Regional Manifold Learning for Disease Classification

1994

Inhomogeneities in the fields of magnetic resonance (MR) systems cause the statistical characteristics of tissue classes to vary within the resulting MR images. These inhomogeneities must be taken into consideration when designing an algorithm for automated tissue classification. The traditional approach in image processing would be to apply a gain field correction technique to remove the inhomogeneities from the images. Statistical solutions would most likely focus on including spatial information in the feature space of the classifier so that it can be trained to model and adjust for the inhomogeneities. This paper will prove that neither of these general approaches offer a complete and viable solution. This paper will prove that neither of these general approaches offers a complete and viable solution. This paper will in fact show that not only do the inhomogeneities modify the local mean and variance of a tissue class as is commonly accepted, but the inhomogeneities also induce a rotation of the covariance matrices. As a result, gain field correction techniques cannot compensate for all of the artifacts associated with inhomogeneities. Additionally, it will be demonstrated that while statistical methods can capture all of the anomalies, the across patient and across time variations of the inhomogeneities necessitate frequent and time consuming retraining of any Bayesian classifier. This paper introduces a two stage process for MR tissue classification which addresses both of these issues by utilizing techniques from both image processing and statistics. First, a band-pass mean field corrector is used to alleviate the mean and variance deformations in each image. Then, using a kernel mixture model classifier couple to an interactive data augmentation tool, the user can selectively refine and explore the class representations for localized regions of the image and thereby capture the rotation of the covariance matrices. This approach is shown to outperform Gaussian classifiers and 4D mixture modeling techniques when both the final accuracy and user time requirements are considered.

Visualization in Biomedical Computing 1994, 1994

Inhomogeneities in the fields of magnetic resonance (MR) systems cause the statistical characteristics of tissue classes to vary within the resulting MR images. These inhomogeneities must be taken into consideration when designing an algorithm for automated tissue classification. The traditional approach in image processing would be to apply a gain field correction technique to remove the inhomogeneities from the images. Statistical solutions would most likely focus on including spatial information in the feature space of the classifier so that it can be trained to model and adjust for the inhomogeneities. This paper will prove that neither of these general approaches offers a complete and viable solution. This paper will in fact show that not only do the inhomogeneities modify the local mean and variance of a tissue class as is commonly accepted, but the inhomogeneities also induce a rotation of the covariance matrices. As a result, gain field correction techniques cannot compensate for all of the artifacts associated with inhomogeneities. Additionally, it will be demonstrated that while statistical methods can capture all of the anomalies, the across patient and across time variations of the inhomogeneities necessitate frequent and time consuming retraining of any Bayesian classifier. This paper introduces a two stage process for MR tissue classification which addresses both of these issues by utilizing techniques from image processing and statistics. First, a band-pass mean field corrector is used to compensate for the mean and variance deformations in each image. Then, using a kernel mixture model classifier coupled to an interactive data augmentation tool, the user can selectively refine and explore the class representations for localized regions of the image and thereby capture the rotation of the covariance matrices. This approach is shown to outperform Gaussian classifiers and 4D mixture modeling techniques when both the final accuracy and user time requirements are considered.

2010

MR image data can provide many features or measures although any single measure is unlikely to comprehensively characterize the underlying morphology. We present a framework in which multiple measures are used in manifold learning steps to generate coordinate embeddings which are then combined to give an improved single representation of the population. An application to neonatal brain MRI data shows that the use of shape and appearance measures in particular leads to biologically plausible and consistent representations correlating well with clinical data. Orthogonality among the correlations suggests the embedding components relate to comparatively independent morphological features. The rapid changes that occur in brain shape and in MR image appearance during neonatal brain development justify the use of shape measures (obtained from a deformation metric) and appearance measures (obtained from image similarity). The benefit of combining separate embeddings is demonstrated by improved correlations with clinical data and we illustrate the potential of the proposed framework in characterizing trajectories of brain development.

Analytics

For robust classification, selecting a proper classifier is of primary importance. However, selecting the best classifiers depends on the problem, as some classifiers work better at some tasks than on others. Despite the many results collected in the literature, the support vector machine (SVM) remains the leading adopted solution in many domains, thanks to its ease of use. In this paper, we propose a new method based on convolutional neural networks (CNNs) as an alternative to SVM. CNNs are specialized in processing data in a grid-like topology that usually represents images. To enable CNNs to work on different data types, we investigate reshaping one-dimensional vector representations into two-dimensional matrices and compared different approaches for feeding standard CNNs using two-dimensional feature vector representations. We evaluate the different techniques proposing a heterogeneous ensemble based on three classifiers: an SVM, a model based on random subspace of rotation boos...

