An Analysis of the Vulnerability of Two Common Deep Learning-Based Medical Image Segmentation Techniques to Model Inversion Attacks.

Recent research in computer vision has shown that original images used for training of deep learning models can be reconstructed using so-called inversion attacks. However, the feasibility of this attack type has not been investigated for complex 3D medical images. Thus, the aim of this study was to examine the vulnerability of deep learning techniques used in medical imaging to model inversion attacks and investigate multiple quantitative metrics to evaluate the quality of the reconstructed images. For the development and evaluation of model inversion attacks, the public LPBA40 database consisting of 40 brain MRI scans with corresponding segmentations of the gyri and deep grey matter brain structures were used to train two popular deep convolutional neural networks, namely a U-Net and SegNet, and corresponding inversion decoders. Matthews correlation coefficient, the structural similarity index measure (SSIM), and the magnitude of the deformation field resulting from non-linear registration of the original and reconstructed images were used to evaluate the reconstruction accuracy. A comparison of the similarity metrics revealed that the SSIM is best suited to evaluate the reconstruction accuray, followed closely by the magnitude of the deformation field. The quantitative evaluation of the reconstructed images revealed SSIM scores of 0.73±0.12 and 0.61±0.12 for the U-Net and the SegNet, respectively. The qualitative evaluation showed that training images can be reconstructed with some degradation due to blurring but can be correctly matched to the original images in the majority of the cases. In conclusion, the results of this study indicate that it is possible to reconstruct patient data used for training of convolutional neural networks and that the SSIM is a good metric to assess the reconstruction accuracy.

Supervised machine learning tools: a tutorial for clinicians.

In an increasingly data-driven world, artificial intelligence is expected to be a key tool for converting big data into tangible benefits and the healthcare domain is no exception to this. Machine learning aims to identify complex patterns in multi-dimensional data and use these uncovered patterns to classify new unseen cases or make data-driven predictions. In recent years, deep neural networks have shown to be capable of producing results that considerably exceed those of conventional machine learning methods for various classification and regression tasks. In this paper, we provide an accessible tutorial of the most important supervised machine learning concepts and methods, including deep learning, which are potentially the most relevant for the medical domain. We aim to take some of the mystery out of machine learning and depict how machine learning models can be useful for medical applications. Finally, this tutorial provides a few practical suggestions for how to properly design a machine learning model for a generic medical problem.

Stroke Lesion Segmentation in FLAIR MRI Datasets Using Customized Markov Random Fields.

Robust and reliable stroke lesion segmentation is a crucial step toward employing lesion volume as an independent endpoint for randomized trials. The aim of this work was to develop and evaluate a novel method to segment sub-acute ischemic stroke lesions from fluid-attenuated inversion recovery (FLAIR) magnetic resonance imaging (MRI) datasets. After preprocessing of the datasets, a Bayesian technique based on Gabor textures extracted from the FLAIR signal intensities is utilized to generate a first estimate of the lesion segmentation. Using this initial segmentation, a customized voxel-level Markov random field model based on intensity as well as Gabor texture features is employed to refine the stroke lesion segmentation. The proposed method was developed and evaluated based on 151 multi-center datasets from three different databases using a leave-one-patient-out validation approach. The comparison of the automatically segmented stroke lesions with manual ground truth segmentation revealed an average Dice coefficient of 0.582, which is in the upper range of previously presented lesion segmentation methods using multi-modal MRI datasets. Furthermore, the results obtained by the proposed technique are superior compared to the results obtained by two methods based on convolutional neural networks and three phase level-sets, respectively, which performed best in the ISLES 2015 challenge using multi-modal imaging datasets. The results of the quantitative evaluation suggest that the proposed method leads to robust lesion segmentation results using FLAIR MRI datasets only as a follow-up sequence.

IMaGe: Iterative Multilevel Probabilistic Graphical Model for Detection and Segmentation of Multiple Sclerosis Lesions in Brain MRI.

In this paper, we present IMaGe, a new, iterative two-stage probabilistic graphical model for detection and segmentation of Multiple Sclerosis (MS) lesions. Our model includes two levels of Markov Random Fields (MRFs). At the bottom level, a regular grid voxel-based MRF identifies potential lesion voxels, as well as other tissue classes, using local and neighbourhood intensities and class priors. Contiguous voxels of a particular tissue type are grouped into regions. A higher, non-lattice MRF is then constructed, in which each node corresponds to a region, and edges are defined based on neighbourhood relationships between regions. The goal of this MRF is to evaluate the probability of candidate lesions, based on group intensity, texture and neighbouring regions. The inferred information is then propagated to the voxel-level MRF. This process of iterative inference between the two levels repeats as long as desired. The iterations suppress false positives and refine lesion boundaries. The framework is trained on 660 MRI volumes of MS patients enrolled in clinical trials from 174 different centres, and tested on a separate multi-centre clinical trial data set with 535 MRI volumes. All data consists of T1, T2, PD and FLAIR contrasts. In comparison to other MRF methods, such as, and a traditional MRF, IMaGe is much more sensitive (with slightly better PPV). It outperforms its nearest competitor by around 20% when detecting very small lesions (3-10 voxels). This is a significant result, as such lesions constitute around 40% of the total number of lesions.

