LMHistNet: Levenberg–Marquardt Based Deep Neural Network for Classification of Breast Cancer Histopathological Images

Breast cancer is a deadly disease commonly affecting women. One method to avoid death from breast cancer is to obtain a diagnosis early. Breast cancer detection is a significant area that benefits from the technological advancements in artificial intelligence. A deep neural network architecture for the classification of microscopic images of breast tumor tissue acquired using excisional biopsy has been proposed. A hybrid convolutional neural network model with asymmetric convolutions and Levenberg-Marquardt optimization named as LMHistNet is used for the classification of breast cancer images into binary as well as eight subclasses. Convolution block attention module is incorporated for adaptive feature refining. Convergence time has been significantly reduced by normalizing the input features on models using batch normalization. Training is improved by reducing the internal covariance shift. Hinge loss function is used for better convergence. Magnification-dependent as well as independent binary and eight class classifications are efficiently performed by extracting diversified features from histopathological images. 7909 histopathological images of which 2480 were images of benign (normal) and the remaining images of malignant patients were used for the study. Further, performance evaluation on the dataset using LMHistNet has been analyzed based on the loss and accuracy curve. The experimental findings show that the proposed model obtained good performance over the various magnifications 40X, 100X, 200X and 400X. The accuracy, precision, recall and F1 score obtained are 88, 89, 88 and 88 respectively for multiclass classification into eight subtypes. The accuracy, precision, recall and f1 score are 99 each for binary classification of breast tumor tissues into benign and malignant classes.