Abstract

Developing a Histopathological Image-Based Tool for Accurate Prognosis of Pancreatic Adenocarcinoma by Implementing Deep Structured Learning

     Pancreatic cancer is one of the deadliest cancers, with a five-year survival rate of 7%, due to its unpredictable metastasis. A major problem is the difficulty in reaching an accurate prognosis, because tumor proliferation and metastasis biomarkers are difficult to find and track due to the complexity of the development of pancreatic cancers. Recent developments in whole slide imaging (WSI) technology allow for the scanning of biopsy samples into multi-resolution images that can be analyzed more accurately than in a microscope.

     Pancreatic adenocarcinoma WSIs were stain and color-corrected. The first method required the partitioning of the WSI into tiles, assigning a feature number based on malignancy, creating as feature map, and using the data to train machine learning algorithms. The next method included separating the WSI into different textures, assigning pathologist-defined malignancy values, creating a feature map, and training the algorithms on the dataset. The final method involved adding surrounding tiles to measure spatial heterogeneity, a key biomarker. This third method was compounded with the previous two to achieve peak AUC. The datasets were 8-fold cross-validated with the dataset randomly partitioned over 100 times. The tiles-only method was more accurate than a conventional human pathologist, and some of the networks performed better with texture information. When spatial heterogeneity information was added, the AUC varied per network, but the highest AUC was achieved. The AUC for the networks was significantly higher than that of a human pathologist in all cases in accurately predicting the proliferation and metastasis based on the WSIs.