Roshni R. Bhagalia

Generic rebooting scheme and model-based probabilistic pruning algorithm for tree-like structure tracking

Tree-like vessel structures are an information-rich source for many image analysis tasks. Hence tracking algorithms extracting such structures have wide applicability. However, due to image artifacts and the minute nature of vessels, these algorithms face several challenges; two of the most common ones are 1) early termination, where tracking stops before the structure ends and 2) leaking, where tracking leaks into nearby closed organs or irrelevant structures. To address these issues, this paper makes two main contributions: a generic rebooting scheme that identifies early terminations and then restarts tracking to track objects in their entirety and a modelbased pruning algorithm that uses global optimization to identify and mitigate leaking. The performance of the proposed algorithm is demonstrated by tracking coronary arteries on 3D cardiac Computed Tomography Angiography (CTA) data from 28 human subjects. Our methods dramatically improve tracking results by detecting and recovering from early terminations and identifying and removing leaking in 98% (63 of 64) branches, with a single erroneously removed valid branch.

Coronary artery motion estimation and compensation: A feasibility study

High temporal resolution and high spatial resolution are required to image the coronary arteries without motion artifacts. Several approaches have been pursued to achieve better temporal resolution including faster rotational speeds, and dual tube systems. In this paper, we present an alternative approach using motion estimation and compensation. The results demonstrate that the proposed methods can significantly reduce motion artifacts in coronary artery imaging.

Organ Localization Using Joint AP/LAT View Landmark Consensus Detection and Hierarchical Active Appearance Models

Parsing 2D radiographs into anatomical regions is a challenging task with many applications. In the clinic, scans routinely include anterior-posterior (AP) and lateral (LAT) view radiographs. Since these orthogonal views provide complementary anatomic information, an integrated analysis can afford the greatest localization accuracy. To solve this integration we propose automatic landmark candidate detection, pruned by a learned geometric consensus detector model and refined by fitting a hierarchical active appearance organ model (H-AAM). Our main contribution is twofold. First, we propose a probabilistic joint consensus detection model which learns how landmarks in either or both views predict landmark locations in a given view. Second, we refine landmarks by fitting a joint H-AAM that learns how landmark arrangement and image appearance can help predict across views. This increases accuracy and robustness to anatomic variation. All steps require just seconds to compute and compared to processing the scouts separately, joint processing reduces mean landmark distance error from 27.3 mm to 15.7 mm in LAT view and from 12.7 mm to 11.2 mm in the AP view. The errors are comparable to human expert inter-observer variability and suitable for clinical applications such as personalized scan planning for dose reduction. We assess our method using a database of scout CT scans from 93 subjects with widely varying pathology.

Learning orientation invariant contextual features for nodule detection in lung CT scans

This work combines model-based local shape analysis and data-driven local contextual feature learning for improved detection of pulmonary nodules in low dose computed tomography (LDCT) chest scans. We reduce orientation-induced appearance variability by performing intensity-weighted principal component analysis (PCA) to estimate the local orientation at each candidate location. Random comparison primitives defined in a local coordinate system are used to describe the local context around a nodule candidate. A random forest is trained to learn and combine a subset of these primitives into discriminative orientation invariant contextual features and classify nodule candidates. Validation using 99 CT scans from the publicly available Lung Image Database Consortium (LIDC) demonstrates the benefit of combining geometric modeling and data-driven machine learning. The proposed method reduces more than 80% of false positives of the baseline model-based method consistently over a wide range of sensitivity levels (70%-90%).

Bi-directional labeled point matching

Robust point matching (RPM) simultaneously estimates correspondences and non-rigid warps between unstructured point-sets. While RPM is robust to outliers in the target (fixed) point-set, its performance degrades when the template (moving) point-set contains outliers. Additionally, RPM does not utilize information about the topological structure or group memberships of the data it is matching. Bi-directional Labeled Point Matching (BLPM) extends the RPM objective function by introducing (i) Bi-Directionality (BD) and (ii) a Label Entropy (LE) term. BD aids in outlier rejection in both point-sets and LE discourages mappings that transform points within a single group in one point-set onto points from multiple distinct groups in the other point-set. The resulting BLPM algorithm translates into simple modifications to the standard RPM update rules.

Improved Robust Point Matching with Label Consistency

Robust point matching (RPM) jointly estimates correspondences and non-rigid warps between unstructured point-clouds. RPM does not, however, utilize information of the topological structure or group memberships of the data it is matching. In numerous medical imaging applications, each extracted point can be assigned group membership attributes or labels based on segmentation, partitioning, or clustering operations. For example, points on the cortical surface of the brain can be grouped according to the four lobes. Estimated warps should enforce the topological structure of such point-sets, e.g. points belonging to the temporal lobe in the two point-sets should be mapped onto each other. We extend the RPM objective function to incorporate group membership labels by including a Label Entropy (LE) term. LE discourages mappings that transform points within a single group in one point-set onto points from multiple distinct groups in the other point-set. The resulting Labeled Point Matching (LPM) algorithm requires a very simple modification to the standard RPM update rules. We demonstrate the performance of LPM on coronary trees extracted from cardiac CT images. We partitioned the point sets into coronary sections without a priori anatomical context, yielding potentially disparate labelings (e.g. [1,2,3] → [a,b,c,d]). LPM simultaneously estimated label correspondences, point correspondences, and a non-linear warp. Non-matching branches were treated wholly through the standard RPM outlier process akin to non-matching points. Results show LPM produces warps that are more physically meaningful than RPM alone. In particular, LPM mitigates unrealistic branch crossings and results in more robust non-rigid warp estimates.

