Autoencoder-Based Feature Generation
from Electroencephalograms of Microsleep Events
David Sommer, Constantin Pohl, Martin Golz
University of Applied Sciences Schmalkalden, Faculty of Computer Science, Germany, [email protected]
Introduction
Neural autoencoder networks are able to compress data by non-linear mapping the high-dimensional input space to a low-dimensional transform space and simultaneously generating an inverse mapping that outputs a representation of the input space. The mean square error between input vectors and estimated input vectors generated by the network serves as objective function to be minimised. Empirically, network architectures can be found so that sufficiently high com- pressions can be achieved with a reasonably minimised objective function. We compare three different variants of this form of feature generation from raw data and evaluate them with the same classification method, namely the support vector machine (SVM).
Methods
A total of 10,356 microsleep events (MSE) and 9,389 counter-examples (SA, sustained attention) were recorded from 70 young adults during night-time studies in our driving simulation laboratory in 2007, 2009, 2016 and 2017. 6-channel EEG (Fp1, Fp2, C3, C4, O1, O2) was processed with the following three variants using autoencoders: (1) raw data, (2) log power spectral densities (LogPSD) and (3) raw data concatenated with LogPSD. The length of raw data segments was 4s. From them, the LogPSD was non-parametrically estimated within the interval 0 to 30Hz. The transform space dimensionality was optimised empirically. Resulting transformed data sets were validated by SVM with Gaussian ker- nel using repeated random subsampling.
Results
A network with 2,048 and 50 neurons in the first and second hidden layer, respectively, and 128 neurons in the com- pression layer turned out to be optimal. On independent validation sets, mean classification accuracies of: (1) 85.6%, (2) 92.7% and (3) 88.4% were achieved.
Conclusion
Although LogPSD features do not contain phase information, more accurate classifications are achievable. The signifi- cantly worse result for time-domain raw data might be caused both by lack of translational invariance and by strong noise caused by the many brain processes running in parallel.
Exploitation of Kronecker Structure in Gaussian Process Regression for Efficient Biomedical Signal Processing
Jannik Prüßmann, Institute for Electrical Engineering in Medicine, Universität zu Lübeck, 23562 Lübeck, Germany, [email protected],
Jan Graßhoff, Institute for Electrical Engineering in Medicine, Universität zu Lübeck, 23562 Lübeck, Germany, [email protected],
Philipp Rostalski, Institute for Electrical Engineering in Medicine, Universität zu Lübeck, 23562 Lübeck, Germany, [email protected]
Introduction
In modern medicine, an ever growing variety of signals can be measured from patients, which often have spatio-temporal structure. Gaussian processes (GPs) provide a powerful tool to implement prior knowledge about such signals. Unfortu- nately, due to the requirements in storage and runtime, standard GP regression is limited to datasets containing at most a few thousand observations and is therefore infeasible on large biomedical datasets. To overcome these limitations, this work evaluates methods for efficient spatio-temporal GP inference and learning. The methods are exemplified by a bio- medical signal separation problem on esophageal manometry data.
Methods
The covariance matrices involved in GP regression are often highly structured. Assuming the spatio-temporal inputs lie on a grid and the covariance function is defined as a product of dimension-related kernels, the resulting covariance matrix has Kronecker structure. By exploiting this structure, fast matrix-vector multiplications can be calculated and the resulting linear system of equations can be solved efficiently. Also, the full covariance matrix can be stored by a set of smaller matrices, exploiting its Kronecker factors, and thereby reducing storage requirements. In combination, these routines enable efficient GP inference and hyperparameter learning.
Results
The performance of Kronecker-exploiting methods is compared to standard GP regression. Unlike standard GP methods, Kronecker methods can be applied to very large biomedical datasets and perform substantially faster. The performance of GP inference and learning is demonstrated on esophageal manometry data. In this example of a biomedical, spatio- temporal dataset, the method was able to efficiently separate cardiac and ventilation-related components of the manometry signal.
Conclusion
The exploitation of Kronecker structure in spatio-temporal GP covariance matrices allows to overcome the limitations of GP regression concerning runtime and storage requirements. As it was exemplarily shown on esophageal manometry data, the investigated methods enable the application of GPs to large spatio-temporal, biomedical datasets.
