Fully Automatic Deep Learning Framework for Pancreatic Ductal Adenocarcinoma Detection on Computed Tomography

Abstract

Early detection improves prognosis in pancreatic ductal adenocarcinoma (PDAC), but is challenging as lesions are often small and poorly defined on contrast-enhanced computed tomography scans (CE-CT). Deep learning can facilitate PDAC diagnosis; however, current models still fail to identify small (<2 cm) lesions. In this study, state-of-the-art deep learning models were used to develop an automatic framework for PDAC detection, focusing on small lesions. Additionally, the impact of integrating the surrounding anatomy was investigated. CE-CT scans from a cohort of 119 pathology-proven PDAC patients and a cohort of 123 patients without PDAC were used to train a nnUnet for automatic lesion detection and segmentation (nnUnet_T). Two additional nnUnets were trained to investigate the impact of anatomy integration: (1) segmenting the pancreas and tumor (nnUnet_TP), and (2) segmenting the pancreas, tumor, and multiple surrounding anatomical structures (nnUnet_MS). An external, publicly available test set was used to compare the performance of the three networks. The nnUnet_MS achieved the best performance, with an area under the receiver operating characteristic curve of 0.91 for the whole test set and 0.88 for tumors <2 cm, showing that state-of-the-art deep learning can detect small PDAC and benefits from anatomy information.

Keywords: deep-learning; early detection; pancreatic ductal adenocarcinoma.

Figure 1

Schematic overview of the proposed automatic PDAC detection framework. The first step in the pipeline is to automatically extract the ROI from the full input CE-CT scan, using the low-resolution pancreas segmentation network. This ROI is then fed to each of the PDAC detection networks: nnUnet_T, nnUnet_TP, and nnUnet_MS. The final tumor likelihood output is derived from the networks’ tumor detection likelihood maps, which in the case of the nnUnet_TP and nnUnet_MS models, are post-processed using the automatically generated pancreas segmentation.

Figure 2

Mean ROC and FROC curves with respective confidence intervals for the external test set.

Figure 3

Mean ROC and FROC curves with respective confidence intervals for the external set considering only the subgroup of tumors < 2 cm in size.

Figure 4

Example of an iso-attenuating tumor from the external test set, which was missed by both the nnUnet_T and nnUnet_TP, but could be correctly localized by nnUnet_MS. (A) Slice of the original ROI input; (B) ground truth segmentation of tumor and pancreas; (C) output of nnUnet_TP, which in this case is only the pancreas segmentation as the tumor is not detected; and (D) output of the nnUnet_MS, which is the segmentation of the detected tumor and surrounding anatomy.