Automatic Nuclear Atypia Scoring of Breast Cancer Pathological Images Based on Deep Residual Network and Meta-decision Tree
Xuedong Zhu, Xi Lu · 2022
Breast cancer has been the most commonly occurred cancer worldwide, and the patient survival is closely related to correct diagnosis and early treatment. Nuclear atypia is the only qualitative index for histological grading of invasive breast cancer, with the lowest consistency of manual diagnosis. In this paper, a model based on deep learning and aggregation strategy is proposed to automatically score the nuclear atypia of breast tumors, in order to provide an efficient and accurate auxiliary diagnosis basis for pathologists. We employed the public dataset ICPR2014 ATYPIA. Base classifiers based on deep residual network were trained separately on the images of each patient at three magnifications (×10, ×20, ×40), and then utilized the meta-decision tree method to combine the outputs of the three classifiers to obtain the diagnosis of breast tumor nuclear atypia for each patient. The classification accuracy of our proposed model reaches 0.8800, and the sensitivity reaches 0.8645. Compared with other algorithms, it is highly competitive in terms of accuracy and efficiency, so it has great potential to be applied in clinical work.