Deep ensemble of texture maps for false positive reduction in mammograms

Abstract

Worldwide, breast cancer is a life-threatening disease attributing to increased mortality rates among women. Mammograms are commonly used for screening breast cancer in asymptomatic stages. However, the subtle nature of abnormalities in early stages makes mammogram analysis a cumbersome task. A computer aided diagnosis (CAD) system can complement subjective diagnosis of physicians with its objective assessment. Mass detection is the most important task in breast cancer diagnosis, as masses are the prominent indicators of the disease. Nevertheless, it is the most challenging task due to the ambiguity between masses and the surrounding normal tissues, especially in dense breasts. Though CAD systems are effective in detecting masses with high sensitivity, the price paid is usually high false positive rates (FPR). Texture analysis is normally employed to reduce the FPR in mass detection, where texture features extracted from suspicious regions are used to build a classifier model to discriminate between actual masses and false positives. Deep learning (DL) is a data-driven model that is gaining increased importance in diverse fields, including medical diagnosis, that involve voluminous amounts of data. In particular, convolutional neural network (CNN) plays an important role in image analysis in various applications, including mammogram analysis. Converting raw images to texture maps can enhance the performance of CNN for false positive reduction. In this work, textural image maps based on Hilbert curve, forest fire model, Radon transform, discrete wavelet transform (DWT) and curvelet transform are analysed using CNN. More specifically, an ensemble of CNNs based on these individual textural image representations is constructed. The proposed work is validated on CBIS-DDSM, a publicly available benchmark dataset, demonstrating 100% accuracy for mass detection with 0% FPR.