gms | German Medical Science

Introduction: Image-based analysis plays an important role in the diagnosis of specific diseases. For some of the rare diseases, in particular neurodegenerative ones, it is necessary to make image-based diagnosis. The white matter degeneration of the brain is proven to be detected by Magnetic Resonance Imaging (MRI). Among the neurodegenerative diseases, Leuko-dystrophies (LD) [1] are a family of genetically determined rare diseases with an estimated prevalence of fewer than 1 in 40,000 people worldwide. LD is a fatal disease characterized by degeneration of the white matter in the central nervous system. The symptoms of LD patients include progressive problems in their motor function, balance, sight, hearing, breathing, and cognition. Due to their similarity to Multiple Sclerosis (MS), LD diseases are commonly misdiagnosed [2].

Creating a pipeline for diagnosing diseases based on medical images would be a great asset for clinicians, education and training purposes. By virtue of machine learning (ML) algorithms, it is possible to find the specific patterns of diseases. Through image processing techniques, the ML model is trained to perform white matter lesion segmentation in the MRI scans of the brain. Based on the detected lesions and their patterns and locations, the diagnostic model has the potential to support the clinician in diagnosis of LD.

In the case of rare diseases, training the machine learning (ML) model is a great challenge since there is not enough amount of clinical data available. Generating synthetic MRI data is a solution for tackling this issue. In addition, access to medical data has privacy restrictions. Hence, it is necessary to perform a distributed data analysis through federated learning techniques.

The goal of the current research is to establish an expert system for diagnostic support for Leuko-dystrophies by analyzing MRI scans. This system is planned to be a pipeline that performs multi-site modeling of LD rare disease by employing federated learning and benefiting from synthetic image data.

Methods: For developing an image-based diagnostic support system, an ML model is trained by the available brain MRI data of LD patients. The MRI data are in three different locations, and the privacy of the matters. The solution is an image processing pipeline that automatically analyses the data in each data location and trains the ML model. We utilize the Personal Health Train [3] (PHT) infrastructure for performing federated learning. PHT is a privacy-preserving distributed analytics platform for healthcare data. In the PHT, the analytical tasks are brought to each location. Hence, the data instances remain in their original location.

The proposed pipeline, based on the PHT, is shown in the figure (top right). In this pipeline, the train ((1) in the figure) takes the analytical tasks to each data location for processing the available MRI scans. In each location, there is a database of the MRI data ((2) in the figure). For preprocessing the data ((3) in the figure), the homogeneity of the MRI scans is corrected and they are registered to a standard brain template. Registering each patient’s brain image to the template facilitates the comparison with other patients. Then, for fulfilling privacy aspects, the brain MRIs are defaced. Since the ML model processes only the brain, the skull and dura matter are stripped from the images.

The preprocessed images are used for training an ML model quantification of the lesions in the brain and identifying their locations. Here, a Deep Convolutional Neural Network architecture is used for obtaining the preliminary results. After the model is trained in the first data location, the train takes the model to the next location and performs similar operations there. After the final location, the model training is over, and the model can be evaluated.

Results and discussion: LD family of diseases is very similar to MS regarding the symptoms and the white matter lesions seen in the MRI scan, and they should be considered as a differential diagnosis. For MS, much more clinical data are available. Accordingly, here we have first implemented the lesion segmentation model on brain MRI of MS patients. The MS images used for training and testing the segmentation model are taken from the ISBI 2015 challenge [4]. The advantage of the challenge data over the LD data is that they are already annotated by a human expert. The unannotated images cannot be used for training and evaluating an ML model. In the figure ([Abb. 1]bottom, a-d), the original FLAIR image of an MS patient, the preprocessed image, the MS lesions segmented by the ML model, the MS lesions delineated by a human expert, and the metrics for evaluating the ML model for MS data are shown. By comparing the two lesion segmentations and the metrics, the model shows promising performance. Moreover, the preprocessing step is applied to the MRI of an LD patient, as shown in the figure ([Abb. 1]bottom, e, f).

Currently, transforming the lesion segmentation algorithm from MS to LD is an open issue that is planned to be carried out in near future. In this regard, we have faced a number of challenges such as harmonizing the MRI scan taken with different devices, imaging parameters and generating synthetic MRI scans of LD.