34th Annual Scientific Meeting proceedings

Objectives:

Plain radiography for diagnosis of HIF is poorly sensitive, however CT is not universally accessible. This study aimed to develop a machine learning model to diagnose HIF from plain radiographs of the canine elbow, and to compare model performance with human radiographic interpretation.

Methods:

A retrospective matched case-control study was conducted. Inclusion criteria were (1) having HIF confirmed by CT, and (2) having corresponding plain elbow radiographs. A single observer carried out blinded evaluation of craniocaudal radiographs, and classified each elbow as HIF-positive or HIF-negative. Radiographs were preprocessed for machine learning and trained a Swin Transformer V2 based model in an 80:10:10 training, validation, and test split. Test performance was compared with McNemar’s test.

Results:

The case group consisted of 52 HIF-positive dogs (70 elbows) and a population-matched control group of 62 HIF-negative dogs (92 elbows) was selected. Spaniels and Spaniel crosses made up 89% of the sample population. Breed distribution (p=0.939646), age distribution (p=0.20054), sex distribution (p=0.142518), and weight distribution (p=0.451816) between case and control groups was not significantly different. Human radiographic interpretation was 74.2% accurate, 53.4% sensitive, and 92.4% specific. The AutoML model was 93.3% accurate, 89.5% sensitive, and 96.2% specific, which was significantly different to that of the human (p<0.001).

Conclusions:

Machine learning is highly sensitive and specific in diagnosing HIF from plain elbow radiographs. It outperforms human radiographic interpretation and may be a clinically useful diagnostic aid or screening tool.