Deep Learning for Detection of Iso-attenuating Pancreatic Adenocarcinoma in Computed Tomography

M. Schuurmans, N. Alves, H. Huisman and J. Hermans

Annual Meeting of the Radiological Society of North America 2022.

Purpose: Investigate the ability of deep learning to localize iso-attenuating pancreatic adenocacinoma (pCa) in computed tomography (CT) images and assess the effect of including prior anatomy information on localization performance. This retrospective study included contrast-enhanced CT scans in the portal venous phase of 44 patients from the public Medical Segmentation Decathlon dataset, who had a visually iso-attenuating tumour. Two previously developed deep learning (DL) algorithms based on the 3D nnUnet framework were applied to the set of iso-attenuating lesions: one considering only tumour information for training (nnUnet_T) and one considering surrounding anatomical structures, namely the pancreas parenchyma, common bile duct, pancreatic duct, arteries and veins in additional to the tumor (nnUNet_MS). Each model creates 10 different outputs, based on 2 random initialisations and 5-fold crossvalidation. The performance of the two models at pCa localization was evaluated with average precision and a permutation test was performed to show statistical significance. Results: The average precision for the nnUNet_MS was 80.03 % ± 8.6%, while for the nnUnet_T model it was 72.87% ± 7.1% (p<0.05). This indicates that surrounding anatomy aids the localization of iso-attenuating pCa lesions. By having access to prior anatomy information during training the network can better focus on the regions of the image where the tumor is located, avoiding confusion with other anatomical structures. Conclusion: In current clinical workup the detection of iso-attenuating pCa lesions on CT is very challenging, as the attenuation is similar to the pancreas parenchyma. The proposed 3D nnUnet_MS model can be used to detect and localize iso-attenuating lesions. Clinical relevance statement: Iso-attenuating pCa is linked to earlier stages of disease and better outcome, but is challenging to detect on CT. DL can accurately detect iso-attenuating lesions and benefits from anatomy information.