The inclusion criteria for participants were between 18 and 65 years old, having SCI that occurred over a year prior to the start of the study and using a manual wheelchair as the primary means of mobility (self-propel at least 40 h per week). Participants were excluded from this study if they had a history of fractures or dislocations in the shoulder, from which they had not fully recovered; upper limb dysthetic pain because of a syrinx or complex regional pain syndrome; and a history of cardiovascular and cardiopulmonary diseases. Participants were excluded if they had pain in an upper extremity that would interfere with normal functioning and daily activities.

Quantitative ultrasound protocol has strong reliability and validity to determine acute changes in tendon appearance and subacromial space with repetitive wheelchair activity [9,10,17–19]. The rater reliability of AHD measures has been reported good to excellent in SCI population [20]. As recommended to improve reliability of the measures, a single examiner (YSL) with experiences in clinical ultrasound shoulder examinations among wheelchair users scanned and recorded all ultrasound images in a randomized order [9,20]. Image field depth was set at 5 cm, and gain was set at 75 dB in each ultrasound system to minimize the variations of echogenicity of raw images that were optimized based on the relative contrasts to differentiate bone surfaces from other surrounding tissues [17]. Shoulder ultrasound images were captured from the Philips HD11 1.0.6 ultrasound machine with a 5–12 MHz linear transducer or the Supersonic Imagine Aixplorer Ultrasound Machine with an SL14-5 transducer placed on the skin above the deltoid muscles of the non-dominant arm. Each video clip was recorded at 10 Hz for approximately 10 s. Operators recorded two shoulder ultrasound clips to ensure that each frame included scapular and humeral structures. AHD measurements were recorded on the non-dominant shoulder in four different positions: shoulder in a neutral resting position without weight, neutral position with a 3-lb (1.4 kg) weight, 45° and 90° arm elevation in the scapular plane with humeral internal rotation (IR) while holding a 3-lb weight, similar to the methods in previous research protocols [20]. AHD ultrasonographic images were recorded at 10 Hz for each scan using an external video recorder. All collected video clips were converted into a series of ultrasound (US) snapshots and processed in the MATLAB software (Mathworks, Natick, United States). The first and the last 2 seconds of each clip were excluded from the analysis due to the transducer’s preparation and separation from the region of interest. A licensed physical therapist (AS), experienced in imaging AHD and blinded from the data collection and experimental setup, performed manual identification of bony landmarks for calculating the AHD.

Our computer-aided approach included two sections. The first section demonstrated the pipeline of echogenicity-based extraction from snapshots of pre-recorded ultrasound video clips. The second section described the CAQUSA based on the probability map used to extract the superior surfaces of the acromion and humeral head for AHD measurement within each ultrasound snapshot. The flowchart of the proposed CAQUSA method for processing an ultrasound video clip to measure AHD is illustrated in Figure 1.

This dataset comprised a total of 116 video clips obtained from 10 manual wheelchair users with SCI.

