Research progress of computer aided diagnosis system for pulmonary nodules in CT images

Abstract

BACKGROUND AND OBJECTIVE:

Since CAD (Computer Aided Diagnosis) system can make it easier and more efficient to interpret CT (Computer Tomography) images, it has gained much attention and developed rapidly in recent years. This article reviews recent CAD techniques for pulmonary nodule detection and diagnosis in CT Images.

METHODS:

CAD systems can be classified into computer-aided detection (CADe) and computer-aided diagnosis (CADx) systems. This review reports recent researches of both systems, including the database, technique, innovation and experimental results of each work. Multi-task CAD systems, which can handle segmentation, false positive reduction, malignancy prediction and other tasks at the same time. The commercial CAD systems are also briefly introduced.

RESULTS:

We have found that deep learning based CAD is the mainstream of current research. The reported sensitivity of deep learning based CADe systems ranged between 80.06% and 94.1% with an average 4.3 false-positive (FP) per scan when using LIDC-IDRI dataset, and between 94.4% and 97.9% with an average 4 FP/scan when using LUNA16 dataset, respectively. The overall accuracy of deep learning based CADx systems ranged between 86.84% and 92.3% with an average AUC of 0.956 reported when using LIDC-IDRI dataset.

CONCLUSIONS:

We summarized the current tendency and limitations as well as future challenges in this field. The development of CAD needs to meet the rigid clinical requirements, such as high accuracy, strong robustness, high efficiency, fine-grained analysis and classification, and to provide practical clinical functions. This review provides helpful information for both engineering researchers and radiologists to learn the latest development of CAD systems.

Keywords

Pulmonary nodules computer-aided detection (CADe)computer-aided diagnosis (CADx)multi-task CAD

1 Introduction

The latest study points out that lung cancer is still the leading cause of cancer death with the highest morbidity and mortality around the world, accounting for 11.6% of the total cancer population [1]. Early detection and early diagnosis can increase the survival rate of lung cancer. The early manifestations of most lung cancer are pulmonary nodules which can be detected by computed tomography (CT), positron emission tomography-computed tomography (PET-CT) and magnetic resonance imaging (MRI). Currently CT scan is the standard imaging modality for pulmonary nodules due to its advantages of high spatial resolution and rapid acquisition [2].

Scientifically, pulmonary nodule is defined as rounded or irregular lesion and its diameter should be less or equal to 3 cm in the chest. Imaging findings of pulmonary nodule include solitary or multiple lesions with clear or obscure border and higher density [2] as shown in Fig. 1. Lesions larger than 3 cm in diameter are pulmonary masses and beyond the scope of this paper. Additionally, pulmonary nodules can be categorized based on their size (large or small), location (well circumscribed, juxta-pleural, juxta-vascular and nodule with pleural tail), density (solid nodule, subsolid nodule and pure ground glass density nodule), shape (spherical, polygonal and irregular), etc.

Fig.1

Different types of pulmonary nodules: (a) Single and small size (≥4 mm, <10 mm), (b) Single and medium size (≥10 mm, <30 mm), (c) Non-spherical shape, (d) Ground-glass-opacity, (e) Multiple and small size.

Most pulmonary nodules are benign. The image features on CT, such as size, calcification, internal structure, lobulation, speculation, texture and subtlety, are important basises for the differential diagnosis of benign and malignant pulmonary nodules. On follow-up CT scan growth rate is the main feature for benign and malignant diagnosis. Clinical screening of lung cancer requires comprehensive evaluation of the individual and clinical characteristics, e.g. age, smoking history, family cancer history, etc. For indeterminate pulmonary nodules, pathological diagnosis is the golden standard.

With the development of CT equipment and the popularity of CT screening in the diagnosis of lung cancer, the amount of image data has increased year by year, greatly increasing the workload of radiologists. Doctors’ fatigue, distraction or inadequate experience would affect the accuracy of clinical diagnosis. The variety of pulmonary nodules also makes detecting and diagnosing difficult. Computer aided diagnosis (CAD) can help the radiologist to reduce reading time, assist in early detection of cancer and increase accuracy in diagnosis [3]. Research [4] has found that CAD can significantly improve the diagnostic performance for interns and residents, but provide minimal benefit for chest radiologist and senior radiologists. The CAD proposed by Huang et al. [5] could increase PPV and reduce FP rate in the early diagnosis of lung cancer. The cloud-based CAD-assisted reading can improve detection of ≥3 mm pulmonary nodules on CT scans, but slightly increase reading time [6]. Although there are differences across the reported performances of different CAD systems, the perception is that using CAD can make it easier to interpret chest CT scans and is one main way of improving lung cancer screening in future [7]. Generally, CAD consists of the following seven major components:

Data acquisition: Many public databases can be utilized to develop CAD systems and to compare the performances persuasively. Listed below are several common databases: LIDC (Lung Image Database Consortium)[8], LIDC-IDRI (Lung Image Database Consortium and Image Database Resource Initiative)[9], LUNA16 (Lung Nodule Analysis 2016) [10], NELSON (Nederland-Leuvens Longkanker Onderzoek) [11], ELCAP (Early Lung Cancer Action Program)[12], ANODE09 (Automatic Nodule Detection 2009)[13], NLST (National Lung Screening Trial)[14]. CT images are often divided into lots of patches or cubes as the input data of CAD. Deep learning based CAD requires vast amounts of data to train the network parameters. Insufficient data will lead to problems such as over-fitting or non-convergence. Data augmentation can effectively prevent over-fitting and misleading by adding in-variances to existing data [15]. Transfer learning is another common way to alleviate the insufficient training data problem [16]. For incomplete labeled data, semi-supervised learning is an effective way to train the models [17].

