annotation of bone surfaces is laborious and time consuming, particularly in 3D data, as well as error-prone potentially affecting DDH diagnosis.

In this paper, we test the hypothesis that deep-learned features can localize bone more accurately than current state-of-the-art phase- and shadow-based features for this task, and whether this technique performs equally well when applied to different ultrasound probes.

4. Methods

4.1 Hand-Crafted Features

We include CSPS and SP in our comparisons given their consistently good performance on neonatal hip ultrasound.

To improve performance we incorporate the spatial prior that the ilium and acetabulum are a continuous bone structure that always appears as the most medial (or deepest with respect to the probe) and superior bone in the image.

To apply this spatial prior, we start with the observation that SP only detects one structure along each vertical scan line, which due to its high acoustic impedance is likely to be bone.

To find this region, we first define the set of k regions $CC=\{{CC}_1,...,{CC}_i,...,{CC}_k\}$ obtained by applying connected component analysis to the SP binary segmentation mask, with 8-connectivity test in 2D and 26-connectivity in 3D.

4.2 Deep Learned Features

4.2.1 Architecture

We use the U-Net architecture for our task of segmenting the ilium and acetabulum.

We choose U-Net for its proven performance on medical image data, ability to train on very few training samples, and for direct comparison with a recent paper that has attempted to use U-Net for DDH segmentation.

We explore training U-Net with two types of inputs:
(1) Raw B-mode image data, and (2) a multi-channel input based on results from several recent papers on bone segmentation that have shown much improved accuracy of bone localization with this method.

In our implementation, the multi-channel input includes the B-mode image, the corresponding SPS, and shadow confidence map features in the R, G, B channels, respectively (see Fig. 1)

4.2.2 U-Net Training

To prepare data, we start with 231,384 coronal slices, obtained with the Ultrasonix probe from 59 neonate scans as potential training data.
In summary, we end up with 439 adequate, labelled samples in our training set from the Ultrasonix set.

We intentionally did not include Clarius samples in the training set to test generalizability of U-Net on different domains.