Improving automated ultrasound infant hip screening using an integrated clinical classification loss

Angle measurements in medical imaging tasks primarily rely on landmark localisation. Developmental dysplasia of the hip (DDH) is a disorder amongst infants where the hip does not form properly such that the femoral head is poorly located within the socket. The Graf method is an ultrasound-based screening technique in which lines are drawn through anatomical landmarks to characterise the deformity. The most important feature in the Graf method is an angle termed ‘alpha’, which is used to define severity and guide treatment decisions. In the Graf method, ultrasound images are annotated by clinicians to determine the optimal treatment. However, the subjective nature of ultrasound image interpretation results in significant intra– and inter– observer variability in measuring alpha. Deep learning has shown reasonable performance in predicting the Graf class compared to clinicians. However, these automated methods lack the evaluation of geometric criteria (landmark detection and angle measurements) and classification metrics. Until now, no work has evaluated the effect of incorporating the clinical classification within the loss function. To ensure the clinical adoption of automated methods, it is important to replicate the clinical pipeline. This paper shows improved performance by adding a weighted class into the loss function for most metrics. This work illustrates the importance of considering all metrics in the clinical pipeline to determine the best methods. The developed methods for evaluating geometric criteria can be applied to other angle-based medical imaging classification tasks.