A Fixed-Point Model for Pancreas Segmentation in Abdominal CT Scans

Abstract

Abstract. Deep neural networks have been widely adopted for automatic organ segmentation from abdominal CT scans. However, the segmentation accuracy of some small organs (e.g., the pancreas) is sometimes below satisfaction, arguably because deep networks are easily disrupted by the complex and variable background regions which occupies a large fraction of the input volume. In this paper, we formulate this problem into a fixed-point model which uses a predicted segmentation mask to shrink the input region. This is motivated by the fact that a smaller input region often leads to more accurate segmentation. In the training process, we use the ground-truth annotation to generate accurate input regions and optimize network weights. On the testing stage, we fix the network parameters and update the segmentation results in an iterative manner. We evaluate our approach on the NIH pancreas segmentation dataset, and outperform the state-of-the-art by more than 4%, measured by the average Dice-Sørensen Coefficient (DSC). In addition, we report 62.43% DSC in the worst case, which guarantees the reliability of our approach in clinical applications.

Deep Supervision for Pancreatic Cyst Segmentation in Abdominal CT Scans

“…We first investigate pathological pancreas segmentation which serves as the first stage of our approach. With the baseline approach described in [17], we obtain an average DSC of 79.23 ± 9.72%. Please note that this number is lower than 82.37 ± 5.68%, which was reported by the same approach in the NIH pancreas segmentation dataset with 82 healthy samples.…”

“…The implementation details follow our recent work [17], which achieves the state-of-the-art performance in the NIH pancreas segmentation dataset [13]. Due to the limited amount of training data, instead of applying 3D networks directly, we cut each 3D volume into a series of 2D pieces, and feed them into a fully-convolutional network (FCN) [10].…”

“…, 60K} iterations, are used to test each volume, and the final result is obtained by computing the union of nine predicted foreground voxel sets. Regarding other technical details, we simply follow [17], including using the DSC loss layer instead of the voxel-wise loss layer to prevent the bias towards background [11], and applying a flexible convergence condition for fine-tuning at the testing stage.…”

“…This paper focuses on segmenting pancreatic cyst from abdominal CT scan. Pancreas is one of the abdominal organs that are very difficult to be segmented even in the healthy cases [13][12] [17], mainly due to the low contrast in the boundary and the high variability in its geometric properties. In the pathological cases, the difference in the pancreatic cancer stage also impacts both the morphology of the pancreas and the cyst [4] [8].…”

“…Using a reasonable transformation function, the optimization process can be factorized into two stages, in which we first find the pancreas, and then localize and segment the cyst based on the predicted pancreas mask. Our approach works efficiently based on a recent published coarse-to-fine segmentation approach [17]. We perform experiments on a newly collected dataset with 131 pathological samples from CT scan.…”