Feature selection in the pattern classification problem of digital chest radiograph segmentation

Deep Residual Separable Convolutional Neural Network for lung tumor segmentation

Lung cancer is one of the deadliest types of cancers. Computed Tomography (CT) is a widely used technique to detect tumors present inside the lungs. Delineation of such tumors is particularly essential for analysis and treatment purposes. With the advancement in hardware technologies, Machine Learning and Deep Learning methods are outperforming the traditional methods in the field of medical imaging. In order to delineate lung cancer tumors, we have proposed a deep learning-based methodology which includes a maximum intensity projection based pre-processing method, two novel deep learning networks and an ensemble strategy. The two proposed networks named Deep Residual Separable Convolutional Neural Network 1 and 2 (DRS-CNN1 and DRS-CNN2) achieved better performance over the state-of-the-art U-net network and other segmentation networks. For fair comparison, we have evaluated the performances of all networks on Medical Segmentation Decathlon (MSD) and StructSeg 2019 datasets. The DRS-CNN2 achieved a mean Dice Similarity Coefficient (DSC) of 0.649, mean 95 Hausdorff Distance (HD95) of 18.26, mean Sensitivity 0.737 and a mean Precision of 0.765 on independent test sets.

A Generic Approach to Lung Field Segmentation From Chest Radiographs Using Deep Space and Shape Learning

Computer-aided diagnosis (CAD) techniques for lung field segmentation from chest radiographs (CXR) have been proposed for adult cohorts, but rarely for pediatric subjects. Statistical shape models (SSMs), the workhorse of most state-of-the-art CXR-based lung field segmentation methods, do not efficiently accommodate shape variation of the lung field during the pediatric developmental stages. The main contributions of our work are: 1) a generic lung field segmentation framework from CXR accommodating large shape variation for adult and pediatric cohorts; 2) a deep representation learning detection mechanism, ensemble space learning, for robust object localization; and 3) marginal shape deep learning for the shape deformation parameter estimation. Unlike the iterative approach of conventional SSMs, the proposed shape learning mechanism transforms the parameter space into marginal subspaces that are solvable efficiently using the recursive representation learning mechanism. Furthermore, our method is the first to include the challenging retro-cardiac region in the CXR-based lung segmentation for accurate lung capacity estimation. The framework is evaluated on 668 CXRs of patients between 3 month to 89 year of age. We obtain a mean Dice similarity coefficient of 0.96 ±0.03 (including the retro-cardiac region). For a given accuracy, the proposed approach is also found to be faster than conventional SSM-based iterative segmentation methods. The computational simplicity of the proposed generic framework could be similarly applied to the fast segmentation of other deformable objects.

An Improved Normalized Mutual Information Variable Selection Algorithm for Neural Network-Based Soft Sensors

In this paper, normalized mutual information feature selection (NMIFS) and tabu search (TS) are integrated to develop a new variable selection algorithm for soft sensors. NMIFS is applied to select influential variables contributing to the output variable and avoids selecting redundant variables by calculating mutual information (MI). A TS based strategy is designed to prevent NMIFS from falling into a local optimal solution. The proposed algorithm performs the variable selection by combining the entropy information and MI and validating error information of artificial neural networks (ANNs); therefore, it has advantages over previous MI-based variable selection algorithms. Several simulation datasets with different scales, correlations and noise parameters are implemented to demonstrate the performance of the proposed algorithm. A set of actual production data from a power plant is also used to check the performance of these algorithms. The experiments showed that the developed variable selection algorithm presents better model accuracy with fewer selected variables, compared with other state-of-the-art methods. The application of this algorithm to soft sensors can achieve reliable results.

Prediction of progression in idiopathic pulmonary fibrosis using CT scans at baseline: A quantum particle swarm optimization - Random forest approach

Idiopathic pulmonary fibrosis (IPF) is a fatal lung disease characterized by an unpredictable progressive decline in lung function. Natural history of IPF is unknown and the prediction of disease progression at the time of diagnosis is notoriously difficult. High resolution computed tomography (HRCT) has been used for the diagnosis of IPF, but not generally for monitoring purpose. The objective of this work is to develop a novel predictive model for the radiological progression pattern at voxel-wise level using only baseline HRCT scans. Mainly, there are two challenges: (a) obtaining a data set of features for region of interest (ROI) on baseline HRCT scans and their follow-up status; and (b) simultaneously selecting important features from high-dimensional space, and optimizing the prediction performance. We resolved the first challenge by implementing a study design and having an expert radiologist contour ROIs at baseline scans, depending on its progression status in follow-up visits. For the second challenge, we integrated the feature selection with prediction by developing an algorithm using a wrapper method that combines quantum particle swarm optimization to select a small number of features with random forest to classify early patterns of progression. We applied our proposed algorithm to analyze anonymized HRCT images from 50 IPF subjects from a multi-center clinical trial. We showed that it yields a parsimonious model with 81.8% sensitivity, 82.2% specificity and an overall accuracy rate of 82.1% at the ROI level. These results are superior to other popular feature selections and classification methods, in that our method produces higher accuracy in prediction of progression and more balanced sensitivity and specificity with a smaller number of selected features. Our work is the first approach to show that it is possible to use only baseline HRCT scans to predict progressive ROIs at 6 months to 1year follow-ups using artificial intelligence.

Review: On Segmentation of Nodules from Posterior and Anterior Chest Radiographs

Lung cancer is one of the major types of cancer in the world. Survival rate can be increased if the disease can be identified early. Posterior and anterior chest radiography and computerized tomography scans are the most used diagnosis techniques for detecting tumor from lungs. Posterior and anterior chest radiography requires less radiation dose and is available in most of the diagnostic centers and it costs less compared to the remaining diagnosis techniques. So PA chest radiography became the most commonly used technique for lung cancer detection. Because of superimposed anatomical structures present in the image, sometimes radiologists cannot find abnormalities from the image. To help radiologists in diagnosing tumor from PA chest radiographic images range of CAD scheme has been developed for the past three decades. These computerized tools may be used by radiologists as a second opinion in detecting tumor. Literature survey on detecting tumors from chest graphs is presented in this paper.

Lung field segmentation using weighted sparse shape composition with robust initialization

PURPOSE

Lung field segmentation for chest radiography is critical to pulmonary disease diagnosis. In this paper, we propose a new deformable model using weighted sparse shape composition with robust initialization to achieve robust and accurate lung field segmentation.

METHODS

Our method consists of three steps: initialization, deformation and regularization. The steps of deformation and regularization are iteratively employed until convergence. First, since a deformable model is sensitive to the initial shape, a robust initialization is obtained by using a novel voting strategy, which allows the reliable patches on the image to vote for each landmark of the initial shape. Then, each point of the initial shape independently deforms to the lung boundary under the guidance of the appearance model, which can distinguish lung tissues from nonlung tissues near the boundary. Finally, the deformed shape is regularized by weighted sparse shape composition (SSC) model, which is constrained by both boundary information and the correlations between each point of the deformed shape.

RESULTS

Our method has been evaluated on 247 chest radiographs from well-known dataset Japanese Society of Radiological Technology (JSRT) and achieved high overlap scores (0.955 ± 0.021).

CONCLUSIONS

The experimental results show that the proposed deformable segmentation model is more robust and accurate than the traditional appearance and shape model on the JSRT database. Our method also shows higher accuracy than most state-of-the-art methods.

Preliminary research on the identification system for anthracnose and powdery mildew of sandalwood leaf based on image processing

This paper presents a survey on a system that uses digital image processing techniques to identify anthracnose and powdery mildew diseases of sandalwood from digital images. Our main objective is researching the most suitable identification technology for the anthracnose and powdery mildew diseases of the sandalwood leaf, which provides algorithmic support for the real-time machine judgment of the health status and disease level of sandalwood. We conducted real-time monitoring of Hainan sandalwood leaves with varying severity levels of anthracnose and powdery mildew beginning in March 2014. We used image segmentation, feature extraction and digital image classification and recognition technology to carry out a comparative experimental study for the image analysis of powdery mildew, anthracnose disease and healthy leaves in the field. Performing the actual test for a large number of diseased leaves pointed to three conclusions: (1) Distinguishing effects of BP (Back Propagation) neural network method, in all kinds of classical methods, for sandalwood leaf anthracnose and powdery mildew disease are relatively good; the size of the lesion areas were closest to the actual. (2) The differences between two diseases can be shown well by the shape feature, color feature and texture feature of the disease image. (3) Identifying and diagnosing the diseased leaves have ideal results by SVM, which is based on radial basis kernel function. The identification rate of the anthracnose and healthy leaves was 92% respectively, and that of powdery mildew was 84%. Disease identification technology lays the foundation for remote monitoring disease diagnosis, preparing for remote transmission of the disease images, which is a very good guide and reference for further research of the disease identification and diagnosis system in sandalwood and other species of trees.

Organ Location Determination and Contour Sparse Representation for Multiorgan Segmentation

Organ segmentation on computed tomography (CT) images is of great importance in medical diagnoses and treatment. This paper proposes organ location determination and contour sparse representation methods (OLD-CSR) for multiorgan segmentation (liver, kidney, and spleen) on abdomen CT images using an extreme learning machine classifier. First, a location determination method is designed to obtain location information of each organ, which is used for coarse segmentation. Second, for coarse-to-fine segmentation, a contour gradient and rate change based feature point extraction method is proposed. A sparse optimization model is developed for refining the contour feature points. Experimentations with 153 CT images demonstrate the performance advantages of OLD-CSR as compared with related work.

Hierarchical lung field segmentation with joint shape and appearance sparse learning

Lung field segmentation in the posterior-anterior (PA) chest radiograph is important for pulmonary disease diagnosis and hemodialysis treatment. Due to high shape variation and boundary ambiguity, accurate lung field segmentation from chest radiograph is still a challenging task. To tackle these challenges, we propose a joint shape and appearance sparse learning method for robust and accurate lung field segmentation. The main contributions of this paper are: 1) a robust shape initialization method is designed to achieve an initial shape that is close to the lung boundary under segmentation; 2) a set of local sparse shape composition models are built based on local lung shape segments to overcome the high shape variations; 3) a set of local appearance models are similarly adopted by using sparse representation to capture the appearance characteristics in local lung boundary segments, thus effectively dealing with the lung boundary ambiguity; 4) a hierarchical deformable segmentation framework is proposed to integrate the scale-dependent shape and appearance information together for robust and accurate segmentation. Our method is evaluated on 247 PA chest radiographs in a public dataset. The experimental results show that the proposed local shape and appearance models outperform the conventional shape and appearance models. Compared with most of the state-of-the-art lung field segmentation methods under comparison, our method also shows a higher accuracy, which is comparable to the inter-observer annotation variation.

A recent survey on colon cancer detection techniques

Colon cancer causes deaths of about half a million people every year. Common method of its detection is histopathological tissue analysis, which, though leads to vital diagnosis, is significantly correlated to the tiredness, experience, and workload of the pathologist. Researchers have been working since decades to get rid of manual inspection, and to develop trustworthy systems for detecting colon cancer. Several techniques, based on spectral/spatial analysis of colon biopsy images, and serum and gene analysis of colon samples, have been proposed in this regard. Due to rapid evolution of colon cancer detection techniques, a latest review of recent research in this field is highly desirable. The aim of this paper is to discuss various colon cancer detection techniques. In this survey, we categorize the techniques on the basis of the adopted methodology and underlying data set, and provide detailed description of techniques in each category. Additionally, this study provides an extensive comparison of various colon cancer detection categories, and of multiple techniques within each category. Further, most of the techniques have been evaluated on similar data set to provide a fair performance comparison. Analysis reveals that neither of the techniques is perfect; however, research community is progressively inching toward the finest possible solution.

Automated lung segmentation in digital chest tomosynthesis

PURPOSE

The purpose of this study was to develop an automated lung segmentation method for computerized detection of lung nodules in digital chest tomosynthesis.

METHODS

The authors collected 45 digital tomosynthesis scans and manually segmented reference lung regions in each scan to assess the performance of the method. The authors automated the technique by calculating the edge gradient in an original image for enhancing lung outline and transforming the edge gradient image to polar coordinate space. The authors then employed a dynamic programming technique to delineate outlines of the unobscured lungs in the transformed edge gradient image. The lung outlines were converted back to the original image to provide the final segmentation result. The above lung segmentation algorithm was first applied to the central reconstructed tomosynthesis slice because of the absence of ribs overlapping lung structures. The segmented lung in the central slice was then used to guide lung segmentation in noncentral slices. The authors evaluated the segmentation method by using (1) an overlap rate of lung regions, (2) a mean absolute distance (MAD) of lung borders, (3) a Hausdorff distance of lung borders between the automatically segmented lungs and manually segmented reference lungs, and (4) the fraction of nodules included in the automatically segmented lungs.

RESULTS

The segmentation method achieved mean overlap rates of 85.7%, 88.3%, and 87.0% for left lungs, right lungs, and entire lungs, respectively; mean MAD of 4.8, 3.9, and 4.4 mm for left lungs, right lungs, and entire lungs, respectively; and mean Hausdorrf distance of 25.0 mm, 25.5 mm, and 30.1 mm for left lungs, right lungs, and entire lungs, respectively. All of the nodules inside the reference lungs were correctly included in the segmented lungs obtained with the lung segmentation method.

CONCLUSIONS

The method achieved relatively high accuracy for lung segmentation and will be useful for computer-aided detection of lung nodules in digital tomosynthesis.

Association between ICP pulse waveform morphology and ICP B waves

The study aimed to investigate changes in the shape of ICP pulses associated with different patterns of the ICP slow waves (0.5-2.0 cycles/min) during ICP overnight monitoring in hydrocephalus. Four patterns of ICP slow waves were characterized in 44 overnight ICP recordings (no waves - NW, slow symmetrical waves - SW, slow asymmetrical waves - AS, slow waves with plateau phase - PW). The morphological clustering and analysis of ICP pulse (MOCAIP) algorithm was utilized to calculate a set of metrics describing ICP pulse morphology based on the location of three sub-peaks in an ICP pulse: systolic peak (P(1)), tidal peak (P(2)) and dicrotic peak (P(3)). Step-wise discriminant analysis was applied to select the most characteristic morphological features to distinguish between different ICP slow waves. Based on relative changes in variability of amplitudes of P(2) and P(3) we were able to distinguish between the combined groups NW + SW and AS + PW (p < 0.000001). The AS pattern can be differentiated from PW based on respective changes in the mean curvature of P(2) and P(3) (p < 0.000001); however, none of the MOCAIP feature separates between NW and SW. The investigation of ICP pulse morphology associated with different ICP B waves may provide additional information for analysing recordings of overnight ICP.

Automatic evaluation of cardiac hypertrophy using cardiothoracic area ratio in chest radiograph images

To evaluate the cardiac hypertrophy from chest radiograph images, radiologists usually examine the cardiothoracic ratio (frequently called CTR) which is a standard diagnostic index. The CTR is computed by the maximum transverse diameter of the heart shadow divided by the maximum transverse diameter of right and left lung boundaries. In this paper, we present a method to evaluate the cardiac hypertrophy by comparing the area of heart with that of lung, instead of the cardiothoracic ratio to get more desirable diagnostic results. We introduce a new index, a cardiothoracic area ratio (CTAR), which is computed by dividing the area of heart region by the area of lung region of specific interest. We first segment a chest region of interest in a radiograph image and then automatically compute the traditional CTR and the CTAR to evaluate the cardiac hypertrophy. And finally, we provide the visual presentation of those ratios on the chest radiograph image. The experimental results using a set of radiograph images show that the proposed method can be used effectively for determining the cardiac hypertrophy in a real-time diagnostic environment. It provides the higher discrimination power than the CTR to identify hypertrophied hearts by recognizing the heart enlargement. It also can be used together with the traditional CTR as a complementary measure when it is difficult to determine abnormalities by the CTR, reducing the rate of wrong diagnosis.

Real-time texture analysis for identifying optimum microbubble concentration in 2-D ultrasonic particle image velocimetry

Many recent studies on ultrasonic particle image velocimetry (Echo PIV) showed that the accuracy of two-dimensional (2-D) flow velocity measured depends largely on the concentration of ultrasound contrast agents (UCAs) during imaging. This article presents a texture-based method for identifying the optimum microbubble concentration for Echo PIV measurements in real-time. The texture features, standard deviation of gray level, and contrast, energy and homogeneity of gray level co-occurrence matrix were extracted from ultrasound contrast images of rotational and pulsatile flow (10 MHz) in vitro and in vivo mouse common carotid arterial flow (40 MHz) with UCAs at various concentrations. The results showed that, at concentration of 0.8∼2 × 10³ bubbles/mL in vitro and 1∼5 × 10⁵ bubbles/mL in vivo, image texture features had a peak value or trough value, and velocity vectors with high accuracy can be obtained. Otherwise, poor quality velocity vectors were obtained. When the texture features were used as a feature set, the accuracy of K-nearest neighbor classifier can reach 86.4% in vitro and 87.5% in vivo, respectively. The texture-based method is shown to be able to quickly identify the optimum microbubble concentration and improve the accuracy for Echo PIV imaging.