An interactive system for muscle and fat tissue identification of the lumbar spine using semantic segmentation
Richard Bieck, Innovation Center Computer Assisted Surgery, University Leipzig, Leipzig, Germany, Richard.bieck@ic- cas.de
David Baur, Orthopaedics, Trauma and Plastic Surgery, Spine Center, Leipzig, Germany
Johann Berger, Innovation Center Computer Assisted Surgery, University Leipzig, Leipzig, Germany Tim Stelzner, Innovation Center Computer Assisted Surgery, University Leipzig, Leipzig, Germany Anna Volker, Orthopaedics, Trauma and Plastic Surgery, Spine Center, Leipzig, Germany
Christoph-Eckhard Heyde, Orthopaedics, Trauma and Plastic Surgery, Spine Center, Leipzig, Germany Thomas Neumuth, Innovation Center Computer Assisted Surgery, University Leipzig, Leipzig, Germany
Introduction
In orthopaedic diagnostics, the complex and sometimes irregular distribution of fat and muscle tissue of the lumbar spine is currently only assessable through direct inspection without means for automated measurement. We, therefore, introduce a system that allows the immediate identification and inspection of fat and muscle structures around the lumbar spine.
Methods
The system comprises a backend component that accepts MRI data from a web-based interactive frontend as REST re- quests. The MRI data is passed through a U-net model, fine-tuned on lumbar mri images, to generate segmentation masks of fat and muscle areas. The result is sent back to the frontend that functions as an inspection tool. For the model training, 4000 MRI images from 108 patients were used in a k-fold crossvalidation study with k = 10. The model training was performed over 25-30 epochs. We applied shift, scale and rotation operations as well as elastic deformation and distortion functions for image augmentation and a combined objective function using Dice and Focal loss.
Results
The trained models reached a macro F1-score of 0.92 with a mean area error for muscle and fat tissue of 0.1 and 0.3, respectively. The interactive web-based frontend as an inspection tool was evaluated by clinicians to be suitable for the exploration of patient data as well as the assessment of segmentation results.
Conclusion
We developed a system that uses semantic segmentation to identify fat and muscle tissue areas in MRI images of the lumbar spine. Further improvements should focus on the segmentation accuracy of fat tissue, as it is a determining factor in surgical decision making. To our knowledge, this is the first system that automatically provides semantic information of the respective lumbar tissues.
Automatic Segmentation of a Freezing Provoking Task in Parkinson’s Disease Using Inertial Sensors
Christina Salchow-Hömmen, Klinik für Neurologie, Charité-Universitätsmedizin Berlin, Germany, [email protected]
Jonas Meyer-Ohle, Klinik für Neurologie, Charité-Universitätsmedizin Berlin, Germany, [email protected] Magdalena Jochner, Klinik für Neurologie, Charité-Universitätsmedizin Berlin, Germany,
Andrea Kühn, Klinik für Neurologie, Charité-Universitätsmedizin Berlin, Germany, [email protected] Nikolaus Wenger, Klinik für Neurologie, Charité-Universitätsmedizin Berlin, Germany, [email protected]
Introduction
Freezing of gait (FOG) in Parkinson’s disease (PD) is an episodic symptom that is difficult to provoke and to evaluate in clinical settings. Specific tasks including turns and obstacles have been designed to provoke FOG in controlled conditions.
However, due to their complexity and short walking distances, commercial gait analysis software with wearables often fails to determine gait parameters of these tasks correctly, hindering an objective evaluation. We present a novel automatic segmentation method based on inertial sensors for a freezing provoking task (FOGT) to overcome this issue.
Methods
The selected FOGT consists of standing-up, turning 360° inside a marked, narrow square, opening and passing a door.
Fifteen PD patients (age = 68.2±8.98, ON or OFF medication) performed three repetitions of the task. Three inertial measurement units (IMUs), containing 3D accelerometers and gyroscopes (MucscleLab, Ergotest), were located on both feet and the hip. All trials were recorded on video from two perspectives. Sensor fusion was applied to determine yaw, pitch, and roll angles of each sensor and to estimate gait phases from the foot sensors. Using a combination of these signals, we developed an algorithm that detects start and end point of each sub-task.
