A Unified and Semantic Model Approach for Histopathologic Cancer Detection Based on Deep Double Transfer Learning

Abstract

Accurately predicting the risk of cancer recurrence and metastasis is very important for individual cancer treatment. Currently, doctors usually use a histological grade that pathologists determine by performing a semi-quantitative analysis of the three histopathological and cytological features of hematoxylin-eosin (HE) stained histopathological images. Evaluate the prognosis and treatment options of patients with breast cancer. In order to efficiently and objectively fully utilize the valuable information underlying HE-stained histopathological images, this work has potential as a feature for constructing a classification model of cancer prognosis. So, a calculation method is proposed to extract morphological information. Breast cancer is not a single disease, but it is composed of many different biological entities with different pathological features and clinical significance. With the advent of personalized medicine, pathologists are facing a significant increase in the workload and complexity of digital pathology in cancer diagnosis, and diagnostic protocols need to focus on equal efficiency and accuracy. Computer-aided image processing techniques have been shown to be able to improve the efficiency, accuracy, and consistency of histopathological assessments and provide decision support to ensure diagnostic consistency. First, a method for segmenting tumor lesions based on a pixel-by-pixel deep learning classifier is proposed and a method for segmenting cell nuclei based on marker-driven watersheds. It then subdivides all image objects and extracts a rich set of predefined quantitative morphological object feature. Then a classification model based on these measurements is used to predict disease-free survival in binary patients. Finally, the predictive model is tested in two independent cohorts of breast cancer patients.