An automated neural-fuzzy approach to malignant tumor localization in 2D ultrasonic images of the prostate

In this paper, a new neural-fuzzy approach is proposed for automated region segmentation in transrectal ultrasound images of the prostate. The goal of region segmentation is to identify suspicious regions in the prostate in order to provide decision support for the diagnosis of prostate cancer. The new automated region segmentation system uses expert knowledge as well as both textural and spatial features in the image to accomplish the segmentation. The textural information is extracted by two recurrent random pulsed neural networks trained by two sets of data (a suspicious tissues' data set and a normal tissues' data set). Spatial information is captured by the atlas-based reference approach and is represented as fuzzy membership functions. The textural and spatial features are synthesized by a fuzzy inference system, which provides a binary classification of the region to be evaluated.

Multiresolution texture models for brain tumor segmentation in MRI

In this study we discuss different types of texture features such as Fractal Dimension (FD) and Multifractional Brownian Motion (mBm) for estimating random structures and varying appearance of brain tissues and tumors in magnetic resonance images (MRI). We use different selection techniques including KullBack - Leibler Divergence (KLD) for ranking different texture and intensity features. We then exploit graph cut, self organizing maps (SOM) and expectation maximization (EM) techniques to fuse selected features for brain tumors segmentation in multimodality T1, T2, and FLAIR MRI. We use different similarity metrics to evaluate quality and robustness of these selected features for tumor segmentation in MRI for real pediatric patients. We also demonstrate a non-patient-specific automated tumor prediction scheme by using improved AdaBoost classification based on these image features.

Pattern recognition of overnight intracranial pressure slow waves using morphological features of intracranial pressure pulse

This study aimed to develop a new approach to detect intracranial pressure (ICP) slow waves based on morphological changes of ICP pulse waveforms. A recently proposed Morphological Clustering and Analysis of ICP Pulse (MOCAIP) algorithm was utilized to calculate a set of metrics that characterize ICP pulse morphology. A regularized linear quadratic classifier was used to test the hypothesis that classification between ICP slow wave and flat ICP could be achieved using features composed of mean values and dispersion of 24 MOCAIP metrics. To optimize the classification performance, three feature selection techniques (differential evolution, discriminant analysis and analysis of variance) were applied to find an optimal set of MOCAIP metrics under different criteria. In addition, we selected three sets of metrics common to those found by combination of two selection methods, to be used as classification features (differential evolution and analysis of variance, discriminant analysis and analysis of variance, and combination of differential evolution and discriminant analysis). To test the approach, a total of 276 selections of ICP recordings corresponding to two patterns without waves and containing slow waves were obtained from overnight ICP studies of 44 hydrocephalus patients performed at the UCLA Adult Hydrocephalus Center. Our results showed that the best classification performance of differentiation of slow waves from the ICP recording without slow waves was obtained using the combination of metrics common to both differential evolution and analysis of variance methods; achieving an accuracy of 89%, specificity 96%, and sensitivity 83%.

Automatic detection of scoliotic curves in posteroanterior radiographs

Spinal deformities are diagnosed using posteroanterior (PA) radiographs. Automatic detection of the spine on conventional radiographs would be of interest to quantify curve severity, would help reduce observer variability and would allow large-scale retrospective studies on radiographic databases. The goal of this paper is to present a new method for automatic detection of spinal curves from a PA radiograph. A region of interest (ROI) is first extracted according to the 2-D shape variability of the spine obtained from a set of PA radiographs of scoliotic patients. This region includes 17 bounding boxes delimiting each vertebral level from T1 to L5. An adaptive filter combining shock with complex diffusion is used to individually restore the image of each vertebral level. Then, texture descriptors of small block elements are computed and submitted for training to support vector machines (SVM). Vertebral body's locations are thereby inferred for a particular vertebral level. The classifications of block elements for all 17 SVMs are identified in the image and a voting system is introduced to cumulate correctly predicted blocks. A spline curve is then fitted through the centers of the predicted vertebral regions and compared to a manual identification using a Student t-test. A clinical validation is performed using 100 radiographs of scoliotic patients (not used for training) and the detected spinal curve is found to be statistically similar (p < 0.05) in 93% of cases to the manually identified curve.

A hierarchical method based on active shape models and directed Hough transform for segmentation of noisy biomedical images; application in segmentation of pelvic X-ray images

Background

Traumatic pelvic injuries are often associated with severe, life-threatening hemorrhage, and immediate medical treatment is therefore vital. However, patient prognosis depends heavily on the type, location and severity of the bone fracture, and the complexity of the pelvic structure presents diagnostic challenges. Automated fracture detection from initial patient X-ray images can assist physicians in rapid diagnosis and treatment, and a first and crucial step of such a method is to segment key bone structures within the pelvis; these structures can then be analyzed for specific fracture characteristics. Active Shape Model has been applied for this task in other bone structures but requires manual initialization by the user. This paper describes a algorithm for automatic initialization and segmentation of key pelvic structures - the iliac crests, pelvic ring, left and right pubis and femurs - using a hierarchical approach that combines directed Hough transform and Active Shape Models.

Results

Performance of the automated algorithm is compared with results obtained via manual initialization. An error measures is calculated based on the shapes detected with each method and the gold standard shapes. ANOVA results on these error measures show that the automated algorithm performs at least as well as the manual method. Visual inspection by two radiologists and one trauma surgeon also indicates generally accurate performance.

Conclusion

The hierarchical algorithm described in this paper automatically detects and segments key structures from pelvic X-rays. Unlike various other x-ray segmentation methods, it does not require manual initialization or input. Moreover, it handles the inconsistencies between x-ray images in a clinical environment and performs successfully in the presence of fracture. This method and the segmentation results provide a valuable base for future work in fracture detection.

Efficient segmentation by sparse pixel classification

Segmentation methods based on pixel classification are powerful but often slow. We introduce two general algorithms, based on sparse classification, for optimizing the computation while still obtaining accurate segmentations. The computational costs of the algorithms are derived, and they are demonstrated on real 3-D magnetic resonance imaging and 2-D radiograph data. We show that each algorithm is optimal for specific tasks, and that both algorithms allow a speedup of one or more orders of magnitude on typical segmentation tasks.

