Connected-SegNets: A Deep Learning Model for Breast Tumor Segmentation from X-ray Images

Abstract

Simple Summary The segmentation of breast tumors is an important step in identifying and classifying benign and malignant tumors in X-ray images. Mammography screening has proven to be an effective tool for breast cancer diagnosis. However, the inspection of breast mammograms for early-stage cancer can be a challenging task due to the complicated structure of dense breasts. Several deep learning models have been proposed to overcome this particular issue; however, the false positive and false negative rates are still high. Hence, this study introduced a deep learning model, called Connected-SegNets, that combines two SegNet architectures with skip connections to provide a robust model to reduce false positive and false negative rates for breast tumor segmentation from mammograms. Inspired by Connected-UNets, this study proposes a deep learning model, called Connected-SegNets, for breast tumor segmentation from X-ray images. In the proposed model, two SegNet architectures are connected with skip connections between their layers. Moreover, the cross-entropy loss function of the original SegNet has been replaced by the intersection over union (IoU) loss function in order to make the proposed model more robust against noise during the training process. As part of data preprocessing, a histogram equalization technique, called contrast limit adapt histogram equalization (CLAHE), is applied to all datasets to enhance the compressed regions and smooth the distribution of the pixels. Additionally, two image augmentation methods, namely rotation and flipping, are used to increase the amount of training data and to prevent overfitting. The proposed model has been evaluated on two publicly available datasets, specifically INbreast and the curated breast imaging subset of digital database for screening mammography (CBIS-DDSM). The proposed model has also been evaluated using a private dataset obtained from Cheng Hsin General Hospital in Taiwan. The experimental results show that the proposed Connected-SegNets model outperforms the state-of-the-art methods in terms of Dice score and IoU score. The proposed Connected-SegNets produces a maximum Dice score of 96.34% on the INbreast dataset, 92.86% on the CBIS-DDSM dataset, and 92.25% on the private dataset. Furthermore, the experimental results show that the proposed model achieves the highest IoU score of 91.21%, 87.34%, and 83.71% on INbreast, CBIS-DDSM, and the private dataset, respectively.