Lecture Notes in Computer Science, 2015

We present BundleMAP, a novel method for extracting features from diffusion MRI (dMRI), which can be used to detect disease with supervised classification. BundleMAP uses manifold learning to aggregate measurements over localized segments of nerve fiber bundles, which are natural anatomical units in this data. We obtain a fully integrated machine learning pipeline by combining this idea with mechanisms for outlier removal and feature selection. We demonstrate that it increases accuracy on a clinical dataset for which classification results have been reported previously, and that it pinpoints the anatomical locations relevant to the classification.

2010

We describe the medical image classification task in ImageCLEF 2005ImageCLEF -2009. It evolved from a classification task with 57 classes on a total of 10,000 images into a hierarchical classification task with a very large number of potential classes. Here, we describe how the database and the objectives changed over the years and how state-of-the-art approaches from machine learning and computer vision were shown to outperform the nearest neighbor-based classification schemes working on full-image descriptors that were very successful in 2005. In particular the use of discriminative classification methods such as support vector machines and the use of local image descriptors were empirically shown to be important building blocks for medical image classification.

2007

Diffusion tensor magnetic resonance imaging (DT-MRI) provides a comprehensive characterization of white matter (WM) in the brain and therefore, plays a crucial role in the investigation of diseases in which WM is suspected to be compromised such as multiple sclerosis and neuropsychiatric disorders like schizophrenia. However changes induced by pathology may be subtle and affected regions of the brain can only be revealed by a group-based analysis of patients in comparison with healthy controls. This in turn requires voxel-based statistical analysis of spatially normalized brain DT images, as in the case of conventional MR images. However this process is rendered extremely challenging in DT-MRI due to the high dimensionality of the data and its inherent non-linearity that causes linear component analysis methods to be inapplicable. We therefore propose a novel framework for the statistical analysis of DT-MRI data using manifold-based techniques such as isomap and kernel PCA that determine the underlying manifold structure of the data, embed it to a manifold and help perform high dimensional statistics on the manifold to determine regions of difference between the groups of patients and controls. The framework has been successfully applied to DT-MRI data from patients with schizophrenia, as well as to study developmental changes in small animals, both of which identify regional changes, indicating the need for manifold-based methods for the statistical analysis of DTI.

IBM Journal of Research and Development, 2015

In this work, we study the performance of a two-stage ensemble visual machine learning framework for classification of medical images. In the first stage, models are built for subsets of features and data, and in the second stage, models are combined. We demonstrate the performance of this framework in four contexts: 1) The public ImageCLEF (Cross Language Evaluation Forum) 2013 medical modality recognition benchmark, 2) echocardiography view and mode recognition, 3) dermatology disease recognition across two datasets, and 4) a broad medical image dataset, merged from multiple data sources into a collection of 158 categories covering both general and specific medical conceptsVincluding modalities, body regions, views, and disease states. In the first context, the presented system achieves state-of-art performance of 82.2% multiclass accuracy. In the second context, the system attains 90.48% multiclass accuracy. In the third, state-of-art performance of 90% specificity and 90% sensitivity is obtained on a small standardized dataset of 200 images using a leave-one-out strategy. For a larger dataset of 2,761 images, 95% specificity and 98% sensitivity is obtained on a 20% held-out test set. Finally, in the fourth context, the system achieves sensitivity and specificity of 94.7% and 98.4%, respectively, demonstrating the ability to generalize over domains.

