Classification of Histopathological Images in Automatic Detection of Breast Cancer with Deep Learning Approach

Abstract

Convolutional neural networks have emerged as an essential tool for image classification
and object detection. In the health field, these tools are a crucial factor in saving time and
minimizing the margin of error for the health system and employees. Breast cancer is the
most common type of cancer in women worldwide. In many cases, it can threaten human
life, resulting in death. Although methods have been developed for the early diagnosis of
this health problem, its support with digital systems remains incomplete. In diagnosis,
histopathological images are examined with microscope methods. In cases where the
number of pathologies is insufficient, delay problems may occur and the error rate
increases in manual controls. The study aims to design a deep-learning object detection
method for the pre-detection of breast cancer. The publicly published BreaKHis dataset
is used as the dataset. Model results that generated with VGG16, InceptionV3 and
ResNet50 deep learning architectures have been compared. The highest accuracy rate
have been obtained with the proposed model as 85%. Accuracy, AUC, precision, recall,
F-score performance metrics have been analysed for each model. A decision support
system screen design has been created using the proposed model weight file. With the
study, the computer-assisted clinical support system makes clinicians' life more
manageable and recommends early diagnosis.