Hasso-Plattner-InstitutSDG am HPI
Hasso-Plattner-InstitutDSG am HPI
Login
 

Post-operative Motor Deficit Risk Prediction for Tumor Patients Using Deep Neural Networks

Shahryar Khorasani

Machine Learning & Digital Health
Hasso Plattner Institute

Office: G-2.1.33
Tel.: +49-(0)331 5509-4874
Email: Shahryar.Khorasani(at)hpi.de
Links: Homepage

Supervisor: Prof. Dr. Christoph Lippert

Motivation

Tumors that are located inside or in the proximity of motor cortex or corticospinal tract increase the risk of motor function damage after surgery. Neurosurgeons need to measure this risk in surgical planning. They use navigated transcranial magnetic stimulation (nTMS) to locate motor function in the motor cortex and refine tractography of the corticospinal tract. The tumor can induce changes in the neuronal excitability of the corticospinal tract and this can be measure by nTMS.

Distance between tumor and corticospinal tract, neuronal excitability and the involvement of the motor cortex are the three biomarkers that are currently used for measuring the risk of post-operative motor deficit [1].

Short Comings of Current Workflow

Neuronal excitability is calculated in terms of resting motor threshold (RMT). The ratio between the RMT in the affected hemisphere compared to the healthy hemisphere is called the RMT ratio. An RMT ratio smaller than 0.9 or larger than 1.1 is classified as abnormal [1]. The threshold used to binarize tumor-tract distance is 8mm [1]. If the tumor overlaps with the TMS positive points, it is known to involve the motor cortex. Any of the three binarized measurements can indicate whether a patient has a high risk or a non-increased risk of developing a motor deficit after the surgery. Extracting these biomarkers is a laborious task that can be improved and automated.

In the processing of the aforementioned measurements, continuous variables are categorized and the output is in a binary form, leading to information loss. In [1] Rosenstock et al. developed a way to improve this process by inputting the measurements into an ordinal regression model that indicates the patient’s motor status after the surgery. The motor status is measured based on British medical council research scale (BMRC) which grades muscle strength from 0 to 5.

Our Approach

In this project, our aim is to develop a model that can predict the risk of developing motor function deficit, reduce the information loss and potentially create new biomarkers for risk prediction. This project is in collaboration with Dr. Thomas Picht and Dr. Tizian Rosenstock from the department of neurosurgery in Charite.

 

Figure 1: Brain Tractogram obtained from diffusion imaging coupled with structural MRI and nTMS data provides the necessary information for neurosurgical planning.

Data

Our project partners, the group of Prof. Dr. Thomas Picht have collected the data from 112 tumor patients in the neurosurgery department of Charite, Berlin. The data is comprised of MRI scans, nTMS data, tumor grade, tumor recurrence, motor status pre, post and three month after the surgery. The MRI scans include T1-weighted, T1-weighted with tumor annotation, contrast-enhanced T1-weighted, FLAIR, and diffusion-weighted imaging, (figure 1). The data from nTMS includes TMS positive points located on T1-weighted MRI scans and RMT values in affected and healthy hemispheres.

During the initial processing, the corticospinal tract was extracted from the diffusion MRI scan using the TMS positive points. Then a corticospinal tract annotation was generated on the T1-wighted MRI scan and the tumor-tract distance was manually measured in millimeters. Other measurements generated from the initial processing included the fractional anisotropy (FA) and the apparent diffusion coefficient (ADC). FA represents diffusion asymmetry in each voxel in terms of its’ eigenvalues and ADC measures the magnitude of diffusion within each tissue. These variables have shown association with motor status. Hence, they were provided for the complete corticospinal tract as well as the part of the tract neighboring the tumor. Furthermore, from histological analysis the tumor grade was measured and reported.

Experimental Setup

In the first phase of the project, we focus on developing a neural network model to predict post-operative motor status from the structural MRI. Initially we attempted to train models solely on T1-weighted with contrast agent and without any annotations, which was unsuccessful due to limited data. Subsequently we will include tumor and tract annotations in the training which should reduce the difficulty of the task. The experiments in this phase will be summarized in a publication.

In the second phase, we will create a deep feature for motor status. We will focus on three regions of the neuroimaging data: The tumor, the corticospinal tract and the part of primary motor cortex tagged with TMS positive points. This will require a preprocessing step in which we connect the TMS positive points and generate annotations to be used as part of the model input. Further, the volume of the tumor and the area of the annotation that overlaps with the tumor will be calculated and validated by medical experts. The overlapped area will measure tumor infiltration in the TMS-tagged primary motor cortex. Our hope is that this measurement could represent the motor cortex involvement. The model will be trained on annotations of the corticospinal tract, the tumor and the tagged primary motor cortex. Initially, we will train the model to reconstruct the three annotations and later we will include prediction of pre-operative motor status, tumor-tract distance, tumor infiltration and tumor size. In these experiments, we will use multi-task learning to enforce constrains on the hidden space to be used as deep feature to predict the post-operative motor status.

References

[1] T. Rosenstock, U. Grittner, G. Acker, V. Schwarzer, N. Kulchytska, P. Vajkoczy, and T. Picht. “Risk stratification in motor area–related glioma surgery based on navigated transcranial magnetic stimulation data”. In: Journal of Neurosurgery JNS 126.4 (1Apr. 2017), pages 1227 –1237. doi: 10.3171/2016.4.JNS152896.

Published Research

Kirchler, M.,Khorasani, S.,Kloft, M., Lippert, C. (2019). "Two-sample Testing Using Deep Learning". In: ariXiv: 1910.06239

Konigorski, S., Khorasani S., & Lippert, C. (2018). Intergrating omics and MRI data with kernel-based tests and CNNs to identify rare genetic markers for Alzheimer’s disease. 32nd Conference on Neural Information Processing Systems (NeurIPS), arXiv:1812.00448.

Masoudi, R., Mazaheri-Asadi, L., & Khorasani, S. (2016). Partial and complete microdeletions of Y chromosome in infertile males from South of Iran. Molecular Biology Research Communications, 5(4), 247–255, PMCID: PMC5326488.