Pigs
CT scans, age and weight of boars were retrospectively extracted from boar test archives. The study was approved by the Ethical Committee of the Norwegian University of Life Sciences Faculty of Veterinary Medicine, reference number 14/04723–68.
Data analysis was completed for 201 boars. The pigs were balanced for four breeds: Landrace, Duroc, Synthetic A and B breeds. Included Landrace and Duroc boars were scanned between 2016 and 2020, whereas Synthetic A and B boars were scanned between 2018 and 2020. The four breed groups were treated as a single study population.
CT-scanning
Landrace and Duroc boars were scanned in a 32-slice CT-scanner (LightSpeed Pro32, GE Healthcare, Chicago, Illinois, USA) and Synthetic A and B boars were scanned in a 64-slice CT-scanner (Revolution Evo, GE Healthcare, Chicago, Illinois, USA). The scanning procedure has been described before [13], but briefly: boars were sedated and placed in sternal recumbency with free limb position. The field of view was collimated from the snout to whichever was most caudal of the tail or hocks of the pig, and scan parameters were 120 kV, dynamic mA up to 400, slice thickness of 1.25 mm and pitch 1.
Preparation of scans for labelling
Scans were prepared for labelling using MatLab (v. 9.9.0 [R2020b], The MathWorks Inc., Natick, Massachusetts, USA). The CT scan was treated as an array of ~ 1000 slices with 512 × 512 pixels; the total slice number varied somewhat with the length of the individual pig (Fig. 6a). An artificial intelligence method previously described by Kvam et al. [43] was used to segment the major limb bones. Segmentations were visually inspected and improved by manual adjustment of the two-dimensional masks where necessary.
New scan volumes were constructed around the humerus, femur (Fig. 6b) and tibia based on their image moment invariants [44], see Gangsei et al. [45] for details. The grid sizes for the new volumes were 256 × 128 × 128 voxels for the humerus, 300 × 150 × 150 voxels for the femur (Fig. 6c) and 300 × 128 × 128 voxels for the tibia. Because grid sizes were fixed, voxel resolution varied slightly with the scaling of individual bones, but resolution from the original scan remained associated with each new volume. The proximal and distal ends of the new volumes were used to extract volumes containing the shoulder, elbow, stifle (Fig. 6d) and hock joints for labelling. The new volumes contained transverse images that were perpendicular to the long axis of the humerus, femur (Fig. 6d) or tibia, irrespective of limb position in the original scan (Fig. 6a). The re-stacking of voxels results in some regular irregularities, apparent as stripes in the images, but these were readily distinguishable from focal lesions. The new volumes also contained two planes that were variably aligned with the frontal and sagittal planes, but always orthogonal to the transverse plane and each other. The x, y and z coordinates from the original CT scan remained associated with the new voxels, meaning applied labels could be transferred back to their location within the original CT scan (Fig. 6e).
Labelling application
An application for manual image labelling was developed using MatLab. The application showed all three image planes and had functions for adjusting window width/level, smoothing, scrolling, zooming and correlating image planes using crosshairs. The application also had tools for labelling lesions by clicking on individual pixels or drawing around larger areas to create so-called “masks”. Different colours were used to represent different categories of observations, described below. Once a pixel was labelled with a particular colour in one image plane, it immediately showed up as the same colour in the two orthogonal planes. Each labelled pixel within a mask could then be exported for further analyses as a unit containing information about its colour and x, y and z coordinates, i.e., a classified voxel, from which information about its relationship to neighbouring voxels could also be extracted.
Labelling strategy
Scans were labelled by a veterinary radiologist with 14 years of experience who was blinded to the identity of the pig. A strategy was pilot-tested where different colours represented categories of qualitative change, e.g., primary osteochondrosis defect, secondary marginal sclerosis etc., but it was difficult to label consistently when individual lesions required multiple colours, and the strategy generated few observations spread on many categories unsuitable for machine learning and was therefore abandoned.
