Depthwise Separable Convolutional Neural Network for Knee Segmentation: Data from the Osteoarthritis Initiative

Abstract

Automated knee segmentation plays an important role in knee osteoarthritis diagnosis as this disease exhibits different imaging biomarkers as it progresses. A good knee segmentation model that is practical and computationally efficient allows a more efficient clinical workflow. This paper presents a preliminary study on Depthwise Separable convolutional layers utilizing the end-to-end segmentation network, UNet architecture on knee segmentation. Results showed that DS2D-UNet and DS3D-UNet perform more efficiently with the adoption of Depthwise Separable convolutional layers with fewer cost of computations, without compromising the overall performance. The models produced strong results of Balanced Accuracy ranging between 90–93% and Dice Similarity Coefficient ranging between 91–93%. In conclusion, the potential of Depthwise Separable convolution should be further investigated to optimize the efficiency of 3D deep learning architectures, specifically on knee imaging volumes.