Computer-Aided Detection of Ground Glass Nodules in Thoracic CT Images Using Shape, Intensity and Context Features

Abstract

Ground glass nodules (GGNs) occur less frequent in computed tomography (CT) scans than solid nodules but have a much higher chance of being malignant. Accurate detection of these nodules is therefore highly important. A complete system for computer-aided detection of GGNs is presented consisting of initial segmentation steps, candidate detection, feature extraction and a two-stage classification process. A rich set of intensity, shape and context features is constructed to describe the appearance of GGN candidates. We apply a two-stage classification approach using a linear discriminant classifier and a GentleBoost classifier to efficiently classify candidate regions. The system is trained and independently tested on 140 scans that contained one or more GGNs from around 10,000 scans obtained in a lung cancer screening trial. The system shows a high sensitivity of 73% at only one false positive per scan.

Computer Aided Diagnosis Using Multilevel Image Features on Large-Scale Evaluation

“…In colon CAD validation dataset, with feature integrated, our shape-gated main classifier improves the per-patient actionable polyp detection sensitivity from 82.50% [99/120] to 93.33% [112/120], along with per-volume FP rate dropping to 2.18 from 2.85. Furthermore, it also improves the performance for detecting the relatively more difficult subcategory of flat polyps from 68.29% [28/41] Comparison: Our CAD performances compare favorably against the state-of-thearts [1,2,5,8,12,14]. For Nodule detection, we achieve testing sensitivities of 90% (SN ≥ 3mm), 87.2% (PSN), 84.3% (GGN) and 78.6% (small with 3 ∼ 4 mm) at 4.1 FP/scan, while [2] reports sensitivity of 90.2% at 8.2 FP/scan; [5] obtains 90% sensitivity for noncalcified solid parenchymal nodules (≥ 4mm) at 5.1 FP/Scan.…”

“…For Nodule detection, we achieve testing sensitivities of 90% (SN ≥ 3mm), 87.2% (PSN), 84.3% (GGN) and 78.6% (small with 3 ∼ 4 mm) at 4.1 FP/scan, while [2] reports sensitivity of 90.2% at 8.2 FP/scan; [5] obtains 90% sensitivity for noncalcified solid parenchymal nodules (≥ 4mm) at 5.1 FP/Scan. [8] manually preselects 140 CT volumes with at least one (≥ 4mm) GGN nodule and reports sensitivity of 73% at 1 FP/Scan (77% ∼ 78% at 4 FP/Scan) for GGN nodule detection only. For comparison on dataset scales, [5] uses 60 nodules from 50 CT scans under a single imaging protocol.…”

“…where Our spatially aggregated image features capture the joint intensity/shape and classconditional probability (℘, υ) statistics of all ROI voxels, which has not been explicitly addressed previously [1,2,5,8,12,14]. These features model the high range/order spatial interactions among voxels within ROI, as an empirical distribution.…”

“…In this paper, we follow voxel classification framework [7,8], but our voxel-level labeling is data-driven and statistically learned from the annotated lesion image masks [6,9] on a number of CT scans. Only voxel probability assignment and thresholding (e.g., suppressing background clutter) are employed to obtain a lesion class-probability response map.…”

Discriminative Data Transform for Image Feature Extraction and Classification

“…The various features that are often used include the relatively traditional ones such as intensities [5], wavelets [4] and shapes [6]. The more recently proposed generic descriptors, such as the local binary patterns (LBP) and scale-invariant feature transform (SIFT), have also been incorporated to represent the texture and gradient information [11,12].…”