Project AM-SIT - Data-Driven Analysis of Medical and Simulation Data for Improved Patient Treatment in Rhinology

Project duration

February, 15, 2019 - February, 15, 2023

Project partners

• Jülich Aachen Research Center for Simulation and Data Science (JARA-CSD), FZJ and RWTH Aachen University
• Institute of Aerodynamics and Chair of Fluid Mechanics (AIA), RWTH Aachen University

Funding

The project is jointly funded by RWTH Aachen University and Forschungszentrum Jülich.

Project description

Analyzing patient pathologies and finding optimal treatment strategies are high-dimensional multi-variate problems. The massive amount of involved data ranges from three-dimensional medical images, complex surfaces of nasal cavities extracted from such images, patient histories, medical diagnoses, results of in-vivo maeasurements, subjective patient evaluations derived from, e.g., Rhinosinusitis Disability Index (RSBI) questionnaires, to simulation data and the post-processed results. It is obvious that manual analyses of such data is inefficient and that a sound and robust processing strategy is hard, if not impossible, to derive. To overcome these issues, ML algorithms to analyze patient and simulative data to support the decision process of medical doctors a-prioria surgery and to predict the according surgery outcome will be employed in this project. Thus, the reliability of standardized medical invention methods will be increased. Furthermore, the system may help to improve education in the field of ear-nose, throat (ENT) care.

Data acquisition can be split into three categories:

1. acquisition of medical data in hospital;
2. the acquisition of data in preparation and by means of simulations;
3. the generation of new knowledge with ML techniques.

An associated medical partner holds thousands of CT data sets of pathological nasal cavities and patient meta data and will provide those in anonymized form as input to an ML database. AIA has developed the highly scalable simulation framework Zonal Flow Solver (ZFS) to efficiently predict the flow in the complex human respiratory system with a lattice-Boltzmann method (LBM). In the project Rhinodiagnost, the simulation method is automated and tuned to efficiently run on JSC’s HPC hardware. AIA will extract surfaces of nasal cavities provided by the medical partner and will use its numerical methods to enrich clinical data with simulation data, thus providing enough data to enable training of selected ML algorithms that use state-of-the-art deep learning (DL) tools (i.e. Tensorflow/Keras). The particular DL model to extract hidden features from medical and simulation data for pathology classification and surgery decision support are Convolutional Neural Networks (CNNs). One of the disruptive attractive capabilities of CNNs is ’feature learning’ that enables the DL network to automatically learn features from the image data sets. In contrast, traditional classification approaches such as support vector machines and random forests require a relatively high amount of time-consuming manual feature engineering. Despite this advantage, the high complexity of using CNNs, compared to traditional basic two-layer artificial neural networks (ANNs), relies on choosing the right CNN architecture, i.e., convolutional layer, pooling layer, number of neurons, etc., for the concrete image problems in this project.

Links

Project website at JARA-CSD: AM-SIT

Project publications: