Pancreas and pancreatic tumor in CT scan.

Workflow for training and testing.

Pancreatic cancer (PC) is one of the most difficult cancers to diagnose and the most lethal cancer. Computed tomography (CT) is the major diagnostic imaging tool for PC. A computer-aided diagnosis tool that can rapidly detect and classify pancreatic lesions on CT can reduce the workload of radiologists and facilitate early detection of PC. We developed two models using different methods, deep learning and radiomics, to differentiate PC from normal pancreas (NP) on CT. We are also studying on combining the two algorithms.

  1. Deep Learning
    By building a convolution neural network and training with GPUs, we can classify whether a 2D patch of pancreas on CT contains cancer or not with a patch-based accuracy of 85.5%. After combining prediction from patches, the accuracy in patients-based can reach 97.2%.
  2. Radiomics
    We extracted radiomic features of each patch and put these radiomic features into XGBoost model to perform classification. Through this work, we can accurately differentiate between cancerous and non-cancerous pancreatic masses with 91.8% of accuracy.

After integration into daily practice in PACS system, the model can provide for the possibility of cancerous mass when the clinician chooses an ROI. This model may be useful to assist radiologists in the clinical setting.