Deep learning techniques for bone surface delineation in ultrasound

Abstract

For computer-assisted interventions in orthopedic surgery, automatic bone surface delineation can be of great value. For instance, given such a method, an automatically extracted bone surface from intraoperative imaging modalities can be registered to the bone surfaces from preoperative images, allowing for enhanced visualization and/or surgical guidance. Ultrasound (US) is ideal for imaging bone surfaces intraoperatively, being real-time, non-ionizing, and cost-effective. However, due to its low signal-to-noise ratio and imaging artifacts, extracting bone surfaces automatically from such images remains challenging. In this work, we examine the suitability of deep learning for automatic bone surface extraction from US. Given 1800 manually annotated US frames, we examine the performance of two popular neural networks used for segmentation. Furthermore, we investigate the effect of different preprocessing methods used for manual annotations in training on the final segmentation quality, and demonstrate excellent qualitative and quantitative segmentation results.

Delineating Bone Surfaces in B-Mode Images Constrained by Physics of Ultrasound Propagation

“…Since neural networks require a large training set and are susceptible to domain shift, the UNet would have been quite disadvantaged if trained on diverseUS, due to its limited 37 images, among which only few are of the forearm region. We therefore presented here UNet results trained with another forearm dataset of 1385 images from a different subject, with which UNet then expectedly outperforms all other methods substantially, in agreement with [27]. Given that it is not always practical to annotate and train on large sets of a targeted anatomy, UNet results reported in [27] are herein only presented as a reference.…”

“…We therefore presented here UNet results trained with another forearm dataset of 1385 images from a different subject, with which UNet then expectedly outperforms all other methods substantially, in agreement with [27]. Given that it is not always practical to annotate and train on large sets of a targeted anatomy, UNet results reported in [27] are herein only presented as a reference. Other works in literature using a similar experiment setup (SonixTouch US machine) report UNet results of 2.43 mm [25] and 0.39 mm [26] for MED metric when only B-mode US images were used as input, where training set sizes were 300 and 415 images, respectively.…”

“…Similarly, CBG improves our earlier proposed method BFG 42% on average for these metrics. For all metrics except MSE, both BFG and CBG achieve performance superior to inter-annotator variation reported for this dataset in [27]. Since neural networks require a large training set and are susceptible to domain shift, the UNet would have been quite disadvantaged if trained on diverseUS, due to its limited 37 images, among which only few are of the forearm region.…”

“…We evaluated bone surface delineation performance of the compared methods on the forearmUS dataset. In Table VI we present a quantitative comparison of the proposed CBG with BFG [28], PS [17], CPS [19], and UNet [27]. PS and CPS required no training and were parameter-optimized on diverseUS.…”

“…Later, additional deep learning based methods fused with phase tensors were proposed to accurately segment bone surfaces [25], [26]. In a recent study, state-of-the-art bone surface delineation accuracy is shown to be achieved with a standard CNN-based endto-end learning approach [27], hence substantially reducing the inference time. It is noteworthy to emphasize on the influence of available datasets in works utilizing CNNs, where works with relatively smaller datasets (300+ images) report substantially inferior results when a CNN is trained using US images alone [25], [26].…”