Caroline Petitjean

A review of segmentation methods in short axis cardiac MR images

For the last 15 years, Magnetic Resonance Imaging (MRI) has become a reference examination for cardiac morphology, function and perfusion in humans. Yet, due to the characteristics of cardiac MR images and to the great variability of the images among patients, the problem of heart cavities segmentation in MRI is still open. This paper is a review of fully and semi-automated methods performing segmentation in short axis images using a cardiac cine MRI sequence. Medical background and specific segmentation difficulties associated to these images are presented. For this particularly complex segmentation task, prior knowledge is required. We thus propose an original categorization for cardiac segmentation methods, with a special emphasis on what level of external information is required (weak or strong) and how it is used to constrain segmentation. After reviewing method principles and analyzing segmentation results, we conclude with a discussion and future trends in this field regarding methodological and medical issues.

A Dataset for Breast Cancer Histopathological Image Classification

Today, medical image analysis papers require solid experiments to prove the usefulness of proposed methods. However, experiments are often performed on data selected by the researchers, which may come from different institutions, scanners, and populations. Different evaluation measures may be used, making it difficult to compare the methods. In this paper, we introduce a dataset of 7909 breast cancer histopathology images acquired on 82 patients, which is now publicly available from http://web.inf.ufpr.br/vri/breast-cancer-database. The dataset includes both benign and malignant images. The task associated with this dataset is the automated classification of these images in two classes, which would be a valuable computer-aided diagnosis tool for the clinician. In order to assess the difficulty of this task, we show some preliminary results obtained with state-of-the-art image classification systems. The accuracy ranges from 80% to 85%, showing room for improvement is left. By providing this dataset and a standardized evaluation protocol to the scientific community, we hope to gather researchers in both the medical and the machine learning field to advance toward this clinical application.

Breast cancer histopathological image classification using Convolutional Neural Networks

The performance of most conventional classification systems relies on appropriate data representation and much of the efforts are dedicated to feature engineering, a difficult and time-consuming process that uses prior expert domain knowledge of the data to create useful features. On the other hand, deep learning can extract and organize the discriminative information from the data, not requiring the design of feature extractors by a domain expert. Convolutional Neural Networks (CNNs) are a particular type of deep, feedforward network that have gained attention from research community and industry, achieving empirical successes in tasks such as speech recognition, signal processing, object recognition, natural language processing and transfer learning. In this paper, we conduct some preliminary experiments using the deep learning approach to classify breast cancer histopathological images from BreaKHis, a publicly dataset available at http://web.inf.ufpr.br/vri/breast-cancer-database. We propose a method based on the extraction of image patches for training the CNN and the combination of these patches for final classification. This method aims to allow using the high-resolution histopathological images from BreaKHis as input to existing CNN, avoiding adaptations of the model that can lead to a more complex and computationally costly architecture. The CNN performance is better when compared to previously reported results obtained by other machine learning models trained with hand-crafted textural descriptors. Finally, we also investigate the combination of different CNNs using simple fusion rules, achieving some improvement in recognition rates.

One class random forests

One class classification is a binary classification task for which only one class of samples is available for learning. In some preliminary works, we have proposed One Class Random Forests (OCRF), a method based on a random forest algorithm and an original outlier generation procedure that makes use of classifier ensemble randomization principles. In this paper, we propose an extensive study of the behavior of OCRF, that includes experiments on various UCI public datasets and comparison to reference one class namely, Gaussian density models, Parzen estimators, Gaussian mixture models and One Class SVMs-with statistical significance. Our aim is to show that the randomization principles embedded in a random forest algorithm make the outlier generation process more efficient, and allow in particular to break the curse of dimensionality. One Class Random Forests are shown to perform well in comparison to other methods, and in particular to maintain stable performance in higher dimension, while the other algorithms may fail.

Cardiac MRI assessment of right ventricular function in acquired heart disease: factors of variability.

