Daniela Stan Raicu

Predicting Radiological Panel Opinions Using a Panel of Machine Learning Classifiers

This paper uses an ensemble of classifiers and active learning strategies to predict radiologists’ assessment of the nodules of the Lung Image Database Consortium (LIDC). In particular, the paper presents machine learning classifiers that model agreement among ratings in seven semantic characteristics: spiculation, lobulation, texture, sphericity, margin, subtlety, and malignancy. The ensemble of classifiers (which can be considered as a computer panel of experts) uses 64 image features of the nodules across four categories (shape, intensity, texture, and size) to predict semantic characteristics. The active learning begins the training phase with nodules on which radiologists’ semantic ratings agree, and incrementally learns how to classify nodules on which the radiologists do not agree. Using our proposed approach, the classification accuracy of the ensemble of classifiers is higher than the accuracy of a single classifier. In the long run, our proposed approach can be used to increase consistency among radiological interpretations by providing physicians a “second read”.

A comparison of texture models for automatic liver segmentation

Automatic liver segmentation from abdominal computed tomography (CT) images based on gray levels or shape alone is difficult because of the overlap in gray-level ranges and the variation in position and shape of the soft tissues. To address these issues, we propose an automatic liver segmentation method that utilizes low-level features based on texture information; this texture information is expected to be homogenous and consistent across multiple slices for the same organ. Our proposed approach consists of the following steps: first, we perform pixel-level texture extraction; second, we generate liver probability images using a binary classification approach; third, we apply a split-and-merge algorithm to detect the seed set with the highest probability area; and fourth, we apply to the seed set a region growing algorithm iteratively to refine the livers boundary and get the final segmentation results. Furthermore, we compare the segmentation results from three different texture extraction methods (Co-occurrence Matrices, Gabor filters, and Markov Random Fields (MRF)) to find the texture method that generates the best liver segmentation. From our experimental results, we found that the co-occurrence model led to the best segmentation, while the Gabor model led to the worst liver segmentation. Moreover, co-occurrence texture features alone produced approximately the same segmentation results as those produced when all the texture features from the combined co-occurrence, Gabor, and MRF models were used. Therefore, in addition to providing an automatic model for liver segmentation, we also conclude that Haralick cooccurrence texture features are the most significant texture characteristics in distinguishing the liver tissue in CT scans.

BRISC—An Open Source Pulmonary Nodule Image Retrieval Framework

We have created a content-based image retrieval framework for computed tomography images of pulmonary nodules. When presented with a nodule image, the system retrieves images of similar nodules from a collection prepared by the Lung Image Database Consortium (LIDC). The system (1) extracts images of individual nodules from the LIDC collection based on LIDC expert annotations, (2) stores the extracted data in a flat XML database, (3) calculates a set of quantitative descriptors for each nodule that provide a high-level characterization of its texture, and (4) uses various measures to determine the similarity of two nodules and perform queries on a selected query nodule. Using our framework, we compared three feature extraction methods: Haralick co-occurrence, Gabor filters, and Markov random fields. Gabor and Markov descriptors perform better at retrieving similar nodules than do Haralick co-occurrence techniques, with best retrieval precisions in excess of 88%. Because the software we have developed and the reference images are both open source and publicly available they may be incorporated into both commercial and academic imaging workstations and extended by others in their research.

Building an ensemble system for diagnosing masses in mammograms.

PurposeClassification of a suspicious mass (region of interest, ROI) in a mammogram as malignant or benign may be achieved using mass shape features. An ensemble system was built for this purpose and tested.MethodsMultiple contours were generated from a single ROI using various parameter settings of the image enhancement functions for the segmentation. For each segmented contour, the mass shape features were computed. For classification, the dataset was partitioned into four subsets based on the patient age (young/old) and the ROI size (large/small). We built an ensemble learning system consisting of four single classifiers, where each classifier is a specialist, trained specifically for one of the subsets. Those specialist classifiers are also an optimal classifier for the subset, selected from several candidate classifiers through preliminary experiment. In this scheme, the final diagnosis (malignant or benign) of an instance is the classification produced by the classifier trained for the subset to which the instance belongs.ResultsThe Digital Database for Screening Mammography (DDSM) from the University of South Florida was used to test the ensemble system for classification of masses, which achieved a 72% overall accuracy. This ensemble of specialist classifiers achieved better performance than single classification (56%).ConclusionAn ensemble classifier for mammography-detected masses may provide superior performance to any single classifier in distinguishing benign from malignant cases.