Preprocessing: The purposes of preprocessing include improving the quality of raw CT images, making data isotropic, etc. Listed below are common preprocessing algorithms: curvature flow[18] for removing noise; contrast stretching [19], discrete wavelet transform (DWT) and unsharp energy mask (UEM) [20] for enhancing contrast. Converted the raw data into Hounsfield unit (HU) and then clamped to [–1000,1000] HU is another effective way to remove noises [15, 21]. CT images with different inner plane spacing could be re-sampled to a unified inner plane spacing [15]. Preprocessing component is not essential for CAD system, since CT equipment have automatically optimized image quality during image generation in clinic.

Lung segmentation: Accurately extracting lung parenchyma from other organs and tissues is a crucial component in CAD system. Region growth algorithm [22, 23], morphological filter [24], OTSU threshold algorithm [23] and active contour model [25, 26] are commonly used lung segmentation methods. Juxta-pleural nodules may be wrongly segmented into the outside of lung parenchyma, thus many algorithms have been proposed to fix this problem, such as rolling ball algorithm [27] and chain code [28]. To avoid the error caused by anisotropic representation of grids in 3D Cartesian coordinates, 3D CAD systems [25, 29] have resampled the lung volume data by interpolation algorithms, e.g. bicubic interpolation algorithm[23], to obtain isotropic data before lung segmentation.

Pulmonary nodule detection: this step aims to find the locations of suspicious pulmonary nodules from a volume of interest (such as lung parenchyma or 3D lung volume). To ensure the high sensitivity of CAD, detection algorithms should be able to find as many suspicious nodules as possible. Generally, there are two categories of detection algorithms: rule-based schemes and deep learning based schemes. Density and shape are two common rules for detection. Since the density of nodules is usually higher than the lung parenchyma, segmentation algorithms, e.g. VQ algorithm [30], have shown competitive advantage in nodule segmentation. Most nodules are approximate spheres or ellipsoids while vessel has a tubular shape, thus spherical filter [26], skeletonization [25] and 3D tensor filter [23] are common shape-based detection algorithms. However, different types of pulmonary nodules have large variations in size, shape and density, e.g. juxta-pleural nodules usually have semispherical shape, and thus rule-based schemes can only apply to certain types of nodules. Deep learning based schemes, which learn from large scale of training data to adjust parameters of the hierarchical network structure, are general detection schemes. Several deep learning models have been utilized for pulmonary nodule detection problem, e.g. 3D faster R-CNN (Region-Convolutional Neural Networks) [31]. The quality and quantity of training data will greatly affect the performance of deep learning based schemes. In addition, since vessels are noticeable structures in lung parenchyma, several methods have been proposed to weaken the vascular structures to make nodules highlight, e.g. Frangi filter [22] and vessel suppressed function[32]. A high number of false positive nodules in this step have confused the radiologist and hindered the widespread usage of CAD.

False positive reduction: in order to reduce the false positive rate and increase the specificity of CAD system, this step is to classify true pulmonary nodules from all suspicious nodules as much as possible. According to the anatomical characteristics of pulmonary nodules, such as shape and curvedness, a portion of false positives can be removed firstly [23]. In general, false positive reduction consists of feature extraction, feature selection and classification stages. The commonly used features include shape or geometry, intensity and texture [3]. Feature selection aims to build an efficient feature set which can improve the efficiency and accuracy, such as CFS (Correlation Feature Selection). Popular classifiers include SVM (Support Vector Machine) [25], decision tree, random forest [23], neural network classifier [25], etc. Deep learning approaches can learn discriminative features automatically, and common models, such as CNN, Deep Reinforcement Learning [33] and 3D residual CNN [29], have achieved better performance than feature extraction based methods in false positive reduction task.

Characterization of pulmonary nodules: imaging characteristics, such as lobulation, size and so on, are associated with lung cancer. Growth rate is the main characteristic for lung cancer diagnosis on follow up CT scan. Analyzing and quantifying valuable characteristics for diagnosis is the objective of this step. For example, effective radius, elongation, compactness and Geometric moment are commonly used quantitative indicators to the shape characteristic [25]. Radiomics [34] and Delta Radiomics [35] (i.e., changes in features over time) are the process of extracting and analyzing image-based, quantitative features from a region of interest which then can be analyzed to develop decision support tools. These quantitative image-based features characterize size, shape, volume, and texture of pulmonary nodule and improve the malignancy prediction in lung cancer screening [35].

Classification of pulmonary nodules: the objective of this step includes automatically diagnosing the type of pulmonary nodules and predicting the probability of malignancy. With features extracted from pulmonary nodules, traditional classifiers can be trained for this task. In current research, the classification accuracy with deep learning models, such as Resnet [36], VGG [37], Alexnet [38] and MoDenseNet [39], is higher than traditional classifiers.

However, the performance of false positive reduction is not good enough to support subsequent components, thus current CAD system has been studied as two subsystems separately: computer-aided detection (CADe) system which mainly focuses on pulmonary nodule detection and false positive reduction; computer-aided diagnosis (CADx) system which mainly focuses on the characterization and classification of pulmonary nodules. This review reports the recent research progress of those two systems, as well as multi-task CAD system and commercial CAD systems in section 2. In section 3, we discuss the limitations of current CAD and the future challenges. Finally, the achievements of this paper have been concluded in section 4.

2 Detailed report of latest CAD Systems

Two subsystems of CAD are introduced firstly: CADe system identifies the locations of suspicious lesions and then CADx system characterizes and classifies the lesions [3]. We have also investigated multi-task CAD system which can handle segmentation, false positive reduction, malignancy prediction and other tasks at the same time. At the end of this section, typical commercial CAD systems are briefly introduced. Latest high-quality studies are selected and summarized, as listed in Table 1.

