Segmentation of the spinous process and its acoustic shadow in vertebral ultrasound images

Abstract

Spinal ultrasound imaging is emerging as a low-cost, radiation-free alternative to conventional X-ray imaging for the clinical follow-up of patients with scoliosis. Currently, deformity measurement relies almost entirely on manual identification of key vertebral landmarks. However, the interpretation of vertebral ultrasound images is challenging, primarily because acoustic waves are entirely reflected by bone. To alleviate this problem, we propose an algorithm to segment these images into three regions: the spinous process, its acoustic shadow and other tissues. This method consists, first, in the extraction of several image features and the selection of the most relevant ones for the discrimination of the three regions. Then, using this set of features and a LDA classifier, each pixel of the image is classified as belonging to one of the three * Corresponding author Email addresses: florian.berton@polymtl.ca (Florian Berton), farida.cheriet@polymtl.ca (Farida Cheriet), marie-claude_miron@ssss.gouv.qc.ca regions. Finally, the image is segmented by regularizing the pixel-wise classification results to account for some geometrical properties of vertebrae. The feature set was first validated by analyzing the classification results across a learning database. The database contained 107 vertebral ultrasound images acquired with convex and linear probes. Classification rates of 84%, 92% and 91% were achieved for the spinous process, the acoustic shadow and other tissues, respectively. Dice similarity coefficients of 0.72 and 0.88 were obtained respectively for the spinous process and acoustic shadow, confirming that the proposed method accurately segments the spinous process and its acoustic shadow in vertebral ultrasound images. Furthermore, the centroid of the automatically segmented spinous process was located at an average distance of 0.38 mm from that of the manually labeled spinous process, which is on the order of image resolution. This suggests that the proposed method is a promising tool for the measurement of the Spinous Process Angle and, more generally, for assisting ultrasound-based assessment of scoliosis progression.

Spine surface detection from local phase‐symmetry enhanced ridges in ultrasound images

“…The mean value of the mean absolute errors is 0.26 mm, and the 90th percentile of mean absolute errors is 0.44 mm with a maximum possible absolute error of 2.01 mm. This is comparable to the mean absolute errors as reported in long bones by Berton et al [0.38 mm between centroids of the spinous process] and Hacihaliloglu et al [0.31 mm], Kowal et al [0.42 mm for cadavers], Foroughi et al [0.3 mm for cadavers], Daanen et al [0.45 mm for patients and 0.27 mm for cadavers], and Jia et al [0.2 mm].…”

“…The mean value of the mean absolute errors is 0.26 mm, and the 90th percentile of mean absolute errors is 0.44 mm with a maximum possible absolute error of 2.01 mm. This is comparable to the mean absolute errors as reported in long bones by Berton et al 31 [0.38 mm between centroids of the spinous process] and Hacihaliloglu et al 29 [0.31 mm], Kowal et al 14 Figure 16 shows the percentage of automatic segmentation (in length) with respect to the expert segmented lamina length. From this graph, we note that the number of false positives far exceeds the number of false negatives.…”

Ultrasound imaging and segmentation of bone surfaces: A review

“…Hybrid approaches, where local image phase information was combined with image intensity information, were also proposed for extraction of bone surfaces from US data 60,61 . Jia et al .…”

“…Berton et al . 61 combined image intensity, gradient and shadow region features with local phase image features and local binary patterns. The obtained image features were used as an input to a classification method segmenting the image into three regions: bone, soft tissue and acoustic shadow.…”

Deteksi Region of Interest Tulang pada Citra B-mode secara Otomatis Menggunakan Region Proposal Networks

“…Beberapa penelitian sebelumnya fokus ke kegiatan mengekstraksi kontur tulang, baik menggunakan pendekatan probabilistik, energetik, maupun regresi, dengan menggunakan fitur fase dan intensitas piksel [5]˗ [8]. Untuk segmentasi tulang pada citra ultrasound, beberapa pendekatan yang digunakan adalah random forest classifier [13], pemanfaatan fitur fase dan shadowing [14], penggunaan energi lokal dan integrated back scattering dengan pendekatan heuristik [15], kombinasi fitur probabilistik, local binary pattern, dan filter Gabor [16], penerapan model snake dan tensor product B-splines approximation [17], dan pendekatan dynamic programming yang diaplikasikan pada beberapa fitur gambar. Di antara penelitian-penelitian tersebut, metode yang paling banyak digunakan adalah klasifikasi linier yang diterapkan pada beberapa kombinasi fitur nilai intensitas, Laplacian of Gaussian filter, dan shadow effect.…”