Semantics and image content integration for pulmonary nodule interpretation in thoracic computed tomography

Useful diagnosis of lung lesions in computed tomography (CT) depends on many factors including the ability of radiologists to detect and correctly interpret the lesions. Computer-aided Diagnosis (CAD) systems can be used to increase the accuracy of radiologists in this task. CAD systems are, however, trained against ground truth and the mechanisms employed by the CAD algorithms may be distinctly different from the visual perception and analysis tasks of the radiologist. In this paper, we present a framework for finding the mappings between human descriptions and characteristics and computed image features. The data in our study were generated from 29 thoracic CT scans collected by the Lung Image Database Consortium (LIDC). Every case was annotated by up to 4 radiologists by marking the contour of nodules and assigning nine semantic terms to each identified nodule; fifty-nine image features were extracted from each segmented nodule. Correlation analysis and stepwise multiple regression were applied to find correlations among semantic characteristics and image features and to generate prediction models for each characteristic based on image features. From our preliminary experimental results, we found high correlations between different semantic terms (margin, texture), and promising mappings from image features to certain semantic terms (texture, lobulation, spiculation, malignancy). While the framework is presented with respect to the interpretation of pulmonary nodules in CT images, it can be easily extended to find mappings for other modalities in other anatomical structures and for other image features.

Mapping LIDC, RadLex™, and lung nodule image features.

AbstractIdeally, an image should be reported and interpreted in the same way (e.g., the same perceived likelihood of malignancy) or similarly by any two radiologists; however, as much research has demonstrated, this is not often the case. Various efforts have made an attempt at tackling the problem of reducing the variability in radiologists’ interpretations of images. The Lung Image Database Consortium (LIDC) has provided a database of lung nodule images and associated radiologist ratings in an effort to provide images to aid in the analysis of computer-aided tools. Likewise, the Radiological Society of North America has developed a radiological lexicon called RadLex. As such, the goal of this paper is to investigate the feasibility of associating LIDC characteristics and terminology with RadLex terminology. If matches between LIDC characteristics and RadLex terms are found, probabilistic models based on image features may be used as decision-based rules to predict if an image or lung nodule could be characterized or classified as an associated RadLex term. The results of this study were matches for 25 (74%) out of 34 LIDC terms in RadLex. This suggests that LIDC characteristics and associated rating terminology may be better conceptualized or reduced to produce even more matches with RadLex. Ultimately, the goal is to identify and establish a more standardized rating system and terminology to reduce the subjective variability between radiologist annotations. A standardized rating system can then be utilized by future researchers to develop automatic annotation models and tools for computer-aided decision systems.

Enhancing undergraduate education: a REU model for interdisciplinary research

This paper presents a successful model for undergraduate research where student participants work on interdisciplinary research projects; in our case, at the frontier between computer science and medicine. Students are part of research teams comprised of other undergraduates, graduate students, faculty and medical experts, participate in professional development and training activities within the larger group, and disseminate their results at the host institutions or conferences specific to the interdisciplinary focus. The model outcomes at the end of the first three years (2005-2007) indicate that the interdisciplinary model successfully 1) expanded the student participation in research by recruiting students who might not otherwise have research opportunities, 2) attracted a diversified pool of talented students into science, 3) promoted interdisciplinary undergraduate studies in computer science and medical informatics as well as in future graduate studies; and 4) trained students in all phases of research, including writing and presenting research papers at conferences.

A classification approach for anatomical regions segmentation

In this paper, a supervised pixel-based classifier approach for segmenting different anatomical regions in abdominal computed tomography (CT) studies is presented. The approach consists of three steps: texture extraction, classifier creation, and anatomical regions identification. First, a set of co-occurrence texture descriptors is calculated for each pixel from the image data sample; second, a decision tree classifier is built using the texture descriptors and the names of the tissues as class labels. At the conclusion of the classification process, a set of decision rules is generated to be used for classification of new pixels and identification of different anatomical regions by joining adjacent pixels with similar classifications. It is expected that the proposed approach will also help automate different semi-automatic segmentation techniques by providing initial boundary points for deformable models or seed points for split and merge segmentation algorithms. Preliminary results obtained for normal CT studies are presented.

Texture-based image retrieval for computerized tomography databases

In this paper we propose a content-based image retrieval (CBIR) system for retrieval of normal anatomical regions present in computed tomography (CT) studies of the chest and abdomen. We implement and compare eight similarity measures using local and global cooccurrence texture descriptors. The preliminary results are obtained using a CT database consisting of 344 CT images representing the segmented heart and great vessels, liver, renal and splenic parenchyma, and backbone from two different patients. We evaluate the results with respect to the retrieval precision metric for each of the similarity measures when calculated per organ and overall.

Directional Invariance of Co-occurrence Matrices within the Liver

Co-occurrence matrices are one of three texture algorithms commonly used on computed tomography (CT) images. In this paper we analyze the directional invariance of Co-occurrence matrices for the purpose of reducing their runtime by reducing the number of directions analyzed without negatively affecting the quality of the texture data extracted.