Detection and local histological staging of prostate cancer foci in H&E whole slide images using convolutional neural networks

A pathologist’s histological diagnosis is the gold standard of prostate cancer diagnostic measures. In addition to cancer grade, the stage of cancer determines the follow-up and possible adjuvant therapies after surgery. Not only glandular pattern recognition but also the assessment of three-dimensional extent and size of cancer, in relation to the whole organ, are parts of the subjective diagnosis, which may vary among pathologists. Furthermore, evaluation of lymph nodes for any possible metastases is time-consuming and missing cancer in lymph nodes can lead to undertreatment of the patient. For the abovementioned reasons, our team is developing an algorithm that calculates the amount of cancer tissue objectively, assisting the pathologist in the diagnostic procedure. In addition, we are creating an assistant for cancer detection in lymph nodes, which are removed for histological evaluation. All in all, we aim to create a tool that will help pathologists to a better, faster, more secure and more accurate diagnosis. Our material consists of full sets of scanned whole slide images from 302 prostates that were retrieved by radical prostatectomy. After supervised learning procedure, convolutional neural networks were employed for the classification of cancerous and non-cancerous regions in the images. A tiling based approach was used in which a slide was divided into square shaped small tiles. Millions of cancerous and non-cancerous tiles were sampled from the dataset for training and validation. During the convolutional neural network training, several different tile sizes were used, i.e., 256x256, 512x512, 1024x1024 pixels. Four different types of architectures were fine-tuned and trained for the task of tile-wise binary classification, namely InceptionV3, Xception, ResNet50 and a custom convolutional neural network architecture. In our preliminary assays for cancer detection, in both pixel-wise and tile-wise evaluation, InceptionV3 performed outstandingly well with an AUC score of 0.97 and 0.951, respectively. In conclusion, our algorithm has developed very well thus far with an accuracy in cancer detection of 97%. It is not only a versatile assisting tool, aiding pathologists to a more objective, standardized and accurate diagnosis, but also serves as a second opinion in difficult and challenging diagnostic cases. Citation Format: Carolin Sturenberg, Umair Khan, Kevin Sandeman, Oguzhan Gencoglu, Adrian Malen, Andrew Erickson, Timo Heikkinen, Antti Rannikko, Tuomas Mirtti. Detection and local histological staging of prostate cancer foci in HE 2019 Mar 29-Apr 3; Atlanta, GA. Philadelphia (PA): AACR; Cancer Res 2019;79(13 Suppl):Abstract nr 1396.