TRD 2: Deep Learning

William Wells Tina Kapur A Fedorov Sarah Frisken
William M. Wells PhD Tina Kapur PhD Andrey Fedorov PhD Sarah Frisken PhD

Led by William Wells, the focus areas of the Deep Learning TRD are:

  • Information theoretic approaches for weakly-supervised deep learning for image registration and segmentation.

  • Infrastructure for curating and organizing image data, annotations, and metadata for deep learning.

  • Application areas: multimodal image registration for MRI and ultrasound guided brain tumor resection, thoracoscopic lung surgery; cancer map prediction from sparse biopsy samples to guide transperineal in-bore MRI-guided prostate biopsy.

Select Publications

Frisken SF. SurfaceNets for Multi-Label Segmentations with Preservation of Sharp Boundaries. J Comput Graph Tech. 2022;11 (1) :34-54.Abstract
We extend 3D SurfaceNets to generate surfaces of segmented 3D medical images composed of multiple materials represented as indexed labels. Our extension generates smooth, high-quality triangle meshes suitable for rendering and tetrahedralization, preserves topology and sharp boundaries between materials, guarantees a user-specified accuracy, and is fast enough that users can interactively explore the trade-off between accuracy and surface smoothness. We provide open-source code in the form of an extendable C++ library with a simple API, and a Qt and OpenGL-based application that allows users to import or randomly generate multi-label volumes to experiment with surface fairing parameters. In this paper, we describe the basic SurfaceNets algorithm, our extension to handle multiple materials, our method for preserving sharp boundaries between materials, and implementation details used to achieve efficient processing.
Sack J, Nitsch J, Meine H, Kikinis R, Halle M, Rutherford A. Quantitative Analysis of Liver Disease Using MRI-Based Radiomic Features of the Liver and Spleen. J Imaging. 2022;8 (10) :277.Abstract
BACKGROUND: Radiomics extracts quantitative image features to identify biomarkers for characterizing disease. Our aim was to characterize the ability of radiomic features extracted from magnetic resonance (MR) imaging of the liver and spleen to detect cirrhosis by comparing features from patients with cirrhosis to those without cirrhosis. METHODS: This retrospective study compared MR-derived radiomic features between patients with cirrhosis undergoing hepatocellular carcinoma screening and patients without cirrhosis undergoing intraductal papillary mucinous neoplasm surveillance between 2015 and 2018 using the same imaging protocol. Secondary analyses stratified the cirrhosis cohort by liver disease severity using clinical compensation/decompensation and Model for End-Stage Liver Disease (MELD). RESULTS: Of 167 patients, 90 had cirrhosis with 68.9% compensated and median MELD 8. Combined liver and spleen radiomic features generated an AUC 0.94 for detecting cirrhosis, with shape and texture components contributing more than size. Discrimination of cirrhosis remained high after stratification by liver disease severity. CONCLUSIONS: MR-based liver and spleen radiomic features had high accuracy in identifying cirrhosis, after stratification by clinical compensation/decompensation and MELD. Shape and texture features performed better than size features. These findings will inform radiomic-based applications for cirrhosis diagnosis and severity.
Bridge CP, Gorman C, Pieper S, Doyle SW, Lennerz JK, Kalpathy-Cramer J, Clunie DA, Fedorov AY, Herrmann MD. Highdicom: a Python Library for Standardized Encoding of Image Annotations and Machine Learning Model Outputs in Pathology and Radiology. J Digit Imaging. 2022.Abstract
Machine learning (ML) is revolutionizing image-based diagnostics in pathology and radiology. ML models have shown promising results in research settings, but the lack of interoperability between ML systems and enterprise medical imaging systems has been a major barrier for clinical integration and evaluation. The DICOM® standard specifies information object definitions (IODs) and services for the representation and communication of digital images and related information, including image-derived annotations and analysis results. However, the complexity of the standard represents an obstacle for its adoption in the ML community and creates a need for software libraries and tools that simplify working with datasets in DICOM format. Here we present the highdicom library, which provides a high-level application programming interface (API) for the Python programming language that abstracts low-level details of the standard and enables encoding and decoding of image-derived information in DICOM format in a few lines of Python code. The highdicom library