Helicobacter pylori-related gastric histology classification using support-vector-machine-based feature selection

This study presents a computer-aided diagnosis system using sequential forward floating selection (SFFS) with support vector machine (SVM) to diagnose gastric histology of Helicobacter pylori (H. pylori) from endoscopic images. To achieve this goal, candidate image features associated with clinical symptoms are extracted from endoscopic images. With these candidate features, the SFFS method is applied to select feature subsets, which perform the best classification results under SVM with respect to different histological features. By using the classifiers obtained from the feature subsets, a new diagnosis system is implemented to provide physicians with H. pylori -related histological results from endoscopic images.

Segmenting lung fields in serial chest radiographs using both population-based and patient-specific shape statistics

This paper presents a new deformable model using both population-based and patient-specific shape statistics to segment lung fields from serial chest radiographs. There are two novelties in the proposed deformable model. First, a modified scale invariant feature transform (SIFT) local descriptor, which is more distinctive than the general intensity and gradient features, is used to characterize the image features in the vicinity of each pixel. Second, the deformable contour is constrained by both population-based and patient-specific shape statistics, and it yields more robust and accurate segmentation of lung fields for serial chest radiographs. In particular, for segmenting the initial time-point images, the population-based shape statistics is used to constrain the deformable contour; as more subsequent images of the same patient are acquired, the patient-specific shape statistics online collected from the previous segmentation results gradually takes more roles. Thus, this patient-specific shape statistics is updated each time when a new segmentation result is obtained, and it is further used to refine the segmentation results of all the available time-point images. Experimental results show that the proposed method is more robust and accurate than other active shape models in segmenting the lung fields from serial chest radiographs.

Minimal shape and intensity cost path segmentation

A new generic model-based segmentation algorithm is presented, which can be trained from examples akin to the active shape model (ASM) approach in order to acquire knowledge about the shape to be segmented and about the gray-level appearance of the object in the image. Whereas ASM alternates between shape and intensity information during search, the proposed approach optimizes for shape and intensity characteristics simultaneously. Local gray-level appearance information at the landmark points extracted from feature images is used to automatically detect a number of plausible candidate locations for each landmark. The shape information is described by multiple landmark-specific statistical models that capture local dependencies between adjacent landmarks on the shape. The shape and intensity models are combined in a single cost function that is optimized noniteratively using dynamic programming, without the need for initialization. The algorithm was validated for segmentation of anatomical structures in chest and hand radiographs. In each experiment, the presented method had a significant higher performance when compared to the ASM schemes. As the method is highly effective, optimally suited for pathological cases and easy to implement, it is highly useful for many medical image segmentation tasks.

A fully automated method for lung nodule detection from postero-anterior chest radiographs

In the past decades, a great deal of research work has been devoted to the development of systems that could improve radiologists' accuracy in detecting lung nodules. Despite the great efforts, the problem is still open. In this paper, we present a fully automated system processing digital postero-anterior (PA) chest radiographs, that starts by producing an accurate segmentation of the lung field area. The segmented lung area includes even those parts of the lungs hidden behind the heart, the spine, and the diaphragm, which are usually excluded from the methods presented in the literature. This decision is motivated by the fact that lung nodules may be found also in these areas. The segmented area is processed with a simple multiscale method that enhances the visibility of the nodules, and an extraction scheme is then applied to select potential nodules. To reduce the high number of false positives extracted, cost-sensitive support vector machines (SVMs) are trained to recognize the true nodules. Different learning experiments were performed on two different data sets, created by means of feature selection, and employing Gaussian and polynomial SVMs trained with different parameters; the results are reported and compared. With the best SVM models, we obtain about 1.5 false positives per image (fp/image) when sensitivity is approximately equal to 0.71; this number increases to about 2.5 and 4 fp/image when sensitivity is = 0.78 and = 0.85, respectively. For the highest sensitivity (= 0.92 and 1.0), we get 7 or 8 fp/image.

Segmentation of anatomical structures in chest radiographs using supervised methods: a comparative study on a public database

The task of segmenting the lung fields, the heart, and the clavicles in standard posterior-anterior chest radiographs is considered. Three supervised segmentation methods are compared: active shape models, active appearance models and a multi-resolution pixel classification method that employs a multi-scale filter bank of Gaussian derivatives and a k-nearest-neighbors classifier. The methods have been tested on a publicly available database of 247 chest radiographs, in which all objects have been manually segmented by two human observers. A parameter optimization for active shape models is presented, and it is shown that this optimization improves performance significantly. It is demonstrated that the standard active appearance model scheme performs poorly, but large improvements can be obtained by including areas outside the objects into the model. For lung field segmentation, all methods perform well, with pixel classification giving the best results: a paired t-test showed no significant performance difference between pixel classification and an independent human observer. For heart segmentation, all methods perform comparably, but significantly worse than a human observer. Clavicle segmentation is a hard problem for all methods; best results are obtained with active shape models, but human performance is substantially better. In addition, several hybrid systems are investigated. For heart segmentation, where the separate systems perform comparably, significantly better performance can be obtained by combining the results with majority voting. As an application, the cardio-thoracic ratio is computed automatically from the segmentation results. Bland and Altman plots indicate that all methods perform well when compared to the gold standard, with confidence intervals from pixel classification and active appearance modeling very close to those of a human observer. All results, including the manual segmentations, have been made publicly available to facilitate future comparative studies.

