Lung segmentation in chest radiographs using anatomical atlases with nonrigid registration

Abstract

The National Library of Medicine (NLM) is developing a digital chest X-ray (CXR) screening system for deployment in resource constrained communities and developing countries worldwide with a focus on early detection of tuberculosis. A critical component in the computer-aided diagnosis of digital CXRs is the automatic detection of the lung regions. In this paper, we present a nonrigid registration-driven robust lung segmentation method using image retrieval-based patient specific adaptive lung models that detects lung boundaries, surpassing state-of-the-art performance. The method consists of three main stages: 1) a content-based image retrieval approach for identifying training images (with masks) most similar to the patient CXR using a partial Radon transform and Bhattacharyya shape similarity measure, 2) creating the initial patient-specific anatomical model of lung shape using SIFT-flow for deformable registration of training masks to the patient CXR, and 3) extracting refined lung boundaries using a graph cuts optimization approach with a customized energy function. Our average accuracy of 95.4% on the public JSRT database is the highest among published results. A similar degree of accuracy of 94.1% and 91.7% on two new CXR datasets from Montgomery County, MD, USA, and India, respectively, demonstrates the robustness of our lung segmentation approach.

Fig. 1.

Anatomical features in two chest X-ray images and their variations Differing lung shape, strong edges of the rib cage, visible shape of the heart, intensity variation around the clavicle bones and sharp corner at costophrenic angle that challenge automated segmentation algorithms. Both X-ray images are from the India dataset (see Section IV-A).

Fig. 2.

CBIR inspired work flow with nonrigid registration for identifying lung boundaries. The system consists of three stages: Stage-I) finding lung atlases similar to the patient X-ray using profile-based Bhattacharya similarity measures, Stage-II) computing a patient specific lung model by warping the training database of chest X-rays to the patient X-ray using the the SIFT-flow registration algorithm, and Stage-III) lung boundary detection using a graph cuts optimization approach with a customized energy function.

Fig. 3.

Plots show the Radon transform profiles for a query and database image, for $p_i\left(x\right)=R\left(\rho ,\theta =0°\right)$, left image, and $q_i\left(y\right)=R\left(\rho ,\theta =90°\right)$, for the right image.

Fig. 4.

(a)–(b) Pair of X-ray images from the JSRT dataset [13]. The right one (b) is the patient X-ray, and the left one (a) is the most similar X-ray to the patient X-ray in the database. Colored markers indicate corresponding matches based on SIFT-flow features for sample pixels. (c) Transformation mapping is applied to all pixels by shifting them according to spatial distances between the corresponding matches. (d) Warped mask.

Fig. 5.

(a) Red contour is the nonsmoothed boundary after the warping stage. (b)–(d) Blue dots are the critical points at different iterations. The green curve is the cubic spline interpolation of all critical points. The numbers of critical points on each lung are 100, 60, and 30, respectively. At each iteration, a point with the lowest relevance value is removed from the contour. Outer turn angles close to 180° and short line segments have a low relevance value. Note that, decreasing the number of critical points does not smooth the costophrenic angle region because of the sharp angle.

Fig. 6.

(a) Randomly selected chest X-ray image from the JSRT dataset. (b) Patient-specific lung model after registration. Each pixel intensity in the lung model image represents the probability of the pixel being part of the lung region.

Fig. 7.

(a) Top five training images using the partial Radon transform and Bhattacharyya shape similarity measure between the target patient CXR shown in Fig. 6(a) and the (JSRT) database. (b) Training masks corresponding to the five most similar X-rays. (c) Retrieved X-rays are warped using the calculated transformation mappings. Note that this warping is actually not needed in the algorithm, but is shown for illustrative purposes. (d) Training masks are warped to the target patient CXR. The average of these warped masks, shown in Fig. 6(b), forms the patient-specific lung model.

Fig. 8.

Overlap score of each image in the JSRT, Montgomery, and India sets. Each marker in the graph represents an X-ray image in the datasets. JSRT set contains 247, Montgomery set contains 138, and India set contains 200 chest X-rays.

Fig. 9.

Segmentation results on (a) JSRT, (b) Montgomery, and (c) India sets. Green and red contours indicate the gold standard and automatic segmentation results, respectively.

Fig. 10.

Particularly difficult lung segmentation cases. (a) The left diaphragm is elevated and there is a large air-distended colon loop below the lung boundary which is incorrectly combined with the lobe into a single region by the automatic algorithm. (b) Detected lung boundary includes the air cavity below left lung. (c)–(e) The algorithm could not detect the lung boundary correctly due to opacity caused by fluid in the lung space. The radiologist “estimated” the ground truth lung boundary (green contour).

Fig. 11.

(a) Segmentation performance and (b) execution time (in second) of the system with respect to the number of training masks. (Execution time is measured at resolution of 256 × 256.)

Fig. 12.

Segmenting the apex and costophrenic angle regions is more challenging than segmenting the other parts of the lung. These regions correspond approximately to the top 20% and bottom 20% of the lung.

Fig. 13.

Segmentation results for apical regions.

Fig. 14.

Segmentation results for costophrenic angle regions.