As a significant proportion of the acoustic energy was reflected toward the transducer when comparing the reflected signal from the bone relative to surrounding structures, the acromion and humeral head surfaces were more hyperechoic than the deltoid and supraspinatus tendon in ultrasound images (Figure 2A). The greater pixel values above the shaded area were identified as a bone surface with augmented features through our proposed development of post-processing techniques. Figure 2B represented the inadequate image quality of bone surface to proceed the analyses. Boundaries of bone surfaces were effectively isolated by a local phase extraction procedure, as the echo signal beam reflection was enhanced on bone surfaces and attenuated in other regions. Bone surfaces in ultrasounds exhibit ridge-like responses that were well captured by local image phase symmetry features [14]. We implemented the phase symmetry Log-Gabor filter using the monogenic signal to compute the odd and even responses from band-pass quadrature features [21]. This filter could construct the arbitrary bandwidth to characterize the shadowing effect below the bone surface in ultrasound images (Figure 2C). We applied a binary threshold mask that filtered the neighboring pixels and intensity values lower than two standard deviations from the average neighborhood echogenicity to exclude shallow soft tissue structures above the bone surfaces, thereby reducing the processing time and efficiency (Figure 2D). The Laplacian Gaussian filter detected reflections from US images and carried zero filter responses to a boundary of bright, ridge-like structures. US waves did not penetrate the bone structures, showing higher intensity from soft tissues above and lower intensity from those below the bone structures. This filter resorted to dichotomizing positive responses to low pixel intensity and negative responses to high pixel intensity, resulting in brighter surface areas on the acromion and humeral head and other structures The Laplacian of Gaussian filter was then applied to blur images from the Gaussian kernel (Figure 2E). The shadowing effect of the image was used to eliminate unwanted ridge-like structures above a bone area (Figure 2F). We then multiplied the Gaussian kernel using Gaussian weighting function to calculate the pixel intensities below the shadowiness and the weighted sum of the pixels used to smooth images and remove details and noises as described previously [22,23]. To minimize false-positive and false-negative detections from structures bearing similar echogenicity to bone features in US images, we further processed our images with a backscattering to reduce the interference from soft tissues above the bone structure (Figure 2G). The scan lines of ultrasound backscattering energy were considered to enhance the probability of detecting bone areas, as most of the ultrasound energy was higher in areas of the bone structure. The integrated backscattering energy was calculated by the sum of squared intensities above a given pixel on a scan line to reduce the number of incorrect probability responses [14]. The bone-probability map was constructed by the multiplication of normalized products from the Log-Gabor filter with the band-pass quadrature filter to integrate backscattering and bone shadowing features [24] (Figure 2H).

We identified the acromion and humeral head based on the degree of discontinuity acquired from each vertical (column) scan line in US images using a threshold value of 0.5, as described in previous literature, to separate bone surfaces from other unwanted structures [14,21]. We then optimized a bone probability map for a bone-detection process [22]. The Cost (i, j) function (Eq. 1) represents the continuity of pixels to enhance the segmentation of bone structures and eliminate unwanted hyperechogenic or hypoechogenic pixels of segmented bone structures. C o s t   i , j = p i x i + 1 − p i x i 2 + p i x j + 1 − p i x j 2 (1)

In Eq. 1, i and j are rows and columns, respectively, and Pix indicates pixel coordinates in a 2D.

US image. Cost (i, j) represents the distance between pixels in the jth scan line and the adjacent scan line. A higher referred to adjacent pixels being sparser; therefore, it carried more degrees of discontinuity. Detected bone-like surfaces in US images were not always firm. Thus, we enhanced those features by filling in gaps between pixels with a degree of discontinuity lower than a certain number of pixels, n, in Eq. 2. C o n d i t i o n = p a t h ,     i f   c o s t i , j ≤ n C o n t i n u e   i f   c o s t i , j   >   n (2)

Specific conditions for Eq. 2 are shown above, where j is the column of a US image, and n is a constant that is determined by the user based on the specs of the US machine, including resolution and scale. In a user-specified range of j scan lines, the algorithm found two specific pixels that carried Cost (i, j). If Cost (i, j) was less than or equal to the user-specified constant n, the two pixels were assumed to have originated from homogeneous bone-like features. The algorithm connected the two pixels by the creation of a path. If Cost (i, j) was higher than n, the two pixels were assumed to have originated from two heterogeneous bone-like features. Therefore, the algorithm continued to the following scan lines to find other pixels in the following conditions. When a path was created, other pixels between the two specified pixels were assumed to be soft tissue and, therefore, eliminated. According to the principle of ultrasound bean reflection, only one pixel per column on the bone probability map is recognized to initiate the segmentation (Eq. 3). The recognition was performed as follows: ∀ c , ∃ s j ∈ 0 , R , s = ∑ i = r R i ∗ B i , j (3) where j is the column index, R is the number of rows in the US image, and B is the bone probability map described in previous fast automatic bone surface segmentation [25]. After generating the bone probability map, the segmentation can then be obtained as follows: S e g B r , c = 1   i f   r = s c 0   e l s e (4)

Detected pixels were then segmented based on a Cost function that was recognized as a “jump” in Eq. 5. J i = c o s t i , j ≥ n 0       < n (5) S i + 1 = J i + 1 ,   i = 1,2 , … k