leverages NumPy arrays for efficient data representation and ties into the extensive Python ecosystem for image processing and machine learning. Simultaneously, by simplifying creation and parsing of DICOM-compliant files, highdicom achieves interoperability with the medical imaging systems that hold the data used to train and run ML models, and ultimately communicate and store model outputs for clinical use. We demonstrate through experiments with slide microscopy and computed tomography imaging, that, by bridging these two ecosystems, highdicom enables developers and researchers to train and evaluate state-of-the-art ML models in pathology and radiology while remaining compliant with the DICOM standard and interoperable with clinical systems at all stages. To promote standardization of ML research and streamline the ML model development and deployment process, we made the library available free and open-source at .
Yu Y, Safdar S, Bourantas G, Zwick B, Joldes G, Kapur T, Frisken S, Kikinis R, Nabavi A, Golby A, et al. Automatic Framework for Patient-Specific Modelling of Tumour Resection-Induced Brain Shift. Comput Biol Med. 2022;143 :105271.Abstract
Our motivation is to enable non-biomechanical engineering specialists to use sophisticated biomechanical models in the clinic to predict tumour resection-induced brain shift, and subsequently know the location of the residual tumour and its boundary. To achieve this goal, we developed a framework for automatically generating and solving patient-specific biomechanical models of the brain. This framework automatically determines patient-specific brain geometry from MRI data, generates patient-specific computational grid, assigns material properties, defines boundary conditions, applies external loads to the anatomical structures, and solves differential equations of nonlinear elasticity using Meshless Total Lagrangian Explicit Dynamics (MTLED) algorithm. We demonstrated the effectiveness and appropriateness of our framework on real clinical cases of tumour resection-induced brain shift.
Giganti F, Cole AP, Fennessy FM, Clinton T, Moreira PLDF, Bernardes MC, Westin C-F, Krishnaswamy D, Fedorov A, Wollin DA, et al. Promoting the Use of the PI-QUAL Score for Prostate MRI Quality: Results From the ESOR Nicholas Gourtsoyiannis Teaching Fellowship. Eur Radiol. 2022 :1-11.Abstract
OBJECTIVES: The Prostate Imaging Quality (PI-QUAL) score is a new metric to evaluate the diagnostic quality of multiparametric magnetic resonance imaging (MRI) of the prostate. This study assesses the impact of an intervention, namely a prostate MRI quality training lecture, on the participant's ability to apply PI-QUAL. METHODS: Sixteen participants (radiologists, urologists, physicists, and computer scientists) of varying experience in reviewing diagnostic prostate MRI all assessed the image quality of ten examinations from different vendors and machines. Then, they attended a dedicated lecture followed by a hands-on workshop on MRI quality assessment using the PI-QUAL score. Five scans assessed by the participants were evaluated in the workshop using the PI-QUAL score for teaching purposes. After the course, the same participants evaluated the image quality of a new set of ten scans applying the PI-QUAL score. Results were assessed using receiver operating characteristic analysis. The reference standard was the PI-QUAL score assessed by one of the developers of PI-QUAL. RESULTS: There was a significant improvement in average area under the curve for the evaluation of image quality from baseline (0.59 [95 % confidence intervals: 0.50-0.66]) to post-teaching (0.96 [0.92-0.98]), an improvement of 0.37 [0.21-0.41] (p < 0.001). CONCLUSIONS: A teaching course (dedicated lecture + hands-on workshop) on PI-QUAL significantly improved the application of this scoring system to assess the quality of prostate MRI examinations. KEY POINTS: • A significant improvement in the application of PI-QUAL for the assessment of prostate MR image quality was observed after an educational intervention. • Appropriate training on image quality can be delivered to those involved in the acquisition and interpretation of prostate MRI. • Further investigation will be needed to understand the impact on improving the acquisition of high-quality diagnostic prostate MR examinations.
McGarry SD, Brehler M, Bukowy JD, Lowman AK, Bobholz SA, Duenweg SR, Banerjee A, Hurrell SL, Malyarenko D, Chenevert TL, et al. Multi-Site Concordance of Diffusion-Weighted Imaging Quantification for Assessing Prostate Cancer Aggressiveness. J Magn Reson Imaging. 2022;55 (6) :1745-58.Abstract