Research Square (Research Square), 2021

a less explored topic, and segmentation steps are often a prerequisite for diagnosis models to be accurate. The data scarcity problem is even more prominent in this case, as the ground truth information for the presence of disease is of a much more sensitive nature, and clinical registries are lacking. Label imbalance and demographic variabilities become more relevant as well, as different institutions usually contain different types of diseases, while in segmentation the expected ground truth is the same in all cases-i.e. a mask of the segmented parts. From a data science standpoint, segmentation allows for more flexible methods of data augmentation (even GAN-based generation), whereas, in diagnosis tasks, augmentation should be done with extreme care to avoid shifting the true value of corresponding labels. This often requires human-expert validation depending on the augmentation method, lest it risks adding noise. Segmentation also has the advantage of freely using 2D slices and patch-based approaches, which allows one to extract many training samples from a single scan, while in diagnosis, a single 3D volume (or sometimes series of longitudinal data) is used as a single training data point. For these reasons, related research in domains other than segmentation has been more limited, with studies tangential to diagnosis popping up only as early as 2020 in breast density classification 13 , and lung tumor survival prediction 14. With the emergence of the COVID-19 pandemic, the urgent need for diagnostic tools circumvented common obstacles and allowed researchers to collaborate in federated learning diagnosis for the first time 15,16. Very recently, an open-source framework that integrates FL along with the functionality of end-to-end encryption to protect against inversion attacks was developed, trained and tested on paediatric X-ray classification 17. In this work, we focus on the diagnosis of cardiovascular disease (CVD) based on Cardiac MRI. The importance of furthering our understanding on the structure and function of the heart is highlighted by the prevalence of CVD in the population, which accounts for one third of annual deaths 18,19. Cardiac MRI has been the go-to modality for this task, allowing the assessment and delineation of the three heart segments-i.e. the myocardium, and the left and right ventricle blood pools-to identify the presence of anomalies such as myocardic infarctions or cardiomyopathies. Based on this modality, and by leveraging hand crafted features, Machine Learning (ML) diagnostic tools have been developed with some success in single-center datasets 20,21. To enable AI studies in CMR diagnosis, there have been public challenges for cardiac MRI segmentation and diagnosis such as "Automatic Cardiac Diagnosis Challenge" dataset (ACDC) and the "Multi-Centre, Multi-Vendor and Multi-Disease Cardiac Segmentation Challenge" dataset (M&M). Despite it's name, M&M now also includes diagnostic labels. We will be using subsets from these datasets in our study and provide more details on the data acquisition and annotation in "Dataset" section. In terms of the Cardiac MRI modality, there has been a lot of literature on deep learning-based segmentation 22-30 with diagnosis typically being a follow-up step, leveraging the segmentation masks and utilizing models such as random forests 31,32 , support vector machines 33 or a simple diagnostic rule 34. These diagnosis models focus on two timepoints of the cardiac MRI per-patient: the phase of End-Diastole (ED) (maximum heart relaxation) and the phase of End-Systole (ES) (maximum heart contraction). Current such studies have emphasized on the aforementioned ACDC dataset 35 , a single center dataset hosting 100 subjects and five labels (20 subjects each). Khened et al. 31 reported promising results on multi-label diagnosis but evaluated on a limited hold-out test set of 10 samples. On the same task, Cetin et al. 33 used an SVM on top of radiomic features derived from manual segmentations, while Wolterink et al. 32 evaluated a random forest classifier instead, both using a cross validation scheme. Liu et al. 34 used an automated deep-learning based segmentation scheme, following up with a diagnostic rule. Despite the impressive classification performance reported in these studies, these models are both trained and evaluated solely on a single-center dataset (ACDC), and thus, no assumptions can be made regarding their generalization to unseen centers and larger datasets. To deploy such models in the real world, one has to assume that new subjects being tested are IID to those seen during training and evaluation. In the domain of medical imaging, where data is highly heterogeneous between centers, this assumption is far from true 36. Furthermore, as these studies focus on a single center, no privacy preservation measures are studied, which are otherwise necessary for deployment of such models. Despite the widespread interest in automated CMR diagnosis methods, multi-centric and distributed learning studies in the field are currently lacking. In this paper, we conduct our study with four centers, three of which were derived from the M&M dataset 37 , and the fourth being a subset of ACDC 35. We test the CDS and FL collaborative learning frameworks. As FL trains local models in each center, multi-label classification is a very challenging problem in the case of the M&M dataset where many labels have little to no overlap between the centers, 3D CNN Central Server Models trained locally in each center Central Server 3D CNN Data collected in a single server A) CollaboraƟve Data Sharing (CDS) B) Federated Learning (FL)

Journal of Imaging

Cardiac cine magnetic resonance imaging (MRI) is a widely used technique for the noninvasive assessment of cardiac functions. Deep neural networks have achieved considerable progress in overcoming various challenges in cine MRI analysis. However, deep learning models cannot be used for classification because limited cine MRI data are available. To overcome this problem, features from cine image settings are derived by handcrafting and addition of other clinical features to the classical machine learning approach for ensuring the model fits the MRI device settings and image parameters required in the analysis. In this study, a novel method was proposed for classifying heart disease (cardiomyopathy patient groups) using only segmented output maps. In the encoder–decoder network, the fully convolutional EfficientNetB5-UNet was modified to perform the semantic segmentation of the MRI image slice. A two-dimensional thickness algorithm was used to combine the segmentation outputs for the ...