Table 1
Comparison of the reviewed computer-aided detection (CADe) and diagnosis (CADx) of Pulmonary nodules using CT images

Authors Year Database Task Technique Experiential result Nodule type

Gong et al. [23] 2018 (1) LUNA 16 CADe 3D tensor filter (1) sensitivity of 79.3% with 4FP/scan Spherical shape

(2) ANODE09 19 selected optimal features (2) sensitivity of 84.62% with 2.8 FP/scan

Gupta et al. [40] 2018 LIDC-IDRI CADe A rich set of 515 features with neural network classifier Training set Multi-size nodules

Sensitivity of 85.6% with 8FP/scan

Sensitivity of 83.5% with 1 FP/scan

Testing set

Sensitivity of 68.4% with 8FP/scan

GU et al. [40] 2019 154 thin-slice scans in LIDC CADe Vessel tree group Multi-scale dot filter Sensitivity of 87.81% with 1.057 FP/scan Juxta-vascular nodules

Zhang et al. [25] 2018 71 subjects of LIDC CADe 3D skeletonization feature Ten features with SVM Sensitivity of 89.3% with 2.1 FP/scan accuracy of 87.6% Juxta-vascular nodules

Zhang et al. [22] 2018 LIDC-IDRI CADe Four CNN structures Sensitivity of 80.06% with 4.7 FP/scan Four size ranges of nodules

Sensitivity of 94% with 15.1 FP/scan

Tan et al. [30] 2019 LIDC-IDRI CADe Four CNN models Sensitivity of 88% with 1.9 FP/scan Irregularly-shape juxta-pleural nodule

Sensitivity of 94.1% with 4.01 FP/scan

Sakamoto et al. [41] 2018 LUNA2016 CADe Cascaded CNN Sensitivity of 94.4% with 4 FP/scan General nodules

Sensitivity of 95.9% with 8 FP/scan

Jin et al. [29] 2018 LUNA16 CADe Deep 3D residual CNN Sensitivity of 82.3% with 0.125 FP/scan General nodules

Sensitivity of 87.7% with 0.25 FP/scan

Sensitivity of 90.5% with 0.5 FP/scan

Sensitivity of 93.1% with 1 FP/scan

Sensitivity of 96.2% with 2 FP/scan

Sensitivity of 97.9% with 4 FP/scan

Sensitivity of 99.1% with 8 FP/scan

Tang et al. [31] 2018 Alibaba’s 2017 TianChi Competition for Healthcare CADe 3D faster R-CNN Average sensitivity of 75.8% at seven predefined FPs/scan rates: 1/8, 1/4, 1/2, 1, 2, 4, 8. General nodules

3D DCNN

Alahmari et al. [35] 2018 NLST CADx Delta Radiomics Conventional Radiomic Diagnostic: Accuracy of 83.96% and AUC of 0.858 General nodules

Risk prediction: Accuracy of 78.3% and AUC of 0.822

Chen et al. [44] 2018 Private Data CADx Feature signature Sensitivity of 92.85% General nodules

Radiomics features Specificity of 72.73%

Accuracy of 84%

Dey et al. [39] 2018 LIDC-IDRI CADx 3D multi-output DenseNet Accuracy of 86.84% General nodules

AUC of 0.9548

Xie et al. [36] 2019 LIDC-IDRI CADx Pre-trained ResNet-50 Accuracy of 91.60% General nodules

AUC of 95.70%.

Liu et al. [15] 2018 (1) LIDC-IDRI CADx Multi-view CNN model (1) Accuracy of 92.3% Multi-scale nodules

(2) ELCAP (2) Accuracy of 90.3%

Oliveira et al. [45] 2018 LIDC-IDRI CADx 2D CNNs Accuracy of 88% General nodules

Liao et al. [21] 2019 DSB CADe A leaky noisy-OR gate Detection: Average recall of 0.8562 Nodules larger than 6 mm

CADx Modified U-net backbone network Predicted cancer probability task: AUC of 0.87

Paul et al. [37] 2018 NLST CADx ensemble of 5 models (3 CNNs, Radiomics and transfer learning model) Accuracy of 89.45% General nodules

AUC of 0.96

Kaya et al. [38] 2018 LIDC CADx Cascaded classification scheme Accuracy of 88.8% General nodules

Fusion of features Sensitivity of 88.41%

Deep features extracted from Alexnet Specificity of 94.12%

Hussein et al. [46] 2017 LIDC-IDRI Multi-Task 3D CNN-based MTL model Accuracy of 91.26% General nodules

Khosravan et al. [47] 2018 LUNA16 Multi-Task 3D deep multi-task CNN Segmentation: DSC of 91% General nodules

FP reduction: a score of nearly 92%

Wu et al. [48] 2018 LIDC-IDRI Multi-Task Similar 3D-UNet structure Nodule malignancy prediction: General nodules

Accuracy of 97.58%

Attributes prediction:

Accuracy of 89.33%

Nodule segmentation

Accuracy of 73.89%

Authors	Year	Database	Task	Technique	Experiential result	Nodule type
Gong et al. [23]	2018	(1) LUNA 16	CADe	3D tensor filter	(1) sensitivity of 79.3% with 4FP/scan	Spherical shape
		(2) ANODE09		19 selected optimal features	(2) sensitivity of 84.62% with 2.8 FP/scan
Gupta et al. [40]	2018	LIDC-IDRI	CADe	A rich set of 515 features with neural network classifier	Training set	Multi-size nodules
					Sensitivity of 85.6% with 8FP/scan
					Sensitivity of 83.5% with 1 FP/scan
					Testing set
					Sensitivity of 68.4% with 8FP/scan
GU et al. [40]	2019	154 thin-slice scans in LIDC	CADe	Vessel tree group Multi-scale dot filter	Sensitivity of 87.81% with 1.057 FP/scan	Juxta-vascular nodules
Zhang et al. [25]	2018	71 subjects of LIDC	CADe	3D skeletonization feature Ten features with SVM	Sensitivity of 89.3% with 2.1 FP/scan accuracy of 87.6%	Juxta-vascular nodules
Zhang et al. [22]	2018	LIDC-IDRI	CADe	Four CNN structures	Sensitivity of 80.06% with 4.7 FP/scan	Four size ranges of nodules
					Sensitivity of 94% with 15.1 FP/scan
Tan et al. [30]	2019	LIDC-IDRI	CADe	Four CNN models	Sensitivity of 88% with 1.9 FP/scan	Irregularly-shape juxta-pleural nodule
					Sensitivity of 94.1% with 4.01 FP/scan
Sakamoto et al. [41]	2018	LUNA2016	CADe	Cascaded CNN	Sensitivity of 94.4% with 4 FP/scan	General nodules
					Sensitivity of 95.9% with 8 FP/scan
Jin et al. [29]	2018	LUNA16	CADe	Deep 3D residual CNN	Sensitivity of 82.3% with 0.125 FP/scan	General nodules
					Sensitivity of 87.7% with 0.25 FP/scan
					Sensitivity of 90.5% with 0.5 FP/scan
					Sensitivity of 93.1% with 1 FP/scan
					Sensitivity of 96.2% with 2 FP/scan
					Sensitivity of 97.9% with 4 FP/scan
					Sensitivity of 99.1% with 8 FP/scan
Tang et al. [31]	2018	Alibaba’s 2017 TianChi Competition for Healthcare	CADe	3D faster R-CNN	Average sensitivity of 75.8% at seven predefined FPs/scan rates: 1/8, 1/4, 1/2, 1, 2, 4, 8.	General nodules
				3D DCNN
Alahmari et al. [35]	2018	NLST	CADx	Delta Radiomics Conventional Radiomic	Diagnostic: Accuracy of 83.96% and AUC of 0.858	General nodules
					Risk prediction: Accuracy of 78.3% and AUC of 0.822
Chen et al. [44]	2018	Private Data	CADx	Feature signature	Sensitivity of 92.85%	General nodules
				Radiomics features	Specificity of 72.73%
					Accuracy of 84%
Dey et al. [39]	2018	LIDC-IDRI	CADx	3D multi-output DenseNet	Accuracy of 86.84%	General nodules
					AUC of 0.9548
Xie et al. [36]	2019	LIDC-IDRI	CADx	Pre-trained ResNet-50	Accuracy of 91.60%	General nodules
					AUC of 95.70%.
Liu et al. [15]	2018	(1) LIDC-IDRI	CADx	Multi-view CNN model	(1) Accuracy of 92.3%	Multi-scale nodules
		(2) ELCAP			(2) Accuracy of 90.3%
Oliveira et al. [45]	2018	LIDC-IDRI	CADx	2D CNNs	Accuracy of 88%	General nodules
Liao et al. [21]	2019	DSB	CADe	A leaky noisy-OR gate	Detection: Average recall of 0.8562	Nodules larger than 6 mm
			CADx	Modified U-net backbone network	Predicted cancer probability task: AUC of 0.87
Paul et al. [37]	2018	NLST	CADx	ensemble of 5 models (3 CNNs, Radiomics and transfer learning model)	Accuracy of 89.45%	General nodules
					AUC of 0.96
Kaya et al. [38]	2018	LIDC	CADx	Cascaded classification scheme	Accuracy of 88.8%	General nodules
				Fusion of features	Sensitivity of 88.41%
				Deep features extracted from Alexnet	Specificity of 94.12%
Hussein et al. [46]	2017	LIDC-IDRI	Multi-Task	3D CNN-based MTL model	Accuracy of 91.26%	General nodules
Khosravan et al. [47]	2018	LUNA16	Multi-Task	3D deep multi-task CNN	Segmentation: DSC of 91%	General nodules
					FP reduction: a score of nearly 92%
Wu et al. [48]	2018	LIDC-IDRI	Multi-Task	Similar 3D-UNet structure	Nodule malignancy prediction:	General nodules
					Accuracy of 97.58%
					Attributes prediction:
					Accuracy of 89.33%
					Nodule segmentation
					Accuracy of 73.89%

2.1 Computer aided detection system

Latest studies of CADe systems are summarized respectively, and common performance index of CADe is sensitivity with FP (False Positives)/ scan.

Juxta-vascular nodules, especially small juxta-vascular nodules attached to tiny vessels, are harder to be detected than isolated nodules. In morphological appearance, vessels are connected tubular structures, while common pulmonary nodules are approximate spheres or ellipsoids. Skeletonization is one of the most widely used morphological operations in image processing, and often utilized to identify vessels and bronchioles in medical applications. Zhang et al. [25] proposed 3D skeletonization feature based CADe to efficiently classify juxta-vascular nodules and vessels. The proposed CADe system has extracted ten features, including one 3D skeletonization feature and nine traditional 3D features (four gray features and five shape features), to train SVM for false positive reduction. On 71 subjects of LIDC database, the average sensitivity was 89.3% with 2.1 FP/scan, and accuracy was 87.6%. GU et al. [40] firstly constructed vessel tree group by grouping and labeling algorithm, and then extracted nodules candidates in non-vessel tree group and vessel tree group separately, lastly multi-scale dot filter had been applied to cut away juxta-vascular nodules from the vessel tree group. On 154 thin-slice scans with 204 nodules in the LIDC database, the sensitivity of this CADe system was 87.81% with 1.057 FP/scan.

