Anna Jerebko

Supervised learning from multiple experts: whom to trust when everyone lies a bit

We describe a probabilistic approach for supervised learning when we have multiple experts/annotators providing (possibly noisy) labels but no absolute gold standard. The proposed algorithm evaluates the different experts and also gives an estimate of the actual hidden labels. Experimental results indicate that the proposed method is superior to the commonly used majority voting baseline.

Segmentation of pulmonary nodules of various densities with morphological approaches and convexity models

Accurate segmentation of a pulmonary nodule is an important and active area of research in medical image processing. Although many algorithms have been reported in literature for this problem, those that are applicable to various density types have not been available until recently. In this paper, we propose a new algorithm that is applicable to solid, non-solid and part-solid types and solitary, vascularized, and juxtapleural types. First, the algorithm separates lung parenchyma and radiographically denser anatomical structures with coupled competition and diffusion processes. The technique tends to derive a spatially more homogeneous foreground map than an adaptive thresholding based method. Second, it locates the core of a nodule in a manner that is applicable to juxtapleural types using a transformation applied on the Euclidean distance transform of the foreground. Third, it detaches the nodule from attached structures by a region growing on the Euclidean distance map followed by a procedure to delineate the surface of the nodule based on the patterns of the region growing and distance maps. Finally, convex hull of the nodule surface intersected with the foreground constitutes the final segmentation. The performance of the technique is evaluated with two Lung Imaging Database Consortium (LIDC) data sets with 23 and 82 nodules each, and another data set with 820 nodules with manual diameter measurements. The experiments show that the algorithm is highly reliable in segmenting nodules of various types in a computationally efficient manner.

Symmetric curvature patterns for colonic polyp detection

A novel approach for generating a set of features derived from properties of patterns of curvature is introduced as a part of a computer aided colonic polyp detection system. The resulting sensitivity was 84% with 4.8 false positives per volume on an independent test set of 72 patients (56 polyps). When used in conjunction with other features, it allowed the detection system to reach an overall sensitivity of 94% with a false positive rate of 4.3 per volume.

Patchwork reconstruction with resolution modeling for digital breast tomosynthesis

PURPOSE Digital breast tomosynthesis is a relatively new diagnostic x-ray modality that allows high resolution breast imaging while suppressing interference from overlapping anatomical structures. However, proper visualization of microcalcifications remains a challenge. For the subset of systems considered by the authors, the main cause of deterioration is movement of the x-ray source during exposures. They propose a modified grouped coordinate ascent algorithm that includes a specific acquisition model to compensate for this deterioration. METHODS A resolution model based on the movement of the x-ray source during image acquisition is created and combined with a grouped coordinate ascent algorithm. Choosing planes parallel to the detector surface as the groups enables efficient implementation of the position dependent resolution model. In the current implementation, the resolution model is approximated by a Gaussian smoothing kernel. The effect of the resolution model on the iterative reconstruction is evaluated by measuring contrast to noise ratio (CNR) of spherical microcalcifications in a homogeneous background. After this, the new reconstruction method is compared to the optimized filtered backprojection method for the considered system, by performing two observer studies: the first study simulates clusters of spherical microcalcifications in a power law background for a free search task; the second study simulates smooth or irregular microcalcifications in the same type of backgrounds for a classification task. RESULTS Including the resolution model in the iterative reconstruction methods increases the CNR of microcalcifications. The first observer study shows a significant improvement in detection of microcalcifications (p = 0.029), while the second study shows that performance on a classification task remains the same (p = 0.935) compared to the filtered backprojection method. CONCLUSIONS The new method shows higher CNR and improved visualization of microcalcifications in an observer experiment on synthetic data. Further study of the negative results of the classification task showed performance variations throughout the volume linked to the changing noise structure introduced by the combination of the resolution model and the smoothing prior.

Out-of-plane artifact reduction in tomosynthesis based on regression modeling and outlier detection

We propose a method for out-of-plane artifact reduction in digital breast tomosynthesis reconstruction. Because of the limited angular range acquisition in DBT, the reconstructed slices have reduced resolution in z-direction and are affected by artifacts. The out-of-plane blur caused by dense tissue and large masses complicates reconstruction of thick slices volumes. The streak-like out-of-plane artifacts caused by calcifications and metal clips distort the shape of calcifications which is regarded by many radiologists as an important malignancy predictor. Small clinical features such as micro-calcifications could be obscured by bright artifacts. The proposed technique involves reconstructing a set of super-resolution slices and predicting the artifact-free voxel intensity based on the corresponding set of projection pixels using a statistical model learned from a set of training data. Our experiments show that the resulting reconstructed images are de-blurred and streak-like artifacts are reduced, visibility of clinical features, contrast and sharpness are improved and thick-slice reconstruction is possible without the loss of contrast and sharpness.

