SqueezeU-Net-based detection and diagnosis of microcalcification in mammograms

Though rare in male, breast cancer is very frequent in women and is of fatal nature. Microcalcification (MC), an early indication of breast cancer, can reduce the mortality rate by many folds if detected and diagnosed at the earlier stage. But, due to its existence in discrete as well as in clustered form, its identification becomes a challenging task. Considering the recent advances and impact of deep learning techniques in biomedical imaging, in this paper, deep learning architectures, U-Net and its modified version: SqueezeU-Net, are proposed for the detection of MC followed by its characterization as benign or malignant. Being fully convolutional network, U-Net and SqueezeU-Net both are independent of dimensions of the input. SqueezeU-Net has less computational complexity while preserving the accuracy of the system due to delay downsampling. The assessment of the proposed model is performed on 1000 mammograms of DDSM dataset and the analysis is extended with data augmentation. A true positive rate (TPR) of 89.83% with 0.42 false positive per image (FPs/I) is obtained for SqueezeU-Net as compared to 84.67% at 0.5 FPs/I by U-Net architecture. The detected MC were classified and achieved an accuracy of 97.30% and area under the ROC curve (AUC) is 0.97 for SqueezeU-Net while the same for U-Net are 93.64% and 0.93 respectively using five-fold cross validation. The results are further compared with other competing schemes of the state-of-the-art where the proposed architectures take an edge over others.