GMDH-based feature ranking and selection for improved classification of medical data

Medical applications are often characterized by a large number of disease markers and a relatively small number of data records. We demonstrate that complete feature ranking followed by selection can lead to appreciable reductions in data dimensionality, with significant improvements in the implementation and performance of classifiers for medical diagnosis. We describe a novel approach for ranking all features according to their predictive quality using properties unique to learning algorithms based on the group method of data handling (GMDH). An abductive network training algorithm is repeatedly used to select groups of optimum predictors from the feature set at gradually increasing levels of model complexity specified by the user. Groups selected earlier are better predictors. The process is then repeated to rank features within individual groups. The resulting full feature ranking can be used to determine the optimum feature subset by starting at the top of the list and progressively including more features until the classification error rate on an out-of-sample evaluation set starts to increase due to overfitting. The approach is demonstrated on two medical diagnosis datasets (breast cancer and heart disease) and comparisons are made with other feature ranking and selection methods. Receiver operating characteristics (ROC) analysis is used to compare classifier performance. At default model complexity, dimensionality reduction of 22 and 54% could be achieved for the breast cancer and heart disease data, respectively, leading to improvements in the overall classification performance. For both datasets, considerable dimensionality reduction introduced no significant reduction in the area under the ROC curve. GMDH-based feature selection results have also proved effective with neural network classifiers.

Improved classification of medical data using abductive network committees trained on different feature subsets

This paper demonstrates the use of abductive network classifier committees trained on different features for improving classification accuracy in medical diagnosis. In an earlier publication, committee members were trained on different subsets of the training set to ensure enough diversity for improved committee performance. In situations characterized by high data dimensionality, i.e. a large number of features and a relatively few training examples, it may be more advantageous to split the feature set rather than the training set. We describe a novel approach for tentatively ranking the features and forming subsets of uniform predictive quality for training individual members. The abductive network training algorithm is used to select optimum predictors from the feature set at various levels of model complexity specified by the user. Using the resulting tentative ranking, the features are grouped into mutually exclusive subsets of approximately equal predictive power for training the members. The approach is demonstrated on three standard medical diagnosis datasets (breast cancer, heart disease, and diabetes). Three-member committees trained on different feature subsets and using simple output combination methods reduce classification errors by up to 20% compared to the best single model developed with the full feature set. Results are compared with those reported previously with members trained through splitting the training set. Training abductive committee members on feature subsets of approximately equal predictive power achieves both diversity and quality for improved committee performance. Ensemble feature subset selection can be performed using GMDH-based learning algorithms. The approach should be advantageous in situations characterized by high data dimensionality.

Segmentation of magnetic resonance images using a combination of neural networks and active contour models

Segmentation of medical images is very important for clinical research and diagnosis, leading to a requirement for robust automatic methods. This paper reports on the combined use of a neural network (a multilayer perceptron, MLP) and active contour model ('snake') to segment structures in magnetic resonance (MR) images. The perceptron is trained to produce a binary classification of each pixel as either a boundary or a non-boundary point. Subsequently, the resulting binary (edge-point) image forms the external energy function for a snake, used to link the candidate boundary points into a continuous, closed contour. We report here on the segmentation of the lungs from multiple MR slices of the torso; lung-specific constraints have been avoided to keep the technique as general as possible. In initial investigations, the inputs to the MLP were limited to normalised intensity values of the pixels from an (7 x 7) window scanned across the image. The use of spatial coordinates as additional inputs to the MLP is then shown to provide an improvement in segmentation performance as quantified using the effectiveness measure (a weighted product of precision and recall). Training sets were first developed using a lengthy iterative process. Thereafter, a novel cost function based on effectiveness is proposed for training that allows us to achieve dramatic improvements in segmentation performance, as well as faster, non-iterative selection of training examples. The classifications produced using this cost function were sufficiently good that the binary image produced by the MLP could be post-processed using an active contour model to provide an accurate segmentation of the lungs from the multiple slices in almost all cases, including unseen slices and subjects.

Automatic detection of abnormalities in chest radiographs using local texture analysis

A fully automatic method is presented to detect abnormalities in frontal chest radiographs which are aggregated into an overall abnormality score. The method is aimed at finding abnormal signs of a diffuse textural nature, such as they are encountered in mass chest screening against tuberculosis (TB). The scheme starts with automatic segmentation of the lung fields, using active shape models. The segmentation is used to subdivide the lung fields into overlapping regions of various sizes. Texture features are extracted from each region, using the moments of responses to a multiscale filter bank. Additional "difference features" are obtained by subtracting feature vectors from corresponding regions in the left and right lung fields. A separate training set is constructed for each region. All regions are classified by voting among the k nearest neighbors, with leave-one-out. Next, the classification results of each region are combined, using a weighted multiplier in which regions with higher classification reliability weigh more heavily. This produces an abnormality score for each image. The method is evaluated on two databases. The first database was collected from a TB mass chest screening program, from which 147 images with textural abnormalities and 241 normal images were selected. Although this database contains many subtle abnormalities, the classification has a sensitivity of 0.86 at a specificity of 0.50 and an area under the receiver operating characteristic (ROC) curve of 0.820. The second database consist of 100 normal images and 100 abnormal images with interstitial disease. For this database, the results were a sensitivity of 0.97 at a specificity of 0.90 and an area under the ROC curve of 0.986.

Application of the mutual information criterion for feature selection in computer-aided diagnosis

The purpose of this study was to investigate an information theoretic approach to feature selection for computer-aided diagnosis (CAD). The approach is based on the mutual information (MI) concept. MI measures the general dependence of random variables without making any assumptions about the nature of their underlying relationships. Consequently, MI can potentially offer some advantages over feature selection techniques that focus only on the linear relationships of variables. This study was based on a database of statistical texture features extracted from perfusion lung scans. The ultimate goal was to select the optimal subset of features for the computer-aided diagnosis of acute pulmonary embolism (PE). Initially, the study addressed issues regarding the approximation of MI in a limited dataset as it is often the case in CAD applications. The MI selected features were compared to those features selected using stepwise linear discriminant analysis and genetic algorithms for the same PE database. Linear and nonlinear decision models were implemented to merge the selected features into a final diagnosis. Results showed that the MI is an effective feature selection criterion for nonlinear CAD models overcoming some of the well-known limitations and computational complexities of other popular feature selection techniques in the field.