In Eq. 5, J(i) indicates a specific column in a US image; it carries a higher Cost (i, j) than the user-determined value of n and is considered as a “jump.” Meanwhile, S(i) is the number of segments determined by the number of “jumps.” Our study used 3.4 mm, the minimum AHD reported previously [26], as a standard to assess the value n. Therefore, spaces between two pixels that were higher than this number indicated the existence of a jump between two heterogeneous segments. On the other hand, the algorithm did not detect bone surfaces if Cost (i, j) was consistently lower than value n, which occurred primarily in cases without hypoechogenic bone features in the US images. A “jump” defined two segments that were located anterior and superior to the “jump.” When multiple “jumps” were detected, additional features could be determined between jump points in addition to the two segments. Assuming that bone surfaces carried a more significant number of pixels, the algorithm excluded segments composed of lower numbers of pixels. It retained two details bearing the first and second highest number of pixels. These two remaining segments were considered as the acromion and humeral head. Once all unwanted components were eliminated, AHD was calculated by selecting two points, the pixel’s most inferior in the acromion and the most superior in the humeral head. After selecting points in each segment, the point-to-point distance was defined as the AHD.

Sensitivity analyses were conducted by excluding low-quality snapshots containing indistinguishable acromion or humeral head in the US images, examined by the experienced examiner. The agreement and reliability were calculated again. If there were no significant changes, the outcome of the computer-aided algorithm was reliable and insensitive to the lower-quality images. For the validation of the proposed computer-aided segmentation method, we compared the AHD measured by our algorithm to that from manual measurement. We compared the processing time between both methods to determine the quantitative performance evaluation. A licensed physical therapist (AS), blinded to the study participants and their shoulder positions, reviewed handpicked sets of images (n = 116) and manually analyzed the AHD. The intra-measurement reliability between the gold standard and computer-aided analyses of AHD was examined using the intraclass correlation coefficient (ICC), a two-way random method with an average of two measurements. The following ICC scale was used for the evaluation: almost perfect (0.81–1.00), excellent (0.61–0.80), moderate (0.41–0.60), and poor to fair (below 0.40) [27]. Correlations between the two methods of measurement were assessed using a Bland-Altman plot. The confidence interval within 95% of the mean differences between the two methods was evaluated. We further validated our segmentation method by evaluating the percentage of a surface fitting between computer-aided and manually segmented bone pixels. To elaborate clinical tolerance (0.5 mm), we extended the surface of manually segmented bone areas by ± 7 pixels.

The characteristics of manual wheelchair users are summarized in Table 1. A total of 116 measurements were included in the analysis of AHD measurements. The developed CAQUSA demonstrated 95% true positive with 5% false positive when fixed parameters of the threshold and the 4th power of shadowiness were used to achieve the detection rate. Figure 3 demonstrates the bone surface segmentation determined by manual selection and CAQUSA, visualizing the acromion and humeral head morphology in US images from one of the SCI participants.

In Figure 3, we compared bone surfaces determined by two methods. Our evaluation indicated that both the humeral head and acromion showed great approximation errors of 95.9% ± 5.81% and 96.1% ± 5.81%, respectively.

The mean and standard deviation of manual and CAQUSA measurements were 11.95 ± 2.29 mm and 11.33 ± 2.48 mm, respectively. The ICC between the two measures showed excellent intra-rater reliability (ICC = 0.953). In Figure 4A, AHD determined by CAQUSA showed a positive linear association with AHD analyzed by manual detection (r ² = 0.83). Bland-Altman plots were analyzed with reasonable limits of agreement and a mean difference close to zero, indicating similar measurements (Figure 4B). However, one size was over 0.27 mm, and five were below −0.14 mm between the CAQUSA and manual measurements, which suggests that the CAQUSA method achieved excellent sensitivity. Compared to manual measurements, the CAQUSA outside of a confidence interval showed a tendency to underestimate due to the heterogeneity of bone surface reflection in US images that were more hyperechogenic than a bony surface during humeral active elevation.