Pancreatic Segmentation via Ringed Residual U-Net

Computer-aided diagnosis (CAD) has a wide range of clinical applications, and medical image segmentation is essential in CAD. In medical image segmentation, due to the high anatomical variability of the pancreas and the weak contrast of the environment, the segmentation of the pancreas has always been the most challenging task. This study proposes a novel segmentation method for pancreatic segmentation. On the basis of complete convolution, an attentional mechanism is added to enhance the information exchange between downsampling and upsampling. By using the ring residual module, the proposed segmentation method can yield satisfactory results via deep convolution and can consolidate the characteristics of traditional deep learning networks. At the same time, compared to previous methods that used the dice coefficient (DICE) as a loss function, this study proposes a new loss function in the proposed method. This new loss function focuses not only on the area coincidence degree but also on the focus shape similarity. In the present work, ten-fold cross-validation computed tomography (CT) data (82 samples) from the NIH public pancreas dataset was conducted. The average dice similarity coefficient (DSC) of the results is 88.32 +/- 2.84, which is higher than the most advanced available methods and corresponds to higher robustness. Therefore, in practical applications, these methods can be used to provide more reliable auxiliary diagnostic data in the application of clinical medicine.