Computer-aided diagnosis in chest radiography: a survey

The traditional chest radiograph is still ubiquitous in clinical practice, and will likely remain so for quite some time. Yet, its interpretation is notoriously difficult. This explains the continued interest in computer-aided diagnosis for chest radiography. The purpose of this survey is to categorize and briefly review the literature on computer analysis of chest images, which comprises over 150 papers published in the last 30 years. Remaining challenges are indicated and some directions for future research are given.

Improvement of mammographic mass characterization using spiculation meausures and morphological features

We are developing new computer vision techniques for characterization of breast masses on mammograms. We had previously developed a characterization method based on texture features. The goal of the present work was to improve our characterization method by making use of morphological features. Toward this goal, we have developed a fully automated, three-stage segmentation method that includes clustering, active contour, and spiculation detection stages. After segmentation, morphological features describing the shape of the mass were extracted. Texture features were also extracted from a band of pixels surrounding the mass. Stepwise feature selection and linear discriminant analysis were employed in the morphological, texture, and combined feature spaces for classifier design. The classification accuracy was evaluated using the area Az under the receiver operating characteristic curve. A data set containing 249 films from 102 patients was used. When the leave-one-case-out method was applied to partition the data set into trainers and testers, the average test Az for the task of classifying the mass on a single mammographic view was 0.83 +/- 0.02, 0.84 +/- 0.02, and 0.87 +/- 0.02 in the morphological, texture, and combined feature spaces, respectively. The improvement obtained by supplementing texture features with morphological features in classification was statistically significant (p = 0.04). For classifying a mass as malignant or benign, we combined the leave-one-case-out discriminant scores from different views of a mass to obtain a summary score. In this task, the test Az value using the combined feature space was 0.91 +/- 0.02. Our results indicate that combining texture features with morphological features extracted from automatically segmented mass boundaries will be an effective approach for computer-aided characterization of mammographic masses.

Improved method for automatic identification of lung regions on chest radiographs

RATIONALE AND OBJECTIVES

The authors performed this study to evaluate an algorithm developed to help identify lungs on chest radiographs.

MATERIALS AND METHODS

Forty clinical posteroanterior chest radiographs obtained in adult patients were digitized to 12-bit gray-scale resolution. In the proposed algorithm, the authors simplified the current approach of edge detection with derivatives by using only the first derivative of the horizontal and/or vertical image profiles. In addition to the derivative method, pattern classification and image feature analysis were used to determine the region of interest and lung boundaries. Instead of using the traditional curve-fitting method to delineate the lung, the authors applied an iterative contour-smoothing algorithm to each of the four detected boundary segments (costal, mediastinal, lung apex, and hemidiaphragm edges) to form a smooth lung boundary.

RESULTS

The algorithm had an average accuracy of 96.0% for the right lung and 95.2% for the left lung and was especially useful in the delineation of hemidiaphragm edges. In addition, it took about 0.775 second per image to identify the lung boundaries, which is much faster than that of other algorithms noted in the literature.

CONCLUSION

The computer-generated segmentation results can be used directly in the detection and compensation of rib structures and in lung nodule detection.

Automatic segmentation of lung fields in chest radiographs

The delineation of important structures in chest radiographs is an essential preprocessing step in order to automatically analyze these images, e.g., for tuberculosis screening support or in computer assisted diagnosis. We present algorithms for the automatic segmentation of lung fields in chest radiographs. We compare several segmentation techniques: a matching approach; pixel classifiers based on several combinations of features; a new rule-based scheme that detects lung contours using a general framework for the detection of oriented edges and ridges in images; and a hybrid scheme. Each approach is discussed and the performance of nine systems is compared with interobserver variability and results available from the literature. The best performance is obtained by the hybrid scheme that combines the rule-based segmentation algorithm with a pixel classification approach. The combinations of two complementary techniques leads to robust performance; the accuracy is above 94% for all 115 images in the test set. The average accuracy of the scheme is 0.969 +/- 0.0080, which is close to the interobserver variability of 0.984 +/- 0.0048. The methods are fast, and implemented on a standard PC platform.

Feature selection and classifier performance in computer-aided diagnosis: the effect of finite sample size

In computer-aided diagnosis (CAD), a frequently used approach for distinguishing normal and abnormal cases is first to extract potentially useful features for the classification task. Effective features are then selected from this entire pool of available features. Finally, a classifier is designed using the selected features. In this study, we investigated the effect of finite sample size on classification accuracy when classifier design involves stepwise feature selection in linear discriminant analysis, which is the most commonly used feature selection algorithm for linear classifiers. The feature selection and the classifier coefficient estimation steps were considered to be cascading stages in the classifier design process. We compared the performance of the classifier when feature selection was performed on the design samples alone and on the entire set of available samples, which consisted of design and test samples. The area Az under the receiver operating characteristic curve was used as our performance measure. After linear classifier coefficient estimation using the design samples, we studied the hold-out and resubstitution performance estimates. The two classes were assumed to have multidimensional Gaussian distributions, with a large number of features available for feature selection. We investigated the dependence of feature selection performance on the covariance matrices and means for the two classes, and examined the effects of sample size, number of available features, and parameters of stepwise feature selection on classifier bias. Our results indicated that the resubstitution estimate was always optimistically biased, except in cases where the parameters of stepwise feature selection were chosen such that too few features were selected by the stepwise procedure. When feature selection was performed using only the design samples, the hold-out estimate was always pessimistically biased. When feature selection was performed using the entire finite sample space, the hold-out estimates could be pessimistically or optimistically biased, depending on the number of features available for selection, the number of available samples, and their statistical distribution. For our simulation conditions, these estimates were always pessimistically (conservatively) biased if the ratio of the total number of available samples per class to the number of available features was greater than five.