Lung nodule segmentation using deep learned prior based graph cut

We propose an automated framework for lung nodule segmentation from pulmonary CT scan using graph cut with a deep learned prior. The segmentation problem is formulated as a hybrid cost function minimization task, which combines a domain specific data term with a deep learned probability map. The proposed segmentation framework embodies the robustness of deep learning in object localization, while retaining the hallmark of traditional segmentation models in addressing the morphological intricacies of elaborate objects. The proposed solution offers more than 20% performance improvement over a contemporary data driven model, and also outperforms traditional graph cuts especially in situations where model initialization is slightly inaccurate.

Improved characterization of molecular phenotypes in breast lesions using 18F-FDG PET image homogeneity

Positron emission tomography (PET) using uorodeoxyglucose (18F-FDG) is commonly used in the assessment of breast lesions by computing voxel-wise standardized uptake value (SUV) maps. Simple metrics derived from ensemble properties of SUVs within each identified breast lesion are routinely used for disease diagnosis. The maximum SUV within the lesion (SUVmax) is the most popular of these metrics. However these simple metrics are known to be error-prone and are susceptible to image noise. Finding reliable SUV map-based features that correlate to established molecular phenotypes of breast cancer (viz. estrogen receptor (ER), progesterone receptor (PR) and human epidermal growth factor receptor 2 (HER2) expression) will enable non-invasive disease management. This study investigated 36 SUV features based on first and second order statistics, local histograms and texture of segmented lesions to predict ER and PR expression in 51 breast cancer patients. True ER and PR expression was obtained via immunohistochemistry (IHC) of tissue samples from each lesion. A supervised learning, adaptive boosting-support vector machine (AdaBoost-SVM), framework was used to select a subset of features to classify breast lesions into distinct phenotypes. Performance of the trained multi-feature classifier was compared against the baseline single-feature SUVmax classifier using receiver operating characteristic (ROC) curves. Results show that texture features encoding local lesion homogeneity extracted from gray-level co-occurrence matrices are the strongest discriminator of lesion ER expression. In particular, classifiers including these features increased prediction accuracy from 0.75 (baseline) to 0.82 and the area under the ROC curve from 0.64 (baseline) to 0.75.

Cardiac localization in topograms using hierarchical models

A vast number of medical imaging protocols identify anatomical regions of interest (ROI) from two dimensional (2D) localizer images to aid high resolution scan planning. These localizer scans are typically two dimensional projections of three dimensional data and as such have lower image detail due to overlapping tissue. The problem is further complicated by large variations in shape, size, appearance and the high occurrence of anomalies in the human anatomy. Manual ROI delineation is time consuming and error prone. To combat these issues we develop a hierarchical multi-object active appearance model (AAM) framework that is both robust to inaccuracies in model initialization yet sufficiently flexible to handle the large diversity of the human body. The method was successfully applied to automatically determine the extents of the human heart in 99 2D CT topograms yielding significant improvement in accuracy over a single global AAM approach.

Improved noninvasive prostate cancer assessment using multiparametric magnetic resonance imaging

This study explores potential multi-parametric magnetic resonance imaging (mpMRI) biomarkers for improved non-invasive prostate cancer (PCa) management, particularly for clinically significant PCa lesions that are missed by the current standard of care mpMRI exams. T2-weighted MRI and diffusion weighted imaging scans were acquired for subjects with Gleason Score 6 to 9 PCa lesions prior to radical prostatectomy. Lesions were examined by experienced clinicians and categorized as those visible on routine mpMRI radiological assessment, viz. mpMRI-detectable (mpMRI-d) and those that could not be satisfactorily diagnosed using routine mpMRI, viz. mpMRI-undetectable (mpMRI-u). A dense set of voxel-wise mpMRI features quantifying local spatial distribution, texture and statistical properties in the prostate were computed and analyzed for their potential to differentiate disease from healthy tissue. Feature selection and training were performed to build classifiers to produce voxel-wise probability-of-malignancy maps over the entire gland to characterize the extent of PCa. Receiver operating characteristic analysis on a cohort of 48 mpMRI exams showed that our classifier improves the sensitivity of the detection of prostate lesions by over 38% for the whole cohort and 124% for the mpMRI-u lesions with only a modest loss in specificity when compared to baseline routine radiological mpMRI assessment.