Probabilistic multiple sclerosis lesion classification based on modeling regional intensity variability and local neighborhood information.

GOAL: In this paper, a fully automatic probabilistic method for multiple sclerosis (MS) lesion classification is presented, whereby the posterior probability density function over healthy tissues and two types of lesions (T1-hypointense and T2-hyperintense) is generated at every voxel. METHODS: During training, the system explicitly models the spatial variability of the intensity distributions throughout the brain by first segmenting it into distinct anatomical regions and then building regional likelihood distributions for each tissue class based on multimodal magnetic resonance image (MRI) intensities. Local class smoothness is ensured by incorporating neighboring voxel information in the prior probability through Markov random fields. The system is tested on two datasets from real multisite clinical trials consisting of multimodal MRIs from a total of 100 patients with MS. Lesion classification results based on the framework are compared with and without the regional information, as well as with other state-of-the-art methods against the labels from expert manual raters. The metrics for comparison include Dice overlap, sensitivity, and positive predictive rates for both voxel and lesion classifications. RESULTS: Statistically significant improvements in Dice values ( ), for voxel-based and lesion-based sensitivity values ( ), and positive predictive rates ( and respectively) are shown when the proposed method is compared to the method without regional information, and to a widely used method [1]. This holds particularly true in the posterior fossa, an area where classification is very challenging. SIGNIFICANCE: The proposed method allows us to provide clinicians with accurate tissue labels for T1-hypointense and T2-hyperintense lesions, two types of lesions that differ in appearance and clinical ramifications, and with a confidence level in the classification, which helps clinicians assess the classification results.

Hierarchical probabilistic Gabor and MRF segmentation of brain tumours in MRI volumes.

In this paper, we present a fully automated hierarchical probabilistic framework for segmenting brain tumours from multispectral human brain magnetic resonance images (MRIs) using multiwindow Gabor filters and an adapted Markov Random Field (MRF) framework. In the first stage, a customised Gabor decomposition is developed, based on the combined-space characteristics of the two classes (tumour and non-tumour) in multispectral brain MRIs in order to optimally separate tumour (including edema) from healthy brain tissues. A Bayesian framework then provides a coarse probabilistic texture-based segmentation of tumours (including edema) whose boundaries are then refined at the voxel level through a modified MRF framework that carefully separates the edema from the main tumour. This customised MRF is not only built on the voxel intensities and class labels as in traditional MRFs, but also models the intensity differences between neighbouring voxels in the likelihood model, along with employing a prior based on local tissue class transition probabilities. The second inference stage is shown to resolve local inhomogeneities and impose a smoothing constraint, while also maintaining the appropriate boundaries as supported by the local intensity difference observations. The method was trained and tested on the publicly available MICCAI 2012 Brain Tumour Segmentation Challenge (BRATS) Database [1] on both synthetic and clinical volumes (low grade and high grade tumours). Our method performs well compared to state-of-the-art techniques, outperforming the results of the top methods in cases of clinical high grade and low grade tumour core segmentation by 40% and 45% respectively.

Evaluating intensity normalization on MRIs of human brain with multiple sclerosis.

Intensity normalization is an important pre-processing step in the study and analysis of Magnetic Resonance Images (MRI) of human brains. As most parametric supervised automatic image segmentation and classification methods base their assumptions regarding the intensity distributions on a standardized intensity range, intensity normalization takes on a very significant role. One of the fast and accurate approaches proposed for intensity normalization is that of Nyul and colleagues. In this work, we present, for the first time, an extensive validation of this approach in real clinical domain where even after intensity inhomogeneity correction that accounts for scanner-specific artifacts, the MRI volumes can be affected from variations such as data heterogeneity resulting from multi-site multi-scanner acquisitions, the presence of multiple sclerosis (MS) lesions and the stage of disease progression in the brain. Using the distributional divergence criteria, we evaluate the effectiveness of the normalization in rendering, under the distributional assumptions of segmentation approaches, intensities that are more homogenous for the same tissue type while simultaneously resulting in better tissue type separation. We also demonstrate the advantage of the decile based piece-wise linear approach on the task of MS lesion segmentation against a linear normalization approach over three image segmentation algorithms: a standard Bayesian classifier, an outlier detection based approach and a Bayesian classifier with Markov Random Field (MRF) based post-processing. Finally, to demonstrate the independence of the effectiveness of normalization from the complexity of segmentation algorithm, we evaluate the Nyul method against the linear normalization on Bayesian algorithms of increasing complexity including a standard Bayesian classifier with Maximum Likelihood parameter estimation and a Bayesian classifier with integrated data priors, in addition to the above Bayesian classifier with MRF based post-processing to smooth the posteriors. In all relevant cases, the observed results are verified for statistical relevance using significance tests.