Ultrasound Image Segmentation: A Deeply Supervised Network With Attention to Boundaries

OBJECTIVE: Segmentation of anatomical structures in ultrasound images requires vast radiological knowledge and experience. Moreover, the manual segmentation often results in subjective variations, therefore, an automatic segmentation is desirable. We aim to develop a fully convolutional neural network (FCNN) with attentional deep supervision for the automatic and accurate segmentation of the ultrasound images. METHOD: FCNN/CNNs are used to infer high-level context using low-level image features. In this paper, a sub-problem specific deep supervision of the FCNN is performed. The attention of fine resolution layers is steered to learn object boundary definitions using auxiliary losses, whereas coarse resolution layers are trained to discriminate object regions from the background. Furthermore, a customized scheme for downweighting the auxiliary losses and a trainable fusion layer are introduced. This produces an accurate segmentation and helps in dealing with the broken boundaries, usually found in the ultrasound images. RESULTS: The proposed network is first tested for blood vessel segmentation in liver images. It results in F1 score, mean intersection over union, and dice index of 0.83, 0.83, and 0.79, respectively. The best values observed among the existing approaches are produced by U-net as 0.74, 0.81, and 0.75, respectively. The proposed network also results in dice index value of 0.91 in the lumen segmentation experiments on MICCAI 2011 IVUS challenge dataset, which is near to the provided reference value of 0.93. Furthermore, the improvements similar to vessel segmentation experiments are also observed in the experiment performed to segment lesions. CONCLUSION: Deep supervision of the network based on the input-output characteristics of the layers results in improvement in overall segmentation accuracy. SIGNIFICANCE: Sub-problem specific deep supervision for ultrasound image segmentation is the main contribution of this paper. Currently the network is trained and tested for fixed size inputs. It requires image resizing and limits the performance in small size images.