Results
The segmentation times marked by the algorithm were evaluated using the visual times annotated from synchronized videos by two independent examiners. The average difference between automatic and visual duration (ΔT) across all sub- tasks was 209 ms. Correlation (Pearson) between automatic and visual start and end of sub-tasks was high with r=0.999 and r=0.998, respectively.
Conclusion
The suggested methodology allows automatic segmentation of all sub-tasks of a FOGT with three IMUs. The segmenta- tion can be used for providing objective information such as duration of, steps, and (turn-)acceleration during sub-tasks, which can facilitate the evaluation of FOGTs for clinical and research purposes.
Acquisition of Semantics for AI-based Applications in Medical Technologies – An Overview
Thomas Wittenberg, Fraunhofer IIS, Erlangen, Germany, [email protected] Michaela Benz, Fraunhofer IIS, Erlangen, Germany, [email protected]
Andras Foltyn, Fraunhofer IIS, Erlangen, Germany, [email protected] Julia Hetzel, Fraunhofer IIS, Erlangen, Germany, [email protected] Ralf Hackner, Fraunhofer IIS, Erlangen, Germany, [email protected]
Thomas Eixelberger, Fraunhofer IIS, Erlangen, Germany, [email protected]
Introduction
For the development, evaluation and use of AI-based methods, with an application in the healthcare market, ade- quate information about the ‘meaning’ of the data (the ‘semantic’ or ‘knowledge’) is needed. The need of ‘seman- tics’ is independent from the addressed task (as e.g., processing and analysis of patient related data, prediction of critical events, planning of healthcare procedures) or origin and type (image data, bio signals, health records, ma- chine states, …). With this ‘semantic information’ a tight relation between the raw data and the human-under- standable concepts from the ‘real world’ can be established. Nevertheless, as the amount of data needed to develop such AI-based methods is strongly increasing, the acquisition of the related ‘knowledge’ becomes more and more challenging. Within this contribution, an overview of currently available concepts of knowledge acquisition shall be described and evaluated. Furthermore, ideas for possible developments shall be discussed.
Methods
In order to obtain an overview of currently used knowledge-acquisition methods for AI-based technologies, lite- rature in the field of biomedical engineering was scanned. Specifically, to yield a broader view of AI-based appli- cations, recent proceedings volumes of the DG-BMT and BVM were used as a starting point for the research.
Results
From the literature, four main groups of knowledge acquisition related to AI-based technologies have been iden- tified: (1) For image data (2D, 3D, 2D+t, 3D+t) mainly iconic annotation methods are used, where experienced users mark or draw depicted entities in the images and label them using a predefined set of classifications. Simi- larly, temporal bio signals of all kinds (EMG, ECG, Emotions, …) are manually labelled, whereby important
‘events’ along the timeline are marked and related to predefined classes. (2) If no sufficient data is available for the development and evaluation of AI-based applications, data augmentation and simulations are applied yielding data and semantics at the same time. These methods can range from physical modelling approaches (motived from computer games and using such visualization engines) to AI-based generation of such data. (3) In applications, where expensive sensors are replaced by low-cost devices, which then make use of AI-procedures to ‘upgrade’ the captured raw data, the ‘expensive’ high-grade data is directly used as the needed ‘semantics’ for the training of the AI-methods. (4) Finally, classic ‘rule-based‘-approaches are still found, where the factual and procedural know- ledge of humans about the content of the data and its context is translated into machine-understandable procedures, and then implemented in adequate measures.
Conclusion
The currently used approaches to obtain and collect ‘knowledge’ about data needed for the development of AI- based devices are all depending on a strong involvement of human experts, either for data labelling, structured knowledge modelling, designing adequate high-end simulations or applying high-grade devices. To reduce this expensive workload, more intelligent and hybrid approaches are needed, shifting the focus from ‘the-human-in- the-loop’ to the ‘machine-in-the-loop’. These could e.g., be ‘boot-strapping’ or ‘active-learning’ approaches, where already available (machine-learning) methods are used for augmentation of data with labels (knowledge), and where human experts are only needed for verification or corrections. Also, a combination of rule-based data (and knowledge) simulation could be thought of, implicitly including a previous version of an AI-based approach.
Acknowledgement
This work was supported by the Bavarian Ministry of Economic Affairs, Regional Development & Energy by the Center for Analytics-Data -Applications (ADA-Lovelace Center). BAYERN DIGITAL II\ (20-3410-2-9-8).