Regularization parameter selection in maximum a posteriori iterative reconstruction for digital breast tomosynthesis

The method presented in this paper addresses the problem of regularization parameter selection in maximum a posteriori iterative reconstruction for digital breast tomosynthesis The method allows analytically deriving the combination of prior function parameters for noise level expected in the reconstruction without priors and estimated breast density such that it effectively controls the level of noise while preserving the edges of breast structures Results show reduced noise level and improved contrast to noise ratio compared to filtered back projection and maximum–likelihood iterative reconstruction without penalizing term.

Lung Nodule Detection

The quantity of digital medical images that must be reviewed by radiologists as part of routine clinical practice has greatly increased in recent years. New acquisition devices generate images that have higher spatial resolution, both in 2–D as well as 3–D, requiring physicians to use more sophisticated visualization tools. In addition, advanced visualization systems, designed to assist the radiologist, are now part of a standard arsenal of tools which, together with workflow improvements, aid the physicians in their clinical tasks. Computer–Assisted Diagnosis (CAD) systems are one of such class of sophisticated tools to support the radiologists in tedious and time–consuming tasks such as the detection of lesions. Over the past ten years, CAD systems have evolved to reach sensitivity capabilities equivalent to or exceeding that of a radiologist, thus becoming clinically acceptable, but with limited specificity which necessitates their use as a second reader tool. This chapter presents one such system (LungCAD) designed for the detection of nodules in the lung parenchyma. Its performance was evaluated as part of a detection challenge organized by ImageCLEF 2009.

Multi network classification scheme for detection of colonic polyps in CT colonography data sets

A multi-network decision classification scheme for colonic polyp detection is presented. The approach is based on the results of voting over several neural networks using variable subsets selected from a general set. We used 21 features including region density, Gaussian and mean curvature and sphericity, lesion size, colon wall thickness, and their means and standard deviations. The subsets of variables are weighted by their effectiveness calculated on the basis of the training and test sample misclassification rates. The final decision is based on the majority vote across the networks and takes into account the weighted votes of all nets. This method reduces the false positive rate by a factor of 1.7 compared to single net decisions. The overall sensitivity and specificity rates reached are 100% and 95% correspondingly. Back propagation neural nets trained with the Levenberg-Marquardt algorithm were used. Ten-fold cross-validation is applied to better estimate the true error rates.

Data-Driven breast decompression and lesion mapping from digital breast tomosynthesis

Digital breast tomosynthesis (DBT) emerges as a new 3D modality for breast cancer screening and diagnosis. Like in conventional 2D mammography the breast is scanned in a compressed state. For orientation during surgical planning, e.g., during presurgical ultrasound-guided anchor-wire marking, as well as for improving communication between radiologists and surgeons it is desirable to estimate an uncompressed model of the acquired breast along with a spatial mapping that allows localizing lesions marked in DBT in the uncompressed model. We therefore propose a method for 3D breast decompression and associated lesion mapping from 3D DBT data. The method is entirely data-driven and employs machine learning methods to predict the shape of the uncompressed breast from a DBT input volume. For this purpose a shape space has been constructed from manually annotated uncompressed breast surfaces and shape parameters are predicted by multiple multi-variate Random Forest regression. By exploiting point correspondences between the compressed and uncompressed breasts, lesions identified in DBT can be mapped to approximately corresponding locations in the uncompressed breast model. To this end, a thin-plate spline mapping is employed. Our method features a novel completely data-driven approach to breast shape prediction that does not necessitate prior knowledge about biomechanical properties and parameters of the breast tissue. Instead, a particular deformation behavior (decompression) is learned from annotated shape pairs, compressed and uncompressed, which are obtained from DBT and magnetic resonance image volumes, respectively. On average, shape prediction takes 26s and achieves a surface distance of 15.80 +/- 4.70 mm. The mean localization error for lesion mapping is 22.48 +/- 8.67 mm.

Evaluation and optimization of the maximum-likelihood approach for image reconstruction in digital breast tomosynthesis

Digital Breast Tomosynthesis (DBT) suffers from incomplete data and poor quantum statistics limited by the total dose absorbed in the breast. Hence, statistical reconstruction assuming the photon statistics to follow a Poisson distribution may have some advantages. This study investigates state-of-art iterative maximum likelihood (ML) statistical reconstruction algorithms for DBT and compares the results with simple backprojection (BP), filtered backprojection (FBP), and iFBP (FBP with filter derived from iterative reconstruction). The gradient-ascent and convex optimization variants of the transmission ML algorithm are evaluated with phantom and clinical data. Convergence speed is very similar for both iterative statistical algorithms and after approximately 5 iterations all significant details are well displayed, although we notice increasing noise. We found empirically that a relaxation factor between 0.25 and 0.5 provides the optimal trade-off between noise and contrast. The ML-convex algorithm gives smoother results than the ML-gradient algorithm. The low-contrast CNR of the ML algorithms is between CNR for simple backprojection (highest) and FBP (lowest). Spatial resolution of iterative statistical and iFBP algorithms is similar to that of FBP but the quantitative density representation better resembles conventional mammograms. The iFBP algorithm provides the benefits of statistical iterative reconstruction techniques and requires much shorter computation time.