RATIONALE AND OBJECTIVES To evaluate intra- and inter-observer variability of right ventricular (RV) functional parameters as evaluated by cardiac magnetic resonance imaging (MRI) in patients with acquired heart disease (AHD), and to identify factors associated with an increased variability. MATERIALS AND METHODS Sixty consecutive patients were enrolled. Right and left ventricular (LV) volumes, ejection fraction, and mass were determined from short-axis cine sequences. All analyzes were performed twice by three observers with various training-degree in cardiac MRI. Intra- and inter-observer variability was evaluated. The impact on variability of each of the following parameters was assessed: observers experience, basal and apical slices selection, end-systolic phase selection, and delineation. RESULTS Mean segmentation time ranged 9.8-19.0 minutes for RV and 6.4-9.2 minutes for LV. Variability of RV functional parameters measurement was strongly influenced by previous observers experience: it was two to three times superior to that of LV, even for the most experienced observer. High variability in the measurement of RV mass was observed. For both ventricles, selection of the basal slice and delineation were major determinants of variability. CONCLUSION As compared to LV, RV function assessment with cardiac MRI in AHD patients is much more variable and time-consuming. Observers experience, selection of basal slice, and delineation are determinant.

Automatic cardiac ventricle segmentation in MR images: a validation study.

AbstractPurposeSegmenting the cardiac ventricles in magnetic resonance (MR) images is required for cardiac function assessment. Numerous segmentation methods have been developed and applied to MR ventriculography. Quantitative validation of these segmentation methods with ground truth is needed prior to clinical use, but requires manual delineation of hundreds of images. We applied a well-established method to this problem and rigorously validated the results.MethodsAn automatic method based on active contours without edges was used for left and the right ventricle cavity segmentation. A large database of 1,920 MR images obtained from 59 patients who gave informed consent was evaluated. Two standard metrics were used for quantitative error measurement.ResultsSegmentation results are comparable to previously reported values in the literature. Since different points in the cardiac cycle and different slice levels were used in this study, a detailed error analysis is possible. Better performance was obtained at end diastole than at end systole, and on mid-ventricular slices than apical slices. Localization of segmentation errors were highlighted through a study of their spatial distribution.ConclusionsVentricular segmentation based on region-driven active contours provided satisfactory results in MRI, without the use of a priori knowledge. The study of error distribution allows identification of potential improvements in algorithm performance.

Medical Image Synthesis with Context-Aware Generative Adversarial Networks

Computed tomography (CT) is critical for various clinical applications, e.g., radiation treatment planning and also PET attenuation correction in MRI/PET scanner. However, CT exposes radiation during acquisition, which may cause side effects to patients. Compared to CT, magnetic resonance imaging (MRI) is much safer and does not involve radiations. Therefore, recently researchers are greatly motivated to estimate CT image from its corresponding MR image of the same subject for the case of radiation planning. In this paper, we propose a data-driven approach to address this challenging problem. Specifically, we train a fully convolutional network (FCN) to generate CT given the MR image. To better model the nonlinear mapping from MRI to CT and produce more realistic images, we propose to use the adversarial training strategy to train the FCN. Moreover, we propose an image-gradient-difference based loss function to alleviate the blurriness of the generated CT. We further apply Auto-Context Model (ACM) to implement a context-aware generative adversarial network. Experimental results show that our method is accurate and robust for predicting CT images from MR images, and also outperforms three state-of-the-art methods under comparison.

Diagnostic accuracy and variability of three semi-quantitative methods for assessing right ventricular systolic function from cardiac MRI in patients with acquired heart disease

