Deep learning (DL) has emerged as a key technology for enhancing magnetic resonance imaging (MRI). This review summarizes recent advances in DL for MRI reconstruction, focusing on various approaches and architectures that improve image quality, accelerate scans, and address data-related challenges. It explores end-to-end neural networks, pre-trained and generative models, and self-supervised methods, highlighting their contributions to overcoming traditional MRI limitations. The review also discusses DL's role in optimizing acquisition protocols, enhancing robustness against distribution shifts, and tackling biases. It outlines current successes, limitations, and future directions for leveraging DL in MRI reconstruction, emphasizing its potential to significantly impact clinical imaging practices. MRI has long been a critical tool in diagnostic radiology due to its non-invasive nature and exceptional soft tissue contrast. However, its full potential is often limited by imaging speed and resolution. Long acquisition times are a major barrier in clinical MRI, primarily due to the trade-off between image quality and speed. High-resolution images, crucial for accurate diagnosis, require extended scan durations that can be uncomfortable for patients and increase the risk of motion artifacts. Additionally, sequences with long repetition times, extensive field-of-view coverage, and protocols requiring multiple contrasts significantly extend MRI scan durations. The acquisition of 3D images also necessitates longer scan times due to the increased volume of data being collected. Significant research efforts have been dedicated to accelerating MRI. Central to these efforts is the development of methods for image reconstruction from under-sampled data. Techniques based on parallel imaging (PI) and compressed sensing (CS) have been developed to reduce scan times. However, both PI and CS have limitations, including dependency on coil array geometry and computational demands. DL has emerged as a promising solution to these challenges, offering remarkable success in extracting complex patterns from large datasets. DL methods focus on learning from vast amounts of data to transform under-sampled or noisy data into high-fidelity images, demonstrating their ability to mitigate artifacts, enhance resolution, and accelerate the imaging process. The review covers various DL approaches and architectures, including end-to-end neural networks, pre-trained denoisers (plug-and-play methods), generative models, un-trained methods, and self-supervised methods. It also identifies recent architectures such as transformers and dual-domain networks. The top-performing models in the 2020 FastMRI challenge and the 2024 CMRxRecon challenge all used end-to-end neural networks. The review discusses the image formation process, conventional model-based image recovery, and optimization algorithms. It also explores DL reconstruction approaches, including unrolled networks, pretrained plug-and-play methods, generative priors, and self-supervised methods. The review highlights the potential of DL to significantly impact clinical imaging practices by improving image quality, accelerating scans, and addressing data-related challenges.Deep learning (DL) has emerged as a key technology for enhancing magnetic resonance imaging (MRI). This review summarizes recent advances in DL for MRI reconstruction, focusing on various approaches and architectures that improve image quality, accelerate scans, and address data-related challenges. It explores end-to-end neural networks, pre-trained and generative models, and self-supervised methods, highlighting their contributions to overcoming traditional MRI limitations. The review also discusses DL's role in optimizing acquisition protocols, enhancing robustness against distribution shifts, and tackling biases. It outlines current successes, limitations, and future directions for leveraging DL in MRI reconstruction, emphasizing its potential to significantly impact clinical imaging practices. MRI has long been a critical tool in diagnostic radiology due to its non-invasive nature and exceptional soft tissue contrast. However, its full potential is often limited by imaging speed and resolution. Long acquisition times are a major barrier in clinical MRI, primarily due to the trade-off between image quality and speed. High-resolution images, crucial for accurate diagnosis, require extended scan durations that can be uncomfortable for patients and increase the risk of motion artifacts. Additionally, sequences with long repetition times, extensive field-of-view coverage, and protocols requiring multiple contrasts significantly extend MRI scan durations. The acquisition of 3D images also necessitates longer scan times due to the increased volume of data being collected. Significant research efforts have been dedicated to accelerating MRI. Central to these efforts is the development of methods for image reconstruction from under-sampled data. Techniques based on parallel imaging (PI) and compressed sensing (CS) have been developed to reduce scan times. However, both PI and CS have limitations, including dependency on coil array geometry and computational demands. DL has emerged as a promising solution to these challenges, offering remarkable success in extracting complex patterns from large datasets. DL methods focus on learning from vast amounts of data to transform under-sampled or noisy data into high-fidelity images, demonstrating their ability to mitigate artifacts, enhance resolution, and accelerate the imaging process. The review covers various DL approaches and architectures, including end-to-end neural networks, pre-trained denoisers (plug-and-play methods), generative models, un-trained methods, and self-supervised methods. It also identifies recent architectures such as transformers and dual-domain networks. The top-performing models in the 2020 FastMRI challenge and the 2024 CMRxRecon challenge all used end-to-end neural networks. The review discusses the image formation process, conventional model-based image recovery, and optimization algorithms. It also explores DL reconstruction approaches, including unrolled networks, pretrained plug-and-play methods, generative priors, and self-supervised methods. The review highlights the potential of DL to significantly impact clinical imaging practices by improving image quality, accelerating scans, and addressing data-related challenges.