The College of Engineering held its first Virual Research Open House on Thursday, November 5, 2020. Student presenters all prepared a PowerPoint presentation and a three-minute video done in a the style of a Three-Minute Thesis. 

 

Sean Mullan, Graduate Fellow in Computer Science, tied for Best BME graduate presentation, "Using Transparent Deep Learning for Error Detection in Medical Imaging Segmentation".

Authors: Mullan, Sean; Sonka, Milan

Deep learning models have shown remarkable accuracy in many medical image related tasks, but the high cost of failure requires manual checking of automated results. We report a novel machine-learning approach that uses transparent deep learning methods to detect regions of uncertainty in lung segmentations of pulmonary CT volumes. The automatic identification and quantification of these regions can allow for a manual reviewer to quickly identify and correct any potential areas of concern.

Three-Minute Video: https://youtu.be/bChIZNYHBV8

Presentation: mullan_bme_vroh_2020.pdf

 

Lichun Zhang, Graduate Research Assistant in Electrical and Computer Engineering, won best ECE graduate presentation for "Active Learning with FilterNet for Calf Muscle Compartment Segmentation". 

Authors: Zhang, Lichun; Zhihui, Guo; Honghai, Zhang; Eric Axelson; Daniel Thedens; Ellen van der Plas; Peg Nopoulos; Sonka, Milan

With the limited cost for annotation and computation, what instances should be traced to obtain the best performance? We address the question and present a deep active learning framework that combines a novel fully convolutional network (FCN), called FilterNet and active learning to try significantly reducing annotation effort but remains the best performance meanwhile.

Three-Minute Video: https://youtu.be/SIFVEnubbPU

Presentation: zhang_lichun_ece_vroh_2020.pdf

 

 

Other IIBI presenters included:

Chaudhary, Muhammad Faizyab Ali - "CT texture and biomechanical measures for assessing COPD progression"

Authors: Chaudhary, Muhammad F. A.; Pan, Yue; Wang, Di; Bodduluri, Sandeep; Bhatt, Surya P.; Comellas, Alejandro P.; Hoffman, Eric A.; Christensen, Gary E.; Reinhardt, Joseph M.

Chronic Obstructive Pulmonary Disease (COPD) is a heterogeneous collection of airways disease and parenchymal tissue damage. Computed tomography (CT) has become a principal tool for evaluating disease phenotypes. One approach for using CT imaging to assess COPD relies on image registration to match inspiratory and expiratory images to extract biomarkers reflective of local tissue biomechanics. In this study, we propose to evaluate the effectiveness of pulmonary biomechanical features to explain several clinical parameters across four state-of-the-art image registration algorithms. The features would also be used to predict COPD status, severity and exacerbations. Moreover, we propose to evaluate the ability of these feature sets to explain various pulmonary function parameters.

Three-Minute Video: https://youtu.be/IncnumJaLUI

Presentation: chaudhary_bme_vroh_2020.pdf

 

Johnson, Chase - "Preparing the PREDICT-HD dataset for public release through landmark detection and deidentifying data"

Authors: Johnson, Charles

The PREDICT-HD study collected MRI brain scans from participants at risk for Huntington’s Disease from 1999-2016 with the purpose of identifying the disease with medical imaging before symptom expression. Nearing data publication, we have deidentified the data by shifting dates, replacing identifying words/phrases, and defacing images in the dataset to prevent a future researcher from identifying participants. We identified landmarks on each scan using our custom landmark detection software then visually inspected and adjusted any landmarks needing correction. Once published, this neuroimaging dataset will be available to the research community for use in further studies.

Three-Minute Video: https://youtu.be/zdq5D_gHHRI

Presentation: johnson_ece_vroh_2020.pdf

 

Le, Nam Hoang - "LOGISMOS-JEI: Segmentation of Vessel Walls Associated with Brain Aneurysms"

Authors: Le, Nam Hoang; Zhang, Honghai; Hasan, David; Samaniego, Edgar; Derdeyn, Colin; Koscik, Tim; Bathla, Girish; Roa Loor, Jorge; Sonka, Milan

Intracranial aneurysms are swollen parts of blood vessels filled with blood in the brain. In this project, we developed segmentation workflow for the detection of vessel walls which implements the LOGISMOS method for detection of vessel walls. LOGISMOS is a graph search method that optimally segments multiple n-D surfaces which can have mutual geometric relationships. This is achieved by first representing the interested structures by a set of columns filled with graph nodes, then the algorithm tries to find a set of node per columns whose total cost is minimal. The performance of the method and associated software was validated on CT perfusion brain images.