In order to detect different sizes of pulmonary nodules, Gupta et al. [24] have proposed three discrete modules, each of which consisted of a multistage rule-based thresholding and morphological operations, to detect three size ranges of nodules as followings: small candidates (3 mm≤diametric size <6 mm) with eccentricity and compactness rules; medium candidates (6 mm≤diametric size <10 mm) with compactness and elongation rules; large candidates (diametric size ≥10 mm) with circularity rules. The proposed CADe system utilized a rich set of 515 features based on cluster, texture, and voxel-based intensity features to train a neural network classifier. For the LIDC-IDRI training set, sensitivity was 85.6% with 8FP/scan and 83.5% with 1 FP/scan; for test sets, sensitivity was 68.4% with 8FP/scan. Zhang et al. [22] proposed an effective CADe system on multi-group patches cut from the lung images. Four CNN models have been trained with four sizes of sampling patches, including 32×32, 28×28, 24×24 and 20×20 pixels. MC (Multi-Crop) pooling operation, consisted of max-pooling operation for the whole patch and center-pooling operation for the center part of patch, can make the CNN models learn the overall and internal structure of nodules at the same time. On LIDC-IDRI database, the sensitivity was 80.06% with 4.7 FP/scan, and 94% with 15.1 FP/scan.

Different types of pulmonary nodules have large variations in size, shape and density, and thus rule-based schemes can only apply to certain types of nodules. Deep learning based schemes, which learn from large scale of training data to adjust parameters of the hierarchical network structure, are general detection schemes, e.g. 3D faster R-CNN [31]. Tan et al. [30] have designed expert knowledge-infused CNN models for false positive reduction. Specifically, inside, outside, HOG and LBP patches extracted from one candidate patch were considered as internal structure, outer environment, shape and local texture knowledge. Then four CNN models have been trained with those types of patches respectively. The statistical texture features, Haralick features, had also been combined with CNN models to capture global texture knowledge. The proposed CADe system had good robustness for irregularly-shaped juxta-pleural nodules. On LIDC-IDRI database, the sensitivity was 88% with 1.9 FP/scan and 94.1% with 4.01 FP/scan. Jin et al. [29] have trained a deep 3D residual CNN, which was much deeper than the traditional 3D CNNs, to reduce false positives. In the network, a SPC (Spatial Pooling and Cropping) layer was designed to extract the overall and internal contextual information of candidate patches. In the training process, online hard sample selection strategy was employed to make the model better fit nodules which were more difficult to identify. On 888 CT scans of the LUNA16 database, the sensitivity was 82.3% with 0.125 FP/scan, 87.7% with 0.25 FP/scan, 90.5% with 0.5 FP/scan, 93.1% with 1 FP/scan, 96.2% with 2 FP/scan, 97.9% with 4 FP/scan, and 99.1% with 8 FP/scan. DCNN (Deep CNN) based CADe systems show better performance than previous systems. Tang et al. [31] proposed a novel DCNN approach which applied a U-Net-inspired 3D faster R-CNN for suspicious nodules detection and 3D DCNN for false positive reduction. The proposed system has ranked first of 2887 teams in Season one of alibaba’s 2017 TianChi Competition for Healthcare.

Pulmonary nodule classification is a class imbalanced problem, as true nodules are much less than other structures. Classifiers tend to be overwhelmed by majority class and ignore the minority class. Sakamoto et al. [41] have applied cascaded CNN to overcome the imbalanced problem. Specifically, previous m-1 CNN models had been performed as single-sided classifiers to filter out obvious non-nodule structures. By gradually reducing the quantitative difference between true nodules and others, the last m-th CNN model can classify nodules with relatively balanced data. The proposed CADe can achieve few false positives while maintaining high sensitivity. On LUNA2016 database, the sensitivity was 94.4% with 4FP/scan and 95.9% with 8FP/scan.

Discriminative features can be automatically learned from image data is the major advantage of CNN. CNN based methods have shown better performance than traditional methods. As shown in Table 2, with much deeper 3D residual CNN, CADe system proposed by Jin et al. [29] has achieved higher sensitivity than others. With the development of more advanced CNN models, the performance of CADe system can be further improved.

Table 2
Comparison of detection sensitivity of several reviewed CADe systems

Method\FP 1 2 4 8 15

Zhang et al. [25] – 89.3% – – –

GU et al. [40] 87.8% – – –

Gupta et al. [24] – – – 68.4%

Zhang et al. [22] – – 80.1% – 94%

Tan et al. [30] – 88% 94.1% – –

Jin et al. [29] 93.1% 96.2% 97.9% 99.1% –

Sakamoto et al. [41] – – 94.4% 95.9% –

Method\FP	1	2	4	8	15
Zhang et al. [25]	–	89.3%	–	–	–
GU et al. [40]	87.8%	–	–	–
Gupta et al. [24]	–	–	–	68.4%
Zhang et al. [22]	–	–	80.1%	–	94%
Tan et al. [30]	–	88%	94.1%	–	–
Jin et al. [29]	93.1%	96.2%	97.9%	99.1%	–
Sakamoto et al. [41]	–	–	94.4%	95.9%	–

2.2 Computer aided diagnosis system

Currently Radiomics, deep learning schema and feature fusion of those two are research hotspots in CADx systems, and common performance indexes of CADx systems are accuracy and AUC (Area Under the receiver operating characteristic Curve).

2.2.1 Radiomics based CADx

Radiomics is the process of extracting and analyzing image-based quantitative features from a region of interest, including size, shape, volume and texture features. Delta Radiomics reflects the changes in features over time and has utility in predicting treatment response for various cancers. Alahmari et al. [35] have generated delta Radiomics from baseline and follow-up screening intervals to predict risk of cancer for indeterminate pulmonary nodules. The proposed CADx system had utilized three conventional Radiomic features, e.g. 219 Definiens features [42], 23 Rider stable features [43] and 94 Pyradiomics features (http://www.pyradiomics.readthedocs.io/en/latest/features.html), incorporating delta features computed by calculating the difference for those features from two serial screening intervals. On NLST database, the accuracy for the diagnostic experiment was 83.96% and AUC was 0.858, and in the risk prediction experiment the accuracy was 78.3% and AUC was 0.822.