BACKGROUND: Diffusion-weighted imaging (DWI) is commonly used to detect prostate cancer, and a major clinical challenge is differentiating aggressive from indolent disease. PURPOSE: To compare 14 site-specific parametric fitting implementations applied to the same dataset of whole-mount pathologically validated DWI to test the hypothesis that cancer differentiation varies with different fitting algorithms. STUDY TYPE: Prospective. POPULATION: Thirty-three patients prospectively imaged prior to prostatectomy. FIELD STRENGTH/SEQUENCE: 3 T, field-of-view optimized and constrained undistorted single-shot DWI sequence. ASSESSMENT: Datasets, including a noise-free digital reference object (DRO), were distributed to the 14 teams, where locally implemented DWI parameter maps were calculated, including mono-exponential apparent diffusion coefficient (MEADC), kurtosis (K), diffusion kurtosis (DK), bi-exponential diffusion (BID), pseudo-diffusion (BID*), and perfusion fraction (F). The resulting parametric maps were centrally analyzed, where differentiation of benign from cancerous tissue was compared between DWI parameters and the fitting algorithms with a receiver operating characteristic area under the curve (ROC AUC). STATISTICAL TEST: Levene's test, P < 0.05 corrected for multiple comparisons was considered statistically significant. RESULTS: The DRO results indicated minimal discordance between sites. Comparison across sites indicated that K, DK, and MEADC had significantly higher prostate cancer detection capability (AUC range = 0.72-0.76, 0.76-0.81, and 0.76-0.80 respectively) as compared to bi-exponential parameters (BID, BID*, F) which had lower AUC and greater between site variation (AUC range = 0.53-0.80, 0.51-0.81, and 0.52-0.80 respectively). Post-processing parameters also affected the resulting AUC, moving from, for example, 0.75 to 0.87 for MEADC varying cluster size. DATA CONCLUSION: We found that conventional diffusion models had consistent performance at differentiating prostate cancer from benign tissue. Our results also indicated that post-processing decisions on DWI data can affect sensitivity and specificity when applied to radiological-pathological studies in prostate cancer. LEVEL OF EVIDENCE: 1 TECHNICAL EFFICACY: Stage 3.
Zhang F, Wells WM, O'Donnell LJ. Deep Diffusion MRI Registration (DDMReg): A Deep Learning Method for Diffusion MRI Registration. IEEE Trans Med Imaging. 2022;41 (6) :1454-67.Abstract
In this paper, we present a deep learning method, DDMReg, for accurate registration between diffusion MRI (dMRI) datasets. In dMRI registration, the goal is to spatially align brain anatomical structures while ensuring that local fiber orientations remain consistent with the underlying white matter fiber tract anatomy. DDMReg is a novel method that uses joint whole-brain and tract-specific information for dMRI registration. Based on the successful VoxelMorph framework for image registration, we propose a novel registration architecture that leverages not only whole brain information but also tract-specific fiber orientation information. DDMReg is an unsupervised method for deformable registration between pairs of dMRI datasets: it does not require nonlinearly pre-registered training data or the corresponding deformation fields as ground truth. We perform comparisons with four state-of-the-art registration methods on multiple independently acquired datasets from different populations (including teenagers, young and elderly adults) and different imaging protocols and scanners. We evaluate the registration performance by assessing the ability to align anatomically corresponding brain structures and ensure fiber spatial agreement between different subjects after registration. Experimental results show that DDMReg obtains significantly improved registration performance compared to the state-of-the-art methods. Importantly, we demonstrate successful generalization of DDMReg to dMRI data from different populations with varying ages and acquired using different acquisition protocols and different scanners.
Wang S, Zhang F, Huang P, Hong H, Jiaerken Y, Yu X, Zhang R, Zeng Q, Zhang Y, Kikinis R, et al. Superficial White Matter Microstructure Affects Processing Speed in Cerebral Small Vessel Disease. Hum Brain Mapp. 2022;43 (17) :5310-25.Abstract
