Improving performance of computer-aided detection of pulmonary embolisms by incorporating a new pulmonary vascular-tree segmentation algorithm

Abstract

We developed a new pulmonary vascular tree segmentation/extraction algorithm. The purpose of this study was to assess whether adding this new algorithm to our previously developed computer-aided detection (CAD) scheme of pulmonary embolism (PE) could improve the CAD performance (in particular reducing false positive detection rates). A dataset containing 12 CT examinations with 384 verified pulmonary embolism regions associated with 24 threedimensional (3-D) PE lesions was selected in this study. Our new CAD scheme includes the following image processing and feature classification steps. (1) A 3-D based region growing process followed by a rolling-ball algorithm was utilized to segment lung areas. (2) The complete pulmonary vascular trees were extracted by combining two approaches of using an intensity-based region growing to extract the larger vessels and a vessel enhancement filtering to extract the smaller vessel structures. (3) A toboggan algorithm was implemented to identify suspicious PE candidates in segmented lung or vessel area. (4) A three layer artificial neural network (ANN) with the topology 27-10-1 was developed to reduce false positive detections. (5) A k-nearest neighbor (KNN) classifier optimized by a genetic algorithm was used to compute detection scores for the PE candidates. (6) A grouping scoring method was designed to detect the final PE lesions in three dimensions. The study showed that integrating the pulmonary vascular tree extraction algorithm into the CAD scheme reduced false positive rates by 16.2%. For the case based 3D PE lesion detecting results, the integrated CAD scheme achieved 62.5% detection sensitivity with 17.1 false-positive lesions per examination.

PENet—a scalable deep-learning model for automated diagnosis of pulmonary embolism using volumetric CT imaging

“…[21][22][23][24] Prior efforts toward automation of PE diagnosis using CTPA have focused on traditional feature engineering methodologies; while the higher performing among these have reported sensitivity as high as 75% for diagnosing PE on CTPA studies, each have significant drawbacks due to the relatively high burden of development and implementation, including manual feature engineering, complex preprocessing adding significant time and infrastructure costs, and a lack of external validation to understand the generalizability and identify overfitting. [25][26][27][28][29][30] In contrast, advancements in deep learning possess inherent advantages over prior approaches due, in part, to obviating the need for hand crafted feature engineering and flexibility as an "end-to-end" classification solution.…”

Survey on deep learning for pulmonary medical imaging

“…This study involves a structure-based analysis algorithm. For CTPA image classification, the CAD of PE typically consists of the four following stages: (1) extraction of a volume of interest (VOI) from the original dataset via lung [105][106][107] or vessel segmentation [108,109], (2) generation of a set of PE candidates within the VOI using algorithms, such as tobogganing [110] and extracting hand-crafted features from each PE candidate [111,112], and (3) computation of a confidence score for each candidate by using a rule-based classifier, neural networks and a nearest neighbor [106,108,113] or multi-instance classifier [110]. Jinbo Bi [96] proposed a new classification method for the automatic detection of PE.…”

Enhanced Classification Performance using Deep Learning Based Segmentation for Pulmonary Embolism Detection in CT Angiography

“…Developing a general solution for automatic detection of PE has proven challenging because of anatomical variation, motion and breathing artifacts, inter-patient variability in contrast medium concentration, and concurrent pathologies. Over the past two decades, automated PE detection has been attempted using deterministic models, such as image processing and analysis techniques [10,11], or probabilistic/statistical models such as machine learning [12][13][14] and deep convolutional neural networks [15,16]. Yet, the accuracies of these solutions have been insufficient for clinical use due to low sensitivity [10,13,15] and high false positive rate [10,11,13,14], potentially caused by training on small datasets [10,11,[13][14][15].…”

“…Over the past two decades, automated PE detection has been attempted using deterministic models, such as image processing and analysis techniques [10,11], or probabilistic/statistical models such as machine learning [12][13][14] and deep convolutional neural networks [15,16]. Yet, the accuracies of these solutions have been insufficient for clinical use due to low sensitivity [10,13,15] and high false positive rate [10,11,13,14], potentially caused by training on small datasets [10,11,[13][14][15]. The state-of-the-art is a residual neural network (ResNet) classification architecture on 1465 CTPA examinations with sensitivity of 92.7% and specificity of 95.5% [17].…”