Instead, a definitive strategy was adopted where only primary osteochondrosis defects were labelled as lesions (Figs. 1, 2, 3 and 4) [7], and where different colours represented different anatomical regions. Thirteen anatomical regions were used, comprising the glenoid cavity, humeral head, medial and lateral humeral condyles, medial and lateral halves of the proximal radius, medial and lateral trochlear ridges of the femur, medial and lateral condyles of the femur, distal intermediate ridge of the tibia and medial and lateral halves of the talus. Stifle joints were labelled first, followed by the shoulder, elbow and hock joints.
Parameters observed
Voxels were labelled as osteochondrosis lesions if they were located within focal, sharply demarcated and uniformly hypodense, single or multi-lobulated (“stair-step”) defects in or near the ossification front (Fig. 2c). Defects in the ossification front have previously been histologically validated to represent the osteochondrosis manifesta stage of the disease [7, 17], whereas spherical defects that are deeper than manifesta lesions have been validated to represent subchondral pseudo- or true bone cysts [7, 46]. It is not possible to distinguish between true and pseudocysts in CT scans, so such lesions will be referred to as “cyst-like”. Osteochondrosis latens and dissecans stages were not labelled, thus there was no risk of confusion and the labelled osteochondrosis manifesta lesions will be referred to simply as osteochondrosis lesions.
The following changes were not labelled: secondary responses, physeal osteochondrosis lesions, hypodense channels or foci presumed to represent nutrient artery foraminae, synovial fossae or intertrochlear indentations based on their typical location in the centre, midline and at mid-height of epiphyses or the talus [17].
Lesion number
Lesion masks that were not connected at their faces, edges or corners were counted as separate lesions, as determined using the 26-connected version of a MatLab tool for analysing three-dimensional voxel connectivity (https://se.mathworks.com/help/images/pixel-connectivity.html). Only masks that contained four or more connected voxels were counted as lesions, to ensure that lesions spanned at least two voxels in one of the three image planes, thus reducing the risk of labelling normal contour irregularities as lesions.
Both single-lobe and multi-lobulated lesions were counted as one lesion.
Lesion size
Lesion size was defined as the volume of an osteochondrosis or cyst-like lesion that was outlined by bone. The superficial margin of osteochondrosis lesions could not be visualised because it bordered on iso-dense cartilage. By default, the labelled area was therefore extended until it was level with the adjacent, normal ossification front.
Clinical significance
Pigs that develop disease are removed from the boar test, so included pigs did not have clinical signs of osteochondrosis. An attempt was nevertheless made at suggesting the portion of lesions that might progress to clinical disease through comparison with critical defect size, defined as the size above which defects will not heal without intervention [33]. Studies to determine critical defect size are only available for miniature pigs [33]. Critical defect size was translated from miniature to the more regularly-sized, commercially reared pigs that enter the boar test as follows: Chang et al. [34] indicated that the diameter of critical-sized defects was 38.1% of the width of the medial femoral condyle. This width was measured to approximately 22,5 mm for the current included pigs, and 38.1% of that or ≈8.6 mm was used as the diameter for the translated cylindrical defect. Based on Ahern et al. [33], the depth of the defect was set to 6 mm, resulting in a total cylinder volume of 346.30 mm3. According to Ahern et al. [33], the portion of critical-sized defects that is located within bone and therefore detectable by CT is 59%, resulting in a final translated critical-sized bone defect of 204.32 mm3 for the medial femoral condyle of commercial pigs.
Agreement
Intra-observer agreement was informally tested by labelling all regions in 10 pigs two times up to 7 months apart (Supplemental Table 2). This generated 260 paired observations, with agreement on whether a region was positive or negative on 213 occasions (81.9%). Disagreement was handled by keeping the first labelling.
Correlations
Correlations were tested using Pearson’s correlation coefficient with the level of significance set at < 0.05.