Chen et al. [44] have selected the best 4 features as a feature signature from 750 Radiomics features. The significantly different features (p < 0.05) between benign and malignant lesions were selected through method of nonparametric Wilcoxon rank sum test. The sequential forward selection (SFS) was then applied to further evaluate the correlation between the features. SVM was choosed as the classifier, and the leave-one out cross-validation method was applied to get the prediction accuracy for each feature. The most accurate one was selected and followed by the next most accurate feature. Since the inclusion of more than 4 features no longer improved the performance of the classifier, only top 4 features were selected as the signature. The accuracy using the signature in benign or malignant classification was 84% with the sensitivity of 92.85% and the specificity of 72.73%. Radiomics based CADx systems have shown poorer performance than deep learning based systems.

2.2.2 Deep learning based CADx

Both 2D CNN and 3D CNN are commonly used deep learning models in CADx systems. 2D CNN has lower computational complexity, while 3D CNN can better analyze the spatial structure of pulmonary nodules.

Decomposing 3D object into multiple 2D views is a straightforward way to extend the use of 2D CNN to the analysis of volumetric medical images. Xie et al. [36] have proposed MV-KBC algorithm which extracted 2D nodule slices from nine views of planes and then extracted the OA, HVV and HS patches on each slice. OA, HVV and HS patches were extracted from a square ROI that encapsulated the nodules, representing the nodule’s overall appearance, heterogeneity in voxel values and heterogeneity in shapes. Three pre-trained ResNet-50 networks formed a KBC sub-model, and the proposed MV-KBC model consisted of nine KBC sub-models. Fusing sub-models at the decision level enabled the entire model to be trained in an end-to-end manner. On the LIDC-IDRI dataset, the proposed system obtained accuracy of 91.60% and AUC of 95.70%. Instead of on fixing nine views of planes, Liu et al. [15] proposed a CADx system for nodule type classification on the selected planes which were more abundant in information. The system used icosahedra to approximate the spherical surface of nodules, intensity analysis to achieve estimated radius for each nodule, and then sorted circular planes by a high frequency content measure analysis. Specifically, a multi-view CNN model was trained with nodules captured at 3 scales and 4 sorted views. On LIDC-IDRI database, this CADx system had obtained accuracy of 92.3%, and on ELCAP database the accuracy of 90.3%. Oliveira et al. [45] sorted a random vector and random angle to avoid bias to any preferential direction. Trigonometric functions had been used to deliver comprehensive bi-dimensional data representation out of 3D volumetric patches. 2D CNNs trained with this representation had achieved the accuracy of 88% on the LIDC-IDRI database, and that was similar performance with 3D CNNs, but using 16 times less data and running 4 times faster.

The basic 3D CNN would get stuck in a local optimum. To find a better optimum, Dey et al. [39] have introduced early outputs for 3D CNN, which provided immediate feedback from an early evaluation of error functions. DenseNet can strengthen the vanishing gradient by the connections to layers closer to output. The proposed network had adopted the design of 3D DenseNet, and provided early outputs after every pooling layer that followed the dense blocks in both pathways. The feature maps before each intermediate outputs were merged along with the features from the last convolutional layers of the two pathways and then sent to the classifier for the final output. For the lung cancer diagnosis problem, the proposed network obtained the accuracy of 86.84% and AUC of 0.9548 on the LIDC-IDRI dataset. Liao et al. [21] proposed an integrated CAD system which firstly detected all suspicious lesions and then evaluated the lung malignancy. Specifically, the CAD system consisted of two modules. The first one was a 3D region proposal network which outputted all suspicious nodules for a subject. The second one selected the top five nodules based on the detection confidence, evaluated their cancer probabilities, and combined them with a leaky noisy-OR gate to obtain the probability of lung cancer for each subject. The two modules shared the same modified U-net backbone network. The model had been trained end-to-end to achieve efficient optimization. On the validation set of DSB (Data Science Bowl) 2017 (https://www.kaggle.com/c/data-science-bowl-2017), the average recall was 0.8562 for nodule detection task, and for predicted cancer probability task the AUC was 0.87.

2.2.3 Feature fusion based CADx

A fusion of Radiomics and deep learning schema is also a research hotspot in CADx systems. Paul et al. [37] have investigated the performance of an ensemble of classifiers using different feature sets and learning approaches to predict nodule malignancy. From an ensemble of 5 models (3 CNNs, Radiomics and transfer learning model) a best accuracy of 89.45% was obtained using voting, whereas the best AUC of 0.96 was obtained by combining via averaging on the NLST dataset. The imaging characteristics of pulmonary nodules, such as texture, lobulation and calcification, are important diagnostic bases in medical practice. Kaya et al. [38] proposed cascaded classification scheme to use information on nodule characteristics for lung cancer diagnosis. Specifically, at the first level individual classifiers were trained to learn characteristics, and then the first level classifiers were combined for malignancy classification at the second level. The proposed CADx system applied the fusion of 115 image features and deep features extracted from Alexnet, and obtained accuracy of 88.8%, sensitivity of 88.41%, and specificity of 94.12% on LIDC database.

As shown in Fig. 2, the top three CADx systems are based on 3D CNN and multi-view 2D CNN models. CADx systems based on deep learning models have reported the accuracy of higher than 90% [15, 36] and been the mainstream of current research. It is particularly worth mentioning that 3D Deep Leaky Noisy-OR Network proposed by Liao et al. [21] won the first place in the Data Science Bowl 2017 competition.

Fig.2

Accuracy and AUC of the selected CADx systems.

2.3 Multi-task learning system

MTL (Multi-task learning) performs joint learning of relevant tasks while exploiting dependencies in feature space aiming to improve regressing one task using the others. Hussein et al. [46] proposed 3D CNN-based MTL model to extract feature representation for six high-level nodule attributes. Then the features from different CNNs were fused together using graph regularized sparse least square optimization function to obtain coefficient vectors corresponding to each task. During the testing phase, malignancy scores have been computed by multiplying the feature representation of the testing images with the coefficient vectors. On the LIDC-IDRI database, the accuracy was 91.26%.