White matter hyperintensities (WMH) are a typical feature of cerebral small vessel disease (CSVD), which contributes to about 50% of dementias worldwide. Microstructural alterations in deep white matter (DWM) have been widely examined in CSVD. However, little is known about abnormalities in superficial white matter (SWM) and their relevance for processing speed, the main cognitive deficit in CSVD. In 141 CSVD patients, processing speed was assessed using Trail Making Test Part A. White matter abnormalities were assessed by WMH burden (volume on T2-FLAIR) and diffusion MRI measures. SWM imaging measures had a large contribution to processing speed, despite a relatively low SWM WMH burden. Across all imaging measures, SWM free water (FW) had the strongest association with processing speed, followed by SWM mean diffusivity (MD). SWM FW was the only marker to significantly increase between two subgroups with the lowest WMH burdens. When comparing two subgroups with the highest WMH burdens, the involvement of WMH in the SWM was accompanied by significant differences in processing speed and white matter microstructure. Mediation analysis revealed that SWM FW fully mediated the association between WMH volume and processing speed, while no mediation effect of MD or DWM FW was observed. Overall, results suggest that the SWM has an important contribution to processing speed, while SWM FW is a sensitive imaging marker associated with cognition in CSVD. This study extends the current understanding of CSVD-related dysfunction and suggests that the SWM, as an understudied region, can be a potential target for monitoring pathophysiological processes.
Pujol S, Cabeen RP, Yelnik J, François C, Fernandez Vidal S, Karachi C, Bardinet E, Cosgrove RG, Kikinis R. Somatotopic Organization of Hyperdirect Pathway Projections From the Primary Motor Cortex in the Human Brain. Front Neurol. 2022;13 :791092.Abstract
Background: The subthalamic nucleus (STN) is an effective neurosurgical target to improve motor symptoms in Parkinson's Disease (PD) patients. MR-guided Focused Ultrasound (MRgFUS) subthalamotomy is being explored as a therapeutic alternative to Deep Brain Stimulation (DBS) of the STN. The hyperdirect pathway provides a direct connection between the cortex and the STN and is likely to play a key role in the therapeutic effects of MRgFUS intervention in PD patients. Objective: This study aims to investigate the topography and somatotopy of hyperdirect pathway projections from the primary motor cortex (M1). Methods: We used advanced multi-fiber tractography and high-resolution diffusion MRI data acquired on five subjects of the Human Connectome Project (HCP) to reconstruct hyperdirect pathway projections from M1. Two neuroanatomy experts reviewed the anatomical accuracy of the tracts. We extracted the fascicles arising from the trunk, arm, hand, face and tongue area from the reconstructed pathways. We assessed the variability among subjects based on the fractional anisotropy (FA) and mean diffusivity (MD) of the fibers. We evaluated the spatial arrangement of the different fascicles using the Dice Similarity Coefficient (DSC) of spatial overlap and the centroids of the bundles. Results: We successfully reconstructed hyperdirect pathway projections from M1 in all five subjects. The tracts were in agreement with the expected anatomy. We identified hyperdirect pathway fascicles projecting from the trunk, arm, hand, face and tongue area in all subjects. Tract-derived measurements showed low variability among subjects, and similar distributions of FA and MD values among the fascicles projecting from different M1 areas. We found an anterolateral somatotopic arrangement of the fascicles in the corona radiata, and an average overlap of 0.63 in the internal capsule and 0.65 in the zona incerta. Conclusion: Multi-fiber tractography combined with high-resolution diffusion MRI data enables the identification of the somatotopic organization of the hyperdirect pathway. Our preliminary results suggest that the subdivisions of the hyperdirect pathway projecting from the trunk, arm, hand, face, and tongue motor area are intermixed at the level of the zona incerta and posterior limb of the internal capsule, with a predominantly overlapping topographical organization in both regions. Subject-specific knowledge of the hyperdirect pathway somatotopy could help optimize target definition in MRgFUS intervention.
Cole AP, Langbein BJ, Giganti F, Fennessy FM, Tempany CM, Emberton M. Is Perfect the Enemy of Good? Weighing the Evidence for Biparametric MRI in Prostate Cancer. Br J Radiol. 2022;95 (1131) :20210840.Abstract
The role of multiparametric MRI in diagnosis, staging and treatment planning for prostate cancer is well established. However, there remain several challenges to widespread adoption. One such challenge is the duration and cost of the examination. Abbreviated exams omitting contrast-enhanced sequences may help address this challenge. In this review, we will discuss the rationale for biparametric MRI for detection and characterization of clinically significant prostate cancer prior to biopsy and synthesize the published literature. We will weigh up the advantages and disadvantages to this approach and lay out a conceptual cost/benefit analysis regarding adoption of biparametric MRI.