Improving 3D Ultrasound Scan Adequacy Classification Using a Three-Slice Convolutional Neural Network Architecture

Accurately detecting and diagnosing developmental dysplasia of the hip (DDH), a common hip instability condition among infants, requires ultrasound (US) image data that captures the relevant anatomical structures. Our group has recently introduced a technique for automatically processing 3D US scans of the neonatal hip that significantly reduces dysplasia metric measurement variability, but it can be challenging for a clinician to know at the time of acquisition if they have acquired a volume suitable for analysis. We have previously introduced a single-slice-based approach for assessing adequacy of acquired volumes that worked reasonably well, but we believe that more explicitly three-dimensional approaches would be more robust and reliable. Here, we propose a new technique based on a convolutional neural network (CNN) architecture that incorporates inter-slice information and transfer learning. Our classifier labels volumes as adequate or inadequate for subsequent interpretation based on detecting the presence of key hip anatomical structures needed for DDH diagnosis. We validate our approach on 40 datasets from 15 pediatric patients and demonstrate a slice classification rate of 93% (improving on our previous implementation by 3%) with average processing time of 2 seconds per US volume. We expect automatic US scan adequacy assessment to have significant clinical impact with the potential to help in imaging standardization, improving efficiency of measuring DDH metrics, and improving accuracy of clinical decision making.