Classification of malignant and benign masses based on hybrid ART2LDA approach

A new type of classifier combining an unsupervised and a supervised model was designed and applied to classification of malignant and benign masses on mammograms. The unsupervised model was based on an adaptive resonance theory (ART2) network which clustered the masses into a number of separate classes. The classes were divided into two types: one containing only malignant masses and the other containing a mix of malignant and benign masses. The masses from the malignant classes were classified by ART2. The masses from the mixed classes were input to a supervised linear discriminant classifier (LDA). In this way, some malignant masses were separated and classified by ART2 and the less distinguishable benign and malignant masses were classified by LDA. For the evaluation of classifier performance, 348 regions of interest (ROI's) containing biopsy proven masses (169 benign and 179 malignant) were used. Ten different partitions of training and test groups were randomly generated using an average of 73% of ROI's for training and 27% for testing. Classifier design, including feature selection and weight optimization, was performed with the training group. The test group was kept independent of the training group. The performance of the hybrid classifier was compared to that of an LDA classifier alone and a backpropagation neural network (BPN). Receiver operating characteristics (ROC) analysis was used to evaluate the accuracy of the classifiers. The average area under the ROC curve (A(z)) for the hybrid classifier was 0.81 as compared to 0.78 for the LDA and 0.80 for the BPN. The partial areas above a true positive fraction of 0.9 were 0.34, 0.27 and 0.31 for the hybrid, the LDA and the BPN classifier, respectively. These results indicate that the hybrid classifier is a promising approach for improving the accuracy of classification in CAD applications.

Markov random field modeling in posteroanterior chest radiograph segmentation

Previously, the authors presented an algorithm that identifies lung regions in a digitized posteroanterior chest radiograph (DCR) by labeling each pixel as either lung or nonlung. In this manuscript, the inherent flexibility of this algorithm is demonstrated as the algorithm is generalized to identify multiple anatomical regions in a DCR. Specifically, each pixel is classified as belonging to one of six anatomical region types: lung, subdiaphragm, heart, mediastinum, body, or background. The algorithm determines the optimal set of pixel classifications, xOPT, for a given set of DCR pixel gray level values y via a probabilistic approach that defines xOPT as the particular segmentation that maximizes the conditional distribution P(x/y). A spatially varying Markov random field (MRF) model is used that incorporates spatial and textural information of each possible region type. MRF modeling provides the form of P(x/y), and Iterated Conditional Modes is used to converge to the distribution maximum of P(x/y) thus obtaining the optimal segmentation for a given DCR. Results show the algorithm being able to correctly classify 90.0% +/- 3.4% of the pixels in a DCR.

Evaluation of Karhunen-Loève expansion for feature selection in computer-assisted classification of bioprosthetic heart-valve status

This paper analyses the performance of four different feature-selection approaches of the Karhunen-Loève expansion (KLE) method to select the most discriminant set of features for computer-assisted classification of bioprosthetic heart-valve status. First, an evaluation test reducing the number of initial features while maintaining the performance of the original classifier is developed. Secondly, the effectiveness of the classification in a simulated practical situation where a new sample has to be classified is estimated with a validation test. Results from both tests applied to a reference database show that the most efficient feature selection and classification (> or = 97% of correct classifications (CCs)) are performed by the Kittler and Young approach. For the clinical databases, this approach provides poor classification results for simulated 'new samples' (between 50 and 69% of CCs). For both the evaluation and the validation tests, only the Heydorn and Tou approach provides classification results comparable with those of the original classifier (a difference always < or = 7%). However, the degree of feature reduction is particularly variable. The study demonstrates that the KLE feature-selection approaches are highly population-dependent. It also shows that the validation method proposed is advantageous in clinical applications where the data collection is difficult to perform.

A pattern classification approach to characterizing solitary pulmonary nodules imaged on high resolution CT: preliminary results

The purpose of this research is to characterize solitary pulmonary nodules as benign or malignant based on quantitative measures extracted from high resolution CT (HRCT) images. High resolution CT images of 31 patients with solitary pulmonary nodules and definitive diagnoses were obtained. The diagnoses of these 31 cases (14 benign and 17 malignant) were determined from either radiologic follow-up or pathological specimens. Software tools were developed to perform the classification task. On the HRCT images, solitary nodules were identified using semiautomated contouring techniques. From the resulting contours, several quantitative measures were extracted related to each nodule's size, shape, attenuation, distribution of attenuation, and texture. A stepwise discriminant analysis was performed to determine which combination of measures were best able to discriminate between the benign and malignant nodules. A linear discriminant analysis was then performed using selected features to evaluate the ability of these features to predict the classification for each nodule. A jackknifed procedure was performed to provide a less biased estimate of the linear discriminator's performance. The preliminary discriminant analysis identified two different texture measures--correlation and difference entropy--as the top features in discriminating between benign and malignant nodules. The linear discriminant analysis using these features correctly classified 28/31 cases (90.3%) of the training set. A less biased estimate, using jackknifed training and testing, yielded the same results (90.3% correct). The preliminary results of this approach are very promising in characterizing solitary nodules using quantitative measures extracted from HRCT images. Future work involves including contrast enhancement and three-dimensional measures extracted from volumetric CT scans, as well as the use of several pattern classifiers.

Computerized analysis of abnormal asymmetry in digital chest radiographs: evaluation of potential utility

The purpose of this study was to develop and test a computerized method for the fully automated analysis of abnormal asymmetry in digital posteroanterior (PA) chest radiographs. An automated lung segmentation method was used to identify the aerated lung regions in 600 chest radiographs. Minimal a priori lung morphology information was required for this gray-level thresholding-based segmentation. Consequently, segmentation was applicable to grossly abnormal cases. The relative areas of segmented right and left lung regions in each image were compared with the corresponding area distributions of normal images to determine the presence of abnormal asymmetry. Computerized diagnoses were compared with image ratings assigned by a radiologist. The ability of the automated method to distinguish normal from asymmetrically abnormal cases was evaluated by using receiver operating characteristic (ROC) analysis, which yielded an area under the ROC curve of 0.84. This automated method demonstrated promising performance in its ability to detect abnormal asymmetry in PA chest images. We believe this method could play a role in a picture archiving and communications (PACS) environment to immediately identify abnormal cases and to function as one component of a multifaceted computer-aided diagnostic scheme.