Invasive Ductal Carcinoma Grade Classification in Histopathological Images using Transfer Learning Approach

Abstract

Cancer is the greatest cause of mortality in the world wherein; Breast Invasive Ductal Carcinoma (IDC) is the second leading cause of death among women. Computer-assisted diagnostic (CAD) systems embedded with advanced artificial intelligence techniques have demonstrated a significant performance in early detection on a variety of malignancies. Digital histopathological images and deep learning architectures are the preferred approaches in the analysis of biomedical images. In this conjunction, pre-existing deeper networks are trained on large image datasets and adapted for the classification of grades using the transfer learning approach. In this study, a transfer learning approach is proposed for the classification of IDC grade breast. Herein, we have utilized deeper pre-trained models, VGG16, Inception_V3, ResNet152V2, DenseNet201, and NASNetMobile for the feature extraction tasks, and Random Forest (RF) classifier is applied for the final decision on IDC grade classification. A new histopathological microscopy databiox image dataset, containing 922 images from 124 IDC patients that are publicly available is used in this work. The proposed transfer learning approach attained the best performance using DensNet201 (AUC=98%) and VGG16 (AUC=83%) for grade class 0 (Grade-I), DensNet201 (AUC=75%) and NASNetMobile (AUC=72%) for grade class 1 (Grade-II) and Inception_V3 (AUC=65%) and NASNetMobile (AUC=69%) for grade class 2 (Grade-III) among the different pre-trained models. The designed model achieved the highest performance as compared to the existing state-of-the-art techniques.

Topics & Concepts