Comparison of Deep Learning and Classical Image Processing for Skin Segmentation
Skin stiffness correlates with the progression of sclerotic skin diseases. Ultrasound shear wave elasticity imaging techniques can measure skin stiffness, but an accurate skin thickness measurement is required to compute the elastic modulus. We explored different automated methods to segment the skin for use in real-time skin elastography. Local gradient-based methods could not robustly segment the skin on our B-mode images, so we developed a new thresholding method to detect the edges of the skin. We also used our thresholding method to generate labels to train a deep neural network. We compared the performance of thresholding and the trained network for central thickness estimation on a held-out test set. Our thresholding method correctly segmented 58% of images, and the neural network correctly segmented 82%. More than half of thresholding failures on the test set were from overestimation of the bottom skin boundary. The neural network had significantly less overestimation failures and similar rates of failure due to bubbles and underestimation.
