Detecting nodular basal cell carcinoma in pathology imaging using deep learning image segmentation

摘要

With over 4.3 million new cases in the U.S. every year, basal cell carcinoma (BCC), is the most common form of skincancer. Pathologists must examine pathology images to diagnose BCC, potentially resulting in delay, error, andinconsistency. To address the need for standardized, expedited diagnosis, we created an automated diagnostic machine toidentify BCC given pathology images. In MATLAB, we adapted a deep neural network image segmentation model, UNet,to train on BCC images and their corresponding masks, which can learn to highlight these nodules in pathologyimages by outputting a computer-generated mask. We trained the U-Net on one image from the dataset and compared thecomputer-generated mask output from testing on three types of images: an image from a different region of the same imagetaken with the same microscope, an image from a different tissue sample with a different microscope, and an image takenwith a confocal microscope. We observed good, medium and poor results, respectively, illustrating that performancedepends on the similarity between test and training data. In subsequent tests using data augmentation, we achievedsensitivity of 0.82±0.07 and specificity of 0.87±0.16 on N = 6 sample sections from 3 different BCCs imaged with thesame microscope system. These data show that the U-Net performed well with a relatively few number of training images.Examining the errors raised interesting questions regarding what the errors mean and how they possibly arose. By creatinga surgeon interface for rapid pathological assessment and machine learning diagnostics for pathological features, the BCCdiagnosis process will be expedited and standardized.