Brain diseases are increasingly recognized as a major threat to health and lives, affecting millions of people all over the world. Machine learning technologies, especially different deep learning paradigms, have been intensively applied to enable efficient end-to-end diagnosis for brain disease. Currently new models, such as transformer-based networks, have become strong competitors of traditional convolutional neural networks in a variety of predictive tasks using medical images. At the same time, in modern clinical practices, many large-scale research projects are coordinated across multiple collaborated centers. This involves coordinating the different data analysis pipelines standardized at different sites, for example, training and deploying a diverse group of machine learning models with a variety of computing devices (e.g., different edge devices in IoT systems), data modalities (genetic sequencing data and imaging data, such as, CT, MRI, PET, Ultrasound, etc., of different context and image quality respectively), data analysis protocols and learning schemes (weakly supervised, adversarial learning, reinforcement learning, etc. ). This not only introduces performance issues and difficulties in model transferability between sites and devices, but also brings new challenges of data privacy and rationality of the model predictions.
A variety of collaborative learning frameworks have been explored recently for better reliability of machine learning algorithms in such a complicated clinical decision-making scenario. For example, current federated learning techniques allows privacy-preserving training for models adopted by different sites with distributed storage of data (i.i.d or non-i.i.d); model compression enhances applicability of a trained model to different edge devices with different computing power; performance and robustness of independently trained models can be improved through state-of-the-arts model assembling and knowledge distillation methods; development of XAI and transfer learning are also achieving remarkable progresses for better transparency and transferability of present machine learning models; multiple techniques have also been proposed to align multi-modality or heterogeneous in latent spaces with well-defined semantic meanings. The purpose of this Research Topic is to collect original research achievements in machine learning that facilitate practical collaborative decision making for brain disease diagnosis. Any efforts made to narrow the gaps between distinct devices, data, models and different machine learning practices between sites are welcomed, but the correlation between machine learning methods and the downstream diagnostic purposes should be explicitly discussed.
We welcome articles including, but are not limited to the following topics:
• Federated learning methods for brain imaging and post-processing;
• Model compression and inference acceleration;
• Comparison between transformer models and traditional CNNs;
• Transfer learning and domain-adaptation;
• Information fusion, joint prediction using multi-modality imaging and other types of data;
• Object detection, recognition, classification, segmentation, registration and reconstruction, enhancement of biomedical brain imaging data;
• Cross-modality image synthesis;
• Transparency, explainability and interpretability of deep learning models;
• Privacy and ethics issues in deep learning schemes, etc..
• Deep learning models in modern cyber physical systems
Brain diseases are increasingly recognized as a major threat to health and lives, affecting millions of people all over the world. Machine learning technologies, especially different deep learning paradigms, have been intensively applied to enable efficient end-to-end diagnosis for brain disease. Currently new models, such as transformer-based networks, have become strong competitors of traditional convolutional neural networks in a variety of predictive tasks using medical images. At the same time, in modern clinical practices, many large-scale research projects are coordinated across multiple collaborated centers. This involves coordinating the different data analysis pipelines standardized at different sites, for example, training and deploying a diverse group of machine learning models with a variety of computing devices (e.g., different edge devices in IoT systems), data modalities (genetic sequencing data and imaging data, such as, CT, MRI, PET, Ultrasound, etc., of different context and image quality respectively), data analysis protocols and learning schemes (weakly supervised, adversarial learning, reinforcement learning, etc. ). This not only introduces performance issues and difficulties in model transferability between sites and devices, but also brings new challenges of data privacy and rationality of the model predictions.
A variety of collaborative learning frameworks have been explored recently for better reliability of machine learning algorithms in such a complicated clinical decision-making scenario. For example, current federated learning techniques allows privacy-preserving training for models adopted by different sites with distributed storage of data (i.i.d or non-i.i.d); model compression enhances applicability of a trained model to different edge devices with different computing power; performance and robustness of independently trained models can be improved through state-of-the-arts model assembling and knowledge distillation methods; development of XAI and transfer learning are also achieving remarkable progresses for better transparency and transferability of present machine learning models; multiple techniques have also been proposed to align multi-modality or heterogeneous in latent spaces with well-defined semantic meanings. The purpose of this Research Topic is to collect original research achievements in machine learning that facilitate practical collaborative decision making for brain disease diagnosis. Any efforts made to narrow the gaps between distinct devices, data, models and different machine learning practices between sites are welcomed, but the correlation between machine learning methods and the downstream diagnostic purposes should be explicitly discussed.
We welcome articles including, but are not limited to the following topics:
• Federated learning methods for brain imaging and post-processing;
• Model compression and inference acceleration;
• Comparison between transformer models and traditional CNNs;
• Transfer learning and domain-adaptation;
• Information fusion, joint prediction using multi-modality imaging and other types of data;
• Object detection, recognition, classification, segmentation, registration and reconstruction, enhancement of biomedical brain imaging data;
• Cross-modality image synthesis;
• Transparency, explainability and interpretability of deep learning models;
• Privacy and ethics issues in deep learning schemes, etc..
• Deep learning models in modern cyber physical systems
