The paper introduces AUTOMAP, a deep learning-based framework for image reconstruction that learns a mapping between sensor and image domains from training data. This approach allows for flexible and efficient reconstruction across various imaging modalities, including MRI, CT, PET, and ultrasound. Unlike traditional methods that rely on handcrafted algorithms and require expert parameter tuning, AUTOMAP uses a deep neural network to learn the reconstruction process, enabling it to adapt to different acquisition strategies and noise conditions. The framework is trained on a diverse set of data, including natural images and brain scans, and demonstrates superior performance in noise immunity and artifact reduction compared to conventional methods. It also shows strong generalization capabilities, performing well across different encoding schemes without requiring specific domain knowledge. The method is particularly effective in low-SNR environments, such as low-dose CT and low-light imaging, where traditional reconstruction techniques struggle. The paper also highlights the ability of AUTOMAP to reconstruct phase information from complex-valued sensor data, expanding its applicability to medical imaging. The approach is supported by extensive experiments across multiple imaging modalities, demonstrating its effectiveness in reconstructing images from various sensor domain encodings. The results show that AUTOMAP outperforms conventional methods in terms of noise and sampling defect immunity, with the ability to learn robust features from data without explicit noise modeling. The framework is implemented using a deep neural network with fully connected layers and sparse convolutional autoencoders, and its performance is validated using brain MRI data from the Human Connectome Project. The study also explores the hidden-layer activity of the network during reconstruction, showing that the model effectively extracts robust features and learns sparse representations. The results indicate that AUTOMAP can be used for a wide range of imaging applications, including the development of new acquisition strategies. The paper concludes that AUTOMAP provides a powerful new paradigm for image reconstruction, leveraging deep learning to learn optimal reconstruction functions for any acquisition strategy.The paper introduces AUTOMAP, a deep learning-based framework for image reconstruction that learns a mapping between sensor and image domains from training data. This approach allows for flexible and efficient reconstruction across various imaging modalities, including MRI, CT, PET, and ultrasound. Unlike traditional methods that rely on handcrafted algorithms and require expert parameter tuning, AUTOMAP uses a deep neural network to learn the reconstruction process, enabling it to adapt to different acquisition strategies and noise conditions. The framework is trained on a diverse set of data, including natural images and brain scans, and demonstrates superior performance in noise immunity and artifact reduction compared to conventional methods. It also shows strong generalization capabilities, performing well across different encoding schemes without requiring specific domain knowledge. The method is particularly effective in low-SNR environments, such as low-dose CT and low-light imaging, where traditional reconstruction techniques struggle. The paper also highlights the ability of AUTOMAP to reconstruct phase information from complex-valued sensor data, expanding its applicability to medical imaging. The approach is supported by extensive experiments across multiple imaging modalities, demonstrating its effectiveness in reconstructing images from various sensor domain encodings. The results show that AUTOMAP outperforms conventional methods in terms of noise and sampling defect immunity, with the ability to learn robust features from data without explicit noise modeling. The framework is implemented using a deep neural network with fully connected layers and sparse convolutional autoencoders, and its performance is validated using brain MRI data from the Human Connectome Project. The study also explores the hidden-layer activity of the network during reconstruction, showing that the model effectively extracts robust features and learns sparse representations. The results indicate that AUTOMAP can be used for a wide range of imaging applications, including the development of new acquisition strategies. The paper concludes that AUTOMAP provides a powerful new paradigm for image reconstruction, leveraging deep learning to learn optimal reconstruction functions for any acquisition strategy.