Adaptation and Evaluation of Deep Learning Techniques for Skin Segmentation on Novel Abdominal Dataset

Abstract

Skin segmentation plays an important role in a wide variety of biomedical image processing applications, such as skin cancer identification, skin lesion detection, and wound isolation. However, contemporary research has been mainly based on facial and hand skin datasets, with no other body regions considered for skin pixels sampling. Segmenting skin specifically in the abdominal region can aid in robotic abdominal surgeries and treatment procedures, such as robot-assisted laparoscopic surgeries and abdominal ultrasounds. A robust and highly accurate abdominal skin detection technique thus becomes imperative. To this end, we compiled a novel dataset of 1,400 segmented abdominal pictures and adapted and compared four abdominal skin segmentation techniques: one based on thresholding and three deep learning techniques, namely a fully connected neural network for pixel-level classification, and two convolution-based networks, U-Net and Mask-RCNN. We show that the U-Net model outperforms the other segmentation techniques, resulting in a pixel-to-pixel mean cross-validation accuracy of 95.51% on our Abdominal dataset. The incorporation of the Abdominal dataset in the training helped improve the abdominal skin segmentation accuracy by 10.19%. The U-Net model proved to be computationally the fastest, enabling real time skin segmentation with a processing rate of 37 frames per second.