PanNet: a feature-based attention aggregation model for segmenting pancreatic ductal adenocarcinoma on contrast-enhanced CT images of the abdomen

Pancreatic ductal adenocarcinoma (PDAC) is one of the most lethal forms of cancer with a 5-year survival rate of only 8%. One of the primary reasons for the low survival rate is the late detection of the PDAC cancer. This work focuses on the segmentation of PDAC for improved detection of pancreatic masses which will ultimately promote diagnosis during earlier stages of the disease. We propose a novel automatic segmentation model, PanNet which uses multi-level skip connection along with novel feature-based attention aggregation (FAA) block to improve the accuracy of the PDAC detection. The FAA block improves the interpolation power of the model in the decoder blocks, thereby reducing false positive pixels in the predicted tumor masks. The pixel-wise attention algorithm in the FAA block is applied across all channels in the 3D feature vectors obtained from individual decoder blocks. This leads to a substantial improvement of up to 7.3% in Dice Score (DSC) score on two datasets, each containing a test set of patients with early onset of PDAC. This aggregates to an improvement in the pixels of tumor volume prediction by at most 60.2% in comparison to state-ofthe-art (SOTA) pancreas segmentation models across both the datasets. The proposed PanNet can be utilized for early detection of PDAC cancer, given its consistent and enhanced segmentation performance demonstrated across multiple datasets in this paper.