Guidelines for early diagnosis and treatment of lung cancer need accurate measurements of abnormalities, while accurate segmentation is the premise of accurate measurement. Khosravan et al. [47] had built a 3D deep multi-task CNN to tackle false positive reduction and nodule segmentation jointly. On LUNA16 dataset the proposed system had achieved an average dice similarity coefficient (DSC) of 91% as segmentation accuracy and a score of nearly 92% for FP reduction. Wu et al. [48] proposed end to end multi-task learning CNN to joint learn for pulmonary nodule segmentation, attributes and malignancy prediction. The combination of related tasks was helpful to improve the performance of each single task. Specifically, the proposed CNN had a similar 3D-UNet structure. Dice coefficient loss was chosen for segmentation, categorical cross-entropy loss was exploited for attribute learning and malignancy prediction. A fixed trade-off parameter between classification and segmentation was applied during training. On the LIDC-IDRI database the proposed model had achieved accuracy of 97.58% for nodule malignancy prediction, 89.33% for attributes prediction and 73.89% for nodule segmentation.

2.4 Commercialized CAD products

At present, several companies have committed to the development of clinical CAD systems abroad including IBM Watson and Google Verily. For example, ClearRead CT™ developed by Riverain Technologies is an US FDA approved CAD system, which aims to assist radiologists to detect pulmonary nodules or early signs of lung cancer. Clinical research [32] had shown that with the help of this system radiologists’ interpretation time was significantly decreased by 26%; sensitivity was increased by 11%, while specificity was decreased from 89.9% to 84.4%. Post-processing platform on Philips IntelliSpace Portal is also approved by US FDA and CFDA of China, and the lung disease diagnosis and treatment module can assist to assess the risk of cancer, grade the disease and recommend the treatment plan.

Currently, many new companies or groups within the established companies have also been established in China, e.g. Deepwise, YITU Technology, 12 SIGMA Technologies, Infervision, United Imaging Healthcare and Tencent MIYING, which apply artificial intelligent (AI) technology to develop new CADe and CADx systems of lung nodules using CT images.

3 Discussion and future work

Features based on shape, cluster, texture, gray and voxel-based intensity are common features for CADe systems [24, 25]. For the detection of juxta-vascular nodules, 3D skeletonization feature [25], multi-scale dot filter and vessel tree group [40] have been utilized to identify vessels and bronchioles. Traditional classifiers, such as SVM, random forest and neural network, have been trained for false positive reduction [23 , 40]. Deep learning based CADe systems can learn discriminative features automatically and achieve better performance than feature extraction based systems [22 , 41]. Deep 3D residual CNN [29], which has deeper network structure, has shown better performance than shallow CNN to reduce false positives. The sensitivity of deep learning based CADe systems reported on LIDC-IDRI is between 80.06% and 94.1% with an average 4.3 FP/scan, and on LUNA16 is between 94.4% and 97.9% with an average 4 FP/scan.

Radiomics based CADx systems [35, 42] have shown poorer performance than deep learning based CADx systems. Feature fusion based CADx systems [37, 38] have reported similar performance to deep learning based CADx systems. Through decomposing 3D pulmonary nodules into multiple 2D views, 2D CNN based CADx systems [15 , 43] can achieve as good performance as 3D CNN CADx systems, such as 3D DenseNet [39]. 3D U-Net shape DCNN is an effective CNN structure for both CADe [31] and CADx systems [21, 46]. The accuracy of deep learning based CADx systems reported on LIDC-IDRI is between 86.84% and 92.3% with an average AUC of 0.956.

However, the performances of current CAD systems have not yet met the rigid requirements of the clinical applications. High sensitivity with a low false positive rate, strong robustness and high efficiency are the goals of the CADe systems. Besides benign-malignant classification, identifying specific types, accurate measurement and assessment of characteristics for pulmonary nodules are also the auxiliary functions needed in clinic. For instance, the size of irregular pulmonary nodules is poorly represented by diameter [49] and needs to be measured better. Therefore, fine-grained analysis and high-precision classification of pulmonary nodules are the goals of CADx systems.

The development of CAD systems needs to be closely aligned with clinical requirements. However, the following clinical scenarios have not been fully considered in current CAD systems and should be the future directions of development:

Follow-up screening. At present, CAD mainly focuses on the first screening of patients, while clinically lots of patients need follow-up screening after the baseline screening. Lung-RADS (Lung Imaging Reporting and Data System) has classified pulmonary nodules based on CT findings, principles of treatment, probability of malignancy and incidence rate of disease in expected population, so that different types of pulmonary nodules can have a relatively clear judgment and different follow-up screening programs. Growth rate obtained by followed-up screening is the strongest predictor for lung cancer [50]. Semantic features defined by border definition and contour have performed similar to lung-RADS at follow-up time point and outperformed lung-RADS at baseline, and thus these semantics alongside of lung-RADS can improve the performance to detect malignancy [51]. On follow-up screening, certain type of pulmonary nodules can be followed to assess the risk of developing cancer, for instance perifissural nodules were less likely to develop cancer [52], while subsoild nodules showed a high risk [53]. In the diagnosis of subsoild nodule, accurate segmentation of the solid core and enclosed vessels [54] on follow-up CT scan is benefit for increasing the specificity for adenocarcinoma invasiveness [55]. Follow-up screening can provide additional basises for detection and diagnosis, thus need to be better analyzed by CAD in the future.

