Core Concepts
This review provides a comprehensive overview of deep learning-based approaches for semantic segmentation of natural and medical images, categorizing the literature into six main groups: architectural improvements, optimization function-based improvements, data synthesis-based improvements, sequenced models, weakly supervised methods, and multi-task models. The review analyzes the contributions and limitations of each group and presents potential future research directions.
Abstract
The review covers deep learning-based semantic image segmentation approaches for both natural and medical images. It starts by discussing architectural improvements, including fully convolutional networks, encoder-decoder models, and attention-based methods. The review then covers optimization function-based improvements, such as cross-entropy, overlap-based, and boundary-based loss functions, and how they have been applied to medical image segmentation.
Next, the review discusses data synthesis-based methods, particularly generative adversarial networks (GANs) for data augmentation. It covers GAN-based approaches for generating labeled image-segmentation mask pairs to address the limited data problem in medical image analysis.
The review also covers sequenced models, such as recurrent neural networks (RNNs) and long short-term memory (LSTMs), which have been used to model temporal dependencies in medical image sequences. Weakly supervised methods, which leverage alternative forms of annotations (e.g., bounding boxes, image-level labels) to train segmentation models, are also discussed.
Finally, the review covers multi-task models that jointly learn segmentation along with other tasks, such as detection and classification, to leverage shared representations and improve overall performance.
Throughout the review, the authors highlight the key contributions, limitations, and potential future research directions for each category of methods.
Stats
"Deep learning has had a tremendous impact on various fields in science. The focus of the current study is on one of the most critical areas of computer vision: medical image analysis (or medical computer vision), particularly deep learning-based approaches for medical image segmentation."
"Image segmentation is formally defined as "the partition of an image into a set of nonoverlapping regions whose union is the entire image" (Haralick and Shapiro, 1992)."
"Deep architectural improvement has been a focus of many researchers for different purposes, e.g., tackling gradient vanishing and exploding of deep models, model compression for efficient small yet accurate models, while other works have tried to improve the performance of deep networks by introducing new optimization functions."
Quotes
"Deep learning has had a tremendous impact on various fields in science. The focus of the current study is on one of the most critical areas of computer vision: medical image analysis (or medical computer vision), particularly deep learning-based approaches for medical image segmentation."
"Image segmentation is formally defined as "the partition of an image into a set of nonoverlapping regions whose union is the entire image" (Haralick and Shapiro, 1992)."
"Deep architectural improvement has been a focus of many researchers for different purposes, e.g., tackling gradient vanishing and exploding of deep models, model compression for efficient small yet accurate models, while other works have tried to improve the performance of deep networks by introducing new optimization functions."