This Research Topic is part of the "From Prototype to Clinical Workflow: Moving Machine Learning for Lesion Quantification into Clinical Practice series". You can find Volume II
hereOver the course of the last decade, we have seen a considerable improvement in the performance of algorithms for lesion segmentation and detection in Neuroimaging. This development can be attributed to the advent of new Machine Learning algorithms, which has been greatly facilitated by a series of challenges in Medical Imaging (e.g., Brain Tumor Segmentation (BRATS) Challenge). In particular, deep learning algorithms (e.g., U-Net architecture) proved to be very effective to solve learning problems involving the segmentation and detection of brain lesions. Such algorithms have a wide variety of potential applications in Neuroradiology ranging from quantification of disease progression in brain tumors and multiple sclerosis to support of differential diagnosis and prognosis in neurovascular and neuroinflammatory diseases.
In contrast to the encouraging research on Machine Learning algorithms for brain lesion quantification, there is a substantial lack of methods which have actually been translated into clinical routine. The integration of brain lesion segmentation and detection methods in clinical workflows is a challenge often left unaddressed by researchers and hence remains poorly understood. When deployed in clinical routine, Machine Learning methods should ideally be robust to changes in input data, produce reliable results, and provide actual support in clinical decision making. Several issues of Machine Learning research play a crucial role for the successful translation of Machine Learning-based brain lesion quantification methods including domain adaptation, uncertainty quantification, and model validation. In addition, new public datasets and software modules are needed to support the translation of Machine Learning algorithms from research into clinical practice.
Therefore, the objective of the current Research Topic is to foster new research in areas relevant to the translation of Machine Learning-based brain lesion quantification methods into clinical routine. This includes but is not limited to the following topics:
• Domain Adaptation (e.g., detection and quantification of dataset shifts)
• Uncertainty quantification (e.g., calibration of segmentation and detection models, novel methods for uncertainty estimation, integration of uncertainty estimates in a clinical workflow)
• Model validation (e.g., research on new metrics for evaluation of brain lesion segmentation, validation studies of existing methods on large datasets, systematic reviews of results from brain lesion segmentation and detection challenges)
• Machine Learning and data privacy (e.g., federated learning or secure multi-party computation with applications in brain lesion quantification)
• Data contributions (e.g., new public datasets to support the development of Machine Learning methods for brain lesion quantification)
• Software contributions (e.g., Python modules offering functionality to support the integration of Machine Learning methods into clinical routine such as automated image quality rating)
• Perspectives on how Machine Learning-based brain lesion quantification has been integrated in the clinical workflow of your institution