ObjectivesTo evaluate the diagnostic accuracy and variability of 3 semi-quantitative (SQt) methods for assessing right ventricular (RV) systolic function from cardiac MRI in patients with acquired heart disease: tricuspid annular plane systolic excursion (TAPSE), RV fractional-shortening (RVFS) and RV fractional area change (RVFAC).MethodsSixty consecutive patients were enrolled. Reference RV ejection fraction (RVEF) was determined from short axis cine sequences. TAPSE, RVFS and RVFAC were measured on a 4-chamber cine sequence. All SQt analyses were performed twice by 3 observers with various degrees of training in cardiac MRI. Correlation with RVEF, intra- and inter-observer variability, and receiver operating characteristic (ROC) curve analysis were performed for each SQt method.ResultsCorrelation between RVFAC and RVEF was good for all observers and did not depend on previous cardiac MRI experience (R range = 0.716–0.741). Conversely, RVFS (R range = 0.534–0.720) and TAPSE (R range = 0.482–0.646) correlated less with RVEF and depended on previous experience. Intra- and inter-observer variability was much lower for RVFAC than for RVFS and TAPSE. ROC analysis demonstrated that RVFAC <41% could predict a RVEF <45% with 90% sensitivity and 94% specificity.ConclusionsRVFAC appears to be more accurate and reproducible than RVFS and TAPSE for SQt assessment of RV function by cardiac MRI.

In Vitro Assessment of a 3D Segmentation Algorithm Based on the Belief Functions Theory in Calculating Renal Volumes by MRI

OBJECTIVE Renal volumetry is an essential part of split renal function assessment in MR urography. The aim of this study was to assess the accuracy and repeatability of a 3D segmentation algorithm based on the belief functions theory for calculating renal volumes from MR images. MATERIALS AND METHODS The true volumes of 20 animal kidneys of various sizes were obtained by fluid displacement. Each kidney was examined using two different MR units. Three-dimensional proton density-weighted acquisitions with an incremental slice thickness were performed. The MR volume was then measured with a segmentation algorithm based on the belief functions theory. Two independent observers performed all segmentations twice. Accuracy, intraobserver variability, and interobserver variability were evaluated by the Bland-Altman method. The number and type of manual corrections were recorded as well as the entire processing time. RESULTS The mean renal volume estimated by fluid displacement was 114 mL (range, 38-224 mL). With regard to the renal volumes obtained from assessments of adjacent axial MR images, the maximal SDs of the difference were 2.2 mL (accuracy), 0.6 mL (intraobserver variability), and 1.8 mL (interobserver variability). Segmentation of axial slices provided better accuracy and reproducibility than coronal slices. Overlapped coronal slices yielded poor results because of the partial volume effect. The mean processing time including optional manual modifications was less than 75 seconds. CONCLUSION The belief functions theory can be considered an accurate and reproducible mathematic method to assess renal volume from MR adjacent images.

Robust feature selection to predict tumor treatment outcome

OBJECTIVE Recurrence of cancer after treatment increases the risk of death. The ability to predict the treatment outcome can help to design the treatment planning and can thus be beneficial to the patient. We aim to select predictive features from clinical and PET (positron emission tomography) based features, in order to provide doctors with informative factors so as to anticipate the outcome of the patient treatment. METHODS In order to overcome the small sample size problem of datasets usually met in the medical domain, we propose a novel wrapper feature selection algorithm, named HFS (hierarchical forward selection), which searches forward in a hierarchical feature subset space. Feature subsets are iteratively evaluated with the prediction performance using SVM (support vector machine). All feature subsets performing better than those at the preceding iteration are retained. Moreover, as SUV (standardized uptake value) based features have been recognized as significant predictive factors for a patient outcome, we propose to incorporate this prior knowledge into the selection procedure to improve its robustness and reduce its computational cost. RESULTS Two real-world datasets from cancer patients are included in the evaluation. We extract dozens of clinical and PET-based features to characterize the patients state, including SUV parameters and texture features. We use leave-one-out cross-validation to evaluate the prediction performance, in terms of prediction accuracy and robustness. Using SVM as the classifier, our HFS method produces accuracy values of 100% and 94% on the two datasets, respectively, and robustness values of 89% and 96%. Without accuracy loss, the prior-based version (pHFS) improves the robustness up to 100% and 98% on the two datasets, respectively. CONCLUSIONS Compared with other feature selection methods, the proposed HFS and pHFS provide the most promising results. For our HFS method, we have empirically shown that the addition of prior knowledge improves the robustness and accelerates the convergence.