Click here to


Are you sure ?

Yes, do it No, cancel

Utilizing the Clique Atrous Spatial Pyramid Pooling for Pancreas Segmentation

M Yang1 , X Qi2 , S Tan1 (1) Huazhong University of Science and Technology, Wuhan, China,(2) UCLA School of Medicine, Los Angeles, CA


(Wednesday, 7/17/2019) 10:30 AM - 11:00 AM

Room: Exhibit Hall | Forum 5

Purpose: Accurate pancreas segmentation is challenging due to its significant shape differences of anatomical structures across patients and atypical pancreas shapes. Built upon Atrous convolution, the pyramid pooling module, which empirically proves to be an effective way to generate multi-scale features, was proposed to encode information using parallel convolution with different dilation rates.

Methods: ASPP(Atrous Spatial Pyramid Pooling) was introduced in the literature to handle objects in semantic segmentation with very different sizes. Motivated by CliqueNet, we proposed a model named CliqueASPP (Clique Atrous Spatial Pyramid Pooling) module. There are both forward and backward connections between any two layers. In Stage-I, we put layers with small dilation rates in lower part, while put layers with large dilation rates in upper part. The output of each atrous layer is concatenated to feed into the following layers. In Stage-II, we concatenate newly updated features to re-update previously updated layer, and reuse the convolution filters with same dilation rates as Stage-I. This recurrent feedback structure not only stacks all dilated layers together (including the features generated by Stage-I and Stage-II) to encode multi-scale information, but also bring higher level visual information back to refine low-level filters and achieve spatial attention.

Results: The proposed method achieved improvements over methods without CliqueASPP module. For a dataset with 12 pancreas volumes, the averaged Dice is 0.678 while the method without CliqueASPP module is 0.675.

Conclusion: Combining Atrous spatial pooling pyramid and Cliquenet’s recurrent feedback structure, the proposed method can generate multi-scale features to remedy the various different sizes in pancreas segmentation to achieve better performance.

Funding Support, Disclosures, and Conflict of Interest: S. Tan and M. Yang were supported in part by the National Natural Science Foundation of China, under Grant Nos. 61375018 and 61672253


Segmentation, CT


IM/TH- image segmentation: CT

Contact Email