Table 3.

Summary of the application of deep learning algorithms in breast cancer histopathology for segmentation.

Model	Strategy	Advantages	Publication
ResNet	Propose segmentation of limited data using rough image-level tags with performance comparable to fully labeled datasets	The main feature of ResNet is the residual block, the purpose of the residual block is to preserve the characteristics of the parameters before the current layer is trained and to pass these parameters onto the subsequent layers together with the trained data	[84]
FCN	Propose a fast segmentation method for breast cancer metastases in pathological images	The FCN replaces the fully connected layer behind the traditional CNN with a convolutional layer so that the output of the network will be a heat map rather than a category; at the same time, the image size is recovered using upsampling in order to address the reduction in image size due to convolution and pooling	[85]
	Propose an automatic method for detecting mitosis		[86]
	Describe a method to automatically segment nuclei from hematoxylin and eosin (H&E)-stained histopathology data with fully convolutional networks		[87]
	Use annotated datasets to create accurate models		[60]
	Propose a histopathological tissue analysis framework based on deep learning and verifies its universality and model generalization under different data distributions		[88]
U-Net	Use histopathological images obtained with hematoxylin and eosin staining for biopsy samples for the diagnosis and segmentation of breast cancer	U-Net networks are able to use valid labeled data more effectively from a very small number of training images, relying on data augmentation	[89]
	Address the task of tissue-level segmentation in intermediate resolution of histopathological breast cancer images		[90]
	Propose a deep learning framework consisting of high-resolution encoder paths, pyramidal pooled bottleneck modules in porous space, and decoders		[91]
	Investigate whether it is possible to further improve the performance of the classifier model at the patch level by integrating multiple extracted histological features into the input image		[92]
CNN	Improve the performance of current Simple Linear Iterative Clustering (SLIC) algorithm to achieve hyperpixel segmentation of high-dimensional features		[93]
	Use a pretrained convolutional neural network (CNN) for segmentation and then another Hybrid-CNN for classification of mitoses		[94]
	Identify a useful cell segmentation approach with histopathological images that uses prominent deep learning algorithms and spatial relationships		[95]
	Propose a framework that combines the effectiveness of attention-based encoder–decoder architecture with an empty space pyramid pool with efficient dimensional convolution (kide-Segnet)		[96]
	Propose a deep learning model for automatic segmentation of complex cores in tissue images by encoder-decoder structure		[97,98]
Transformer	Transformer-encoded global features improve U-Net segmentation performance	Transformer model can be used to encode the global features of pathological images and can improve the performance of current algorithms in many fields	[79]