Predictive Modeling for Breast Cancer Detection using Transfer Learning and Comparative Study of Some Existing Models

Abstract

In the present manuscript, a deep neural network based predictive model has been proposed for breast cancer detection in mammogram images. In this study transfer learning was used to six different pre-trained convolutional neural networks. These neural networks are further trained and fine-tuned to perform breast cancer classification task in order to detect malignant tumor. Results of this experiment suggest that ResNet50 is best suitable model than other five models used to detect tumor in mammogram images with 99.89% testing accuracy. An experimental study has also been performed to observe the accuracies of six models to identify malignant tumors. A class level accuracy of these models has been calculated and we observe that ResNet50 can classify both benign and malignant tumors with almost equal accuracies. Other five models classified benign samples effectively but failed to classify malignant tumors effectively. It is also observed that EfficientNet was the fastest model followed by ResNet50. The two main contributions of this paper is to achieve 99.89% testing accuracy by training used light weight model trained only for 25 epochs and use of whole images without any radiologists intervention. The limitation of this research is the use of a small custom dataset and use of generalized parameters and weights of IMAGENET1K_V1 jnisfor fine-tuning. The proposed light-weight limited layer transfer learning models can be used for further experimentation to improve efficiency of models by adjusting weights along with different segmentation techniques.

Keywords: Breast cancer, Classification, Deep learning, Fine tuning, Transfer learning.

View PDF