SAP‐cGAN: Adversarial learning for breast mass segmentation in digital mammogram based on superpixel average pooling

Abstract

Purpose Breast mass segmentation is a prerequisite step in the use of computer‐aided tools designed for breast cancer diagnosis and treatment planning. However, mass segmentation remains challenging due to the low contrast, irregular shapes, and fuzzy boundaries of masses. In this work, we propose a mammography mass segmentation model for improving segmentation performance. Methods We propose a mammography mass segmentation model called SAP‐cGAN, which is based on an improved conditional generative adversarial network (cGAN). We introduce a superpixel average pooling layer into the cGAN decoder, which utilizes superpixels as a pooling layout to improve boundary segmentation. In addition, we adopt a multiscale input strategy to enable the network to learn scale‐invariant features with increased robustness. The performance of the model is evaluated with two public datasets: CBIS‐DDSM and INbreast. Moreover, ablation analysis is conducted to evaluate further the individual contribution of each block to the performance of the network. Results Dice and Jaccard scores of 93.37% and 87.57%, respectively, are obtained for the CBIS‐DDSM dataset. The Dice and Jaccard scores for the INbreast dataset are 91.54% and 84.40%, respectively. These results indicate that our proposed model outperforms current state‐of‐the‐art breast mass segmentation methods. The superpixel average pooling layer and multiscale input strategy has improved the Dice and Jaccard scores of the original cGAN by 7.8% and 12.79%, respectively. Conclusions Adversarial learning with the addition of a superpixel average pooling layer and multiscale input strategy can encourage the Generator network to generate masks with increased realism and improve breast mass segmentation performance through the minimax game between the Generator network and Discriminator network.

Review of Deep Learning Based Automatic Segmentation for Lung Cancer Radiotherapy

“…Employing deep learning in clinical practice may potentially shorten the unnecessary time and alleviate the workload of relevant staff (48). In recent years, several deep learning based automatic segmentation techniques have been proposed successively (35,(49)(50)(51)(52)(53)(54)(55)(56). In this section, studies related to the deep learning based automatic segmentation of the OARs and GTV in lung cancer are discussed and compared.…”

TrEnD: A transformer‐based encoder‐decoder model with adaptive patch embedding for mass segmentation in mammograms

“…Deep neural networks tend to be overwhelmed by the large class and ignore the small one. 21 To cope with the imbalance problem, most previous studies have opted for (1) segmenting masses within extracted ROIs 6,8,14,[22][23][24] or (2) relying on an additional mass detection stage. 25,26 Zhu et al 15 proposed an adversarial FCN-CRF network for mass segmentation from ROIs.…”

“…Loss = 𝛼L GD + 𝛽L BCE (6) where y j represents the j-th pixel of the ground truth image, p j is the j-th value of the predicted probabilistic map, n is the total amount of pixels in each image, w refers to the proportion of the number of samples in the positive class over that total number of samples, 𝛼 and 𝛽 are constants that control the strength of the two loss terms and are both set to 1. 𝜖 is also a constant that equals to 1.…”

“…Many studies have demonstrated shape and margin are two important indicators in estimating the potential malignancy of breast masses. For that reason, mass segmentation in mammograms is a critical task of breast cancer CAD systems 5,6 . Early segmentation algorithms for breast masses are often based on thresholding, region growing, and clustering methods, which rely heavily on manual features 7–9 .…”

“…For that reason, mass segmentation in mammograms is a critical task of breast cancer CAD systems. 5,6 Early segmentation algorithms for breast masses are often based on thresholding, region growing, and clustering methods, which rely heavily on manual features. [7][8][9] Hence, these algorithms are difficult to automate.…”