Three-Minute Video: https://youtu.be/uS5eReQUBE0

Presentation: le_ece_vroh_2020.pdf

 

Pramanik, Aniket - "Calibrationless Parallel MRI using Model-Based Deep Learn ing (C-MODL)

Authors: Pramanik, Aniket; Jacob, Mathews

We introduce a fast model-based deep-learning (DL) approach for calibrationless Parallel Magnetic Resonance Image (PMRI) reconstruction. It is a non-linear extension of recent calibrationless structured low-rank methods for PMRI, called PSLR, that rely on linear relations in Fourier domain. The proposed scheme pre-learns non-linear relations in the Fourier domain from exemplar data. It is about three orders of magnitude faster than PSLR methods. A challenge with calibration-based methods is the potential for motion artifacts in images due to mismatches between the calibration and main scans. The proposed calibrationless strategy out-performs the calibrated model-based DL approach MoDL while avoiding mismatches.

Three-Minute Video: https://www.youtube.com/watch?v=hivfmGjwArU&feature=youtu.be

Presentation: pramanik_ece_vroh_2020.pptx

 

 

Siemonsma, Stephen - "Magnetic Resonance Fingerprinting Using Model-Based Deep Learning: Exploiting physics-based and deep-learned priors in a novel iterative reconstruction and quantification algorithm applied to simulated brain MRI data"

Authors: Siemonsma, Stephen; Eldar, Yonina; Jacob, Mathews

Since its introduction, magnetic resonance fingerprinting (MRF) has proven itself to be a versatile and increasingly important quantitative MRI method. Traditional MRF dictionary matching techniques are quickly being superseded with convolutional neural network (CNN) approaches. However, most CNN approaches simply use the network as a black box. In this work, although we include a CNN in our algorithm, the data consistency steps ensure that our results remain reliable even at very high acceleration factors. Overall, we are proposing a novel model-based, unrolled algorithm that simultaneously restores the temporal profiles of the data and accurately estimates the underlying tissue parameter maps.

Three-Minute Video: https://youtu.be/_D_V6bRjt4g

Presentation: siemonsma_ece_vroh_2020.pdf

 

Wang, Di - "Fully-UNet: Unsupervised 3D End-to-End Medical Image Registration for Large Deformation"

Authors: Wang, Di; Durumeric, Oguz; Reinhardt, Joseph; Christensen, Gary

We present a learning-based method for deformable image registration (DIR). Many published learning-based methods have shown promising results to predict the deformation vector field (DVF). However, these methods limit to small deformation tasks. To address this shortcoming, we develop lung DIR method using U-Net-like architecture, namely Fully-UNet, to capture large lung motion between inhale-exhale pulmonary CT images. The network is trained end-to-end by optimization of loss metric between pairs of 3D images. Evaluation was performed on public DIRLAB datasets. The results demonstrate that Fully-UNet has achieved excellent performance in terms of TRE and plausibility of DVF among learning-based methods.

Three-Minute Video: https://youtu.be/OCuablwgvZ8

Presentation: wang_ece_vroh_2020.pdf

 

Zhang, Xiaoliu - "CT-Based Characterization of Transverse and Longitudinal Trabeculae and Its Applications"

Authors: Zhang, Xiaoliu; Letuchy, Elena; Levy, Steven; Torner, James; Saha, Punam

Osteoporosis is characterized by reduced bone mineral density (BMD), micro-structural deterioration, and enhanced fracture-risk. There are compelling evidences suggesting that bone micro-structural quality is a strong determinant of bone strength and fracture-risk. Trabecular bone (Tb) consists of transverse and longitudinal microstructures, and there is a hypothesis that transverse trabeculae improve bone strength by arresting the buckling of longitudinal trabeculae. We present a new in vivo CT-based method for characterizing transverse and longitudinal trabeculae, evaluate their repeat CT scan reproducibility, and examine their links with gender, height, weight, and BMI.

Three-Minute Video: https://youtu.be/_3PjSACbCcE

Presentation: zhang_xiaoliu_ece_vroh_2020.pdf

2020 VROH Presentation Winners