Deep learning has become a robust tool for medical image segmentation, widely used for separating homogeneous areas in diagnostic and treatment pipelines. This article critically reviews popular deep learning methods for medical image segmentation, summarizes common challenges, and suggests possible solutions. It focuses on recent machine learning techniques applied in medical image segmentation, analyzing their structures, methods, strengths, and weaknesses. The article is divided into three main sections: approaches (network structures), training techniques, and challenges. The network structure section introduces major popular structures used for image segmentation, their advantages, and shortcomings. It covers 2D CNNs, 2.5D CNNs, 3D CNNs, FCNs, U-Net, V-Net, CRNs, RNNs, and LSTM. Each structure is discussed in terms of its design, performance, and challenges. The training techniques section explores state-of-the-art techniques for training deep neural networks, including deeply supervised learning, weakly supervised learning, and transfer learning. These techniques are crucial for improving segmentation accuracy and handling limited annotated data. The challenges section addresses various issues in medical image segmentation using deep learning, including limited annotated data, data augmentation, sparse annotation, effective negative sets, class imbalance, overfitting, training time, gradient vanishing, and organ appearance. Solutions such as data augmentation, transfer learning, patch-wise training, and class re-weighting are discussed. The article concludes that deep learning approaches will play a significant role in medical image segmentation, offering improved accuracy and efficiency. It emphasizes the importance of choosing appropriate network structures and being aware of potential challenges and solutions.Deep learning has become a robust tool for medical image segmentation, widely used for separating homogeneous areas in diagnostic and treatment pipelines. This article critically reviews popular deep learning methods for medical image segmentation, summarizes common challenges, and suggests possible solutions. It focuses on recent machine learning techniques applied in medical image segmentation, analyzing their structures, methods, strengths, and weaknesses. The article is divided into three main sections: approaches (network structures), training techniques, and challenges. The network structure section introduces major popular structures used for image segmentation, their advantages, and shortcomings. It covers 2D CNNs, 2.5D CNNs, 3D CNNs, FCNs, U-Net, V-Net, CRNs, RNNs, and LSTM. Each structure is discussed in terms of its design, performance, and challenges. The training techniques section explores state-of-the-art techniques for training deep neural networks, including deeply supervised learning, weakly supervised learning, and transfer learning. These techniques are crucial for improving segmentation accuracy and handling limited annotated data. The challenges section addresses various issues in medical image segmentation using deep learning, including limited annotated data, data augmentation, sparse annotation, effective negative sets, class imbalance, overfitting, training time, gradient vanishing, and organ appearance. Solutions such as data augmentation, transfer learning, patch-wise training, and class re-weighting are discussed. The article concludes that deep learning approaches will play a significant role in medical image segmentation, offering improved accuracy and efficiency. It emphasizes the importance of choosing appropriate network structures and being aware of potential challenges and solutions.