A method for labelling lesions for machine learning and some new observations on osteochondrosis in computed tomographic scans of four pig joints

Abstract

Abstract Background Osteochondrosis is a major cause of leg weakness in pigs. Selection against osteochondrosis is currently based on manual scoring of computed tomographic (CT) scans for the presence of osteochondrosis manifesta lesions. It would be advantageous if osteochondrosis could be diagnosed automatically, through artificial intelligence methods using machine learning. The aim of this study was to describe a method for labelling articular osteochondrosis lesions in CT scans of four pig joints to guide development of future machine learning algorithms, and to report new observations made during the labelling process. The shoulder, elbow, stifle and hock joints were evaluated in CT scans of 201 pigs. Results Six thousand two hundred fifty osteochondrosis manifesta and cyst-like lesions were labelled in 201 pigs representing a total volume of 211,721.83 mm3. The per-joint prevalence of osteochondrosis ranged from 64.7% in the hock to 100% in the stifle joint. The lowest number of lesions was found in the hock joint at 208 lesions, and the highest number of lesions was found in the stifle joint at 4306 lesions. The mean volume per lesion ranged from 26.21 mm3 in the shoulder to 42.06 mm3 in the elbow joint. Pigs with the highest number of lesions had small lesions, whereas pigs with few lesions frequently had large lesions, that have the potential to become clinically significant. In the stifle joint, lesion number had a moderate negative correlation with mean lesion volume at r = − 0.54, p < 0.001. Conclusions The described labelling method is an important step towards developing a machine learning algorithm that will enable automated diagnosis of osteochondrosis manifesta and cyst-like lesions. Both lesion number and volume should be considered during breeding selection. The apparent inverse relationship between lesion number and volume warrants further investigation.

Keywords

• Articular osteochondrosis
• Breeding selection
• Helical computed tomography
• Lesion number
• Lesion volume
• Machine learning