Transfer learning with different modified convolutional neural network models for classifying digital mammograms utilizing Local Dataset.

Abstract

Background: Breast cancer is the leading cause of cancer-related mortality among women worldwide. The incidence and mortality increased globally since starting registration in 1990. Artificial intelligence is being widely experimented in aiding in breast cancer detection, radiologically or cytologically. It has a beneficial role in classification when used alone or combined with radiologist evaluation. The objectives of this study are to evaluate the performance and accuracy of different machine learning algorithms in diagnostic mammograms using a local four-field digital mammogram dataset.

Methodology: The dataset of the mammograms was fullfield digital mammography collected from the oncology teaching hospital in Baghdad. All the mammograms of the patients were studied and labeled by an experienced radiologist. Dataset was composed of two views CranioCaudal (CC) and Mediolateral-oblique (MLO) of one or two breasts. The dataset included 383 cases that were classified based on their BIRADS grade. Image processing included filtering, contrast enhancement using contrast limited adaptive histogram equalization (CLAHE), then removal of labels and pectoral muscle for improving performance. Data augmentation was also applied including horizontal and vertical flipping and rotation within 90 degrees. The data set was divided into a training set and a testing set with a ratio 9:1. Transfer learning of many models trained on the Imagenet dataset was used with fine-tuning. The performance of various models was evaluated using metrics including Loss, Accuracy, and Area under the curve (AUC). Python v3.2 was used for analysis with the Keras library. Ethical approval was obtained by the ethical committee from the College of Medicine University of Baghdad Results: NASNetLarge model achieved the highest accuracy and area under curve 0.8475 and 0.8956 respectively. The least performance was achieved using DenseNet169 and InceptionResNetV2. With accuracy 0.72. The longest time spent for analyzing one hundred image was seven seconds.

Discussion and conclusion: This study presents a newly emerging strategy in diagnostic and screening mammography by using AI with the help of transferred learning and fine-tuning. Using these models can achieve acceptable performance in a very fast way which may reduce the workload burden among diagnostic and screening units.