Bone extraction in X-ray images by analysis of line fluctuations

XNet: a convolutional neural network (CNN) implementation for medical x-ray image segmentation suitable for small datasets

“…We also see a significant improvement when benchmarking against state-of-the-art classical image processing techniques. Figure 5 shows a comparison between our architecture and the work of Kazeminia et al, 4 who built upon that of Bandyopadhyay et al 5,6 The XNet segmentation produces smoothly connected boundaries around the bone regions, in addition to differentiating well between bone and soft tissue regions. It should be noted that in producing this output, our algorithm was trained on a set of high resolution TIF images, as produced by the X-Ray scanner, whereas this analysis was run on significantly lower quality JPEG images, meaning our network could not achieve its full potential on this image.…”

“…Despite our dataset being small compared to those used in many machine learning applications, we achieve an overall accuracy significantly higher, and more generalisable, than work using classical image processing techniques. [2][3][4][5][6]12 Additionally, we show that our architecture outperforms leading image segmentation networks developed for other applications. 13 Our paper is structured as follows: after reviewing a selection of the existing literature in Section 2, we discuss our dataset in Section 3 -how we collect and label the data, and the augmentation methods to prevent overfitting; in Section 4 we address the design and structure of our CNN, covering training and testing stages; the results of the network are discussed in Section 5, where we also address the post-processing stage of false positive reduction; a comparison of our results with other works from the classical and machine learning literature is presented in Section 6, after which we discuss future applications and developments in Section 7.…”

“…The authors are part of the Centre for Doctoral Training in Data Intensive Science at Durham University supported by STFC grant number ST/P006744/1. The authors are grateful to IBEX Innovations Ltd. 32 for hosting them and providing data, and to Kazeminia et al 4…”

“…Entropy-based approaches have been employed by Bandyopadhyay et al, 5,6 which have enabled clean boundary identification, yet suffer from extraneous edges such as bone cracks or image distortions. Kazeminia et al 4 build on this work, employing existing edge detection algorithms such as Sobel 15 and Canny, 16 while analysing the intensity fluctuations in pixel rows to more accurately select the bone boundary. This work is less sensitive to noise, however, loses some boundary continuity.…”

“…These applications regularly require the segmentation of images into 3 categories: open beam, soft tissue and bone. Current methods rely heavily on a complex system of classical image processing techniques, such as clustering approaches, 2, 3 line fluctuation analysis, 4 or entropy-based methods, 5,6 which often require the tuning of hyperparameters for each body part class. However, utilising machine learning offers several advantages over these traditional methods since: (i) it naturally addresses noise, (ii) it generalises well to different body parts, and (iii) the segmented regions have continuous boundaries.…”