Multi-mode scan. Multi-mode scan can help to overcome the limitations of CT scan. The limitations of CT scan include radioactivity, respiratory disturbance, disutility for SCLC (small cell lung cancer), etc. Reducing the radiation dose will lead to poor image quality, thus techniques, such as hybrid type iterative reconstruction technique [58], have been proposed to make a positive effect on nodule detection capability of the CADe system even with reduced radiation dose. Recently proposed DIBH PET/CT technique can reduce the blurring effect of respiratory motion to enhance the diagnostic accuracy for pulmonary nodules [59]. Thoracic MRI, which provides not only morphological, but also functional, physiological, pathophysiological, and molecular information, now plays a complementary role for the detection of pulmonary nodules [60]. Due to the fast growth rate, early detection with LDCT imaging had no impact on SCLC outcomes [61], thus ideal screening modality should detect SCLC earlier than when it can be detected on LDCT scans. Thus, multi-mode CAD should also be a valuable research direction.

Co-analysis of text and image. CAD for co-analysis of text and image is another worth studying direction. The current CAD mainly focuses on image data. On CT scan, automatic interpretation of lesions and automatic generation of medical image reports are useful clinical functions and worth to be well developed in the further.

Biomarks. Lung cancer screening can be improved in three ways: (1) by refining selection criteria (risk factor assessment) [56], (2) by using CAD to make it easier to interpret chest CT images, and (3) by using biomarker for early cancer detection [7]. Biomarker is a hot research direction for early detection of lung cancer, e.g. tumor-derived autoantibodies could effectively be served as blood biomarkers for benign-malignant pulmonary nodule diagnosis with specificity of 70% [57]. CADx combined with effective biomarks could increase the accuracy and should be a future research direction.

4 Conclusion

In this paper we have introduced a brief structure of CAD system and provided a detailed review of latest researches for CADe, CADx, multi-task CAD and commercial CAD systems. Deep learning based CAD system is the mainstream of current research, and has achieved better performance than feature extraction based CAD system. For widespread clinical usage, CAD system needs to meet the rigid requirements, such as high sensitivity with a low false positive rate, strong robustness, high efficiency, fine-grained analysis and high-precision classification. CAD systems for follow-up screening, multi-mode scan, co-analysis of text and image, and combining with effective biomarks are worth further studying. As technology advances, it is expected that the utilization of CAD system would improve the performance of lung cancer screening in future clinical practice.

Footnotes

Acknowledgments

This work was supported by grants from funding National Natural Science Foundation of China (No. 61672362, No. U1611263); Beijing Natural Science Foundation (No. 4172012, No. 7192042); Scientific Research Common Program of Beijing Municipal Commission of Education (No. KM201710025011).

References

Ferlay

, Colombet

, Soerjomataram

, et al., Estimating the global cancer incidence and mortality in 2018: Globocan sources and methods, International Journal of Cancer 144 (2019), 1941–1953.

Zhou

, Fan

, Wang

, et al., China national guideline of classification, diagnosis and treatment for lung nodules (2016 version), Chin J Lung Cancer 19 (2016), 793–798.

Zhang

, Jiang

, Yang

, et al., Automatic nodule detection for lung cancer in CT images: A review, Computers in Biology and Medicine 103 (2018), 287–300.

Yamada

, Shiomi

, Hashimoto

, et al., Value of a computer-aided detection system based on chest tomosynthesis imaging for the detection of pulmonary nodules, Radiology 287 (2018), 333–339.

Huang

, Park

, Yan

, et al., Added value of computer-aided CT image features for early lung cancer diagnosis with small pulmonary nodules: A matched case-control study, Radiology 286 (2018), 286–295.

Vassallo

, Traverso

, Agnello

, et al., A cloud-based computer-aided detection system improves identification of lung nodules on computed tomography scans of patients with extra-thoracic malignancies, European Radiology 29 (2018), 144–152.

Benzaquen

, Boutros

, Marquette

, et al., Lung cancer screening, towards a multidimensional approach: Why and how? Cancers 11 (2019), 212.

Armato

S.G.

, McLennan

, Bidaut

, et al., The lung image database consortium (LIDC) and image database resource initiative (IDRI): A completed reference database of lung nodules on CT scans, Medical Physics 38 (2011), 915–931.

Wang

, Luo

, Yang

, Lin

, Data analysis of the lung imaging database consortium and image database resource initiative, Academic Radiology 22 (2015), 488–495.

10.

Setio

, Traverso

, de Bel

, et al., Validation, comparison, and combination of algorithms for automatic detection of pulmonary nodules in computed tomography images: The LUNA16 challenge, Medical Image Analysis 42 (2017), 1–13.

11.

Oudkerk

, Heuvelmans

, Screening for lung cancer by imaging: The NELSON study, Journal of the Belgian Society of Radiology 96 (2013), 163.

12.

Yip

, Henschke

C.I.

, Yankelevitz

D.F.

, Smith

J.P.

, CT screening for lung cancer: Alternative definitions of positive test result based on the national lung screening trial and international early lung cancer action program databases, Radiology 273 (2014), 591–596.

13.

van Ginneken

, Armato

S.G.

, de Hoop

, et al., Comparing and combining algorithms for computer-aided detection of pulmonary nodules in computed tomography scans: The ANODE09 study, Medical Image Analysis 14 (2010), 707–722.

14.

National lung screening trial research team, Reduced lung-cancer mortality with low-dose computed tomographic screening, New England Journal of Medicine 365 (2011), 395–409.

15.

Liu

, Hou

, Qin

, Hao

, Multi-view multi-scale CNNs for lung nodule type classification from CT images, Pattern Recognition 77 (2018), 262–275.

16.

Shin

H.C.

, Roth

H.R.

, Gao

, et al., Deep convolutional neural networks for computer-aided detection: CNN architectures, dataset characteristics and transfer learning, IEEE Transactions on Medical Imaging 35 (2016), 1285–1298.

17.

Papandreou

, Chen

L.C.

, Murphy

K.P.

, Yuille

A.L.

, Weakly-and semi-supervised learning of a deep convolutional network for semantic image segmentation, in 2015 IEEE International Conference on Computer Vision (ICCV), 2015.

18.

Firmino

, Angelo

, Morais

, et al., Computer-aided detection (CADe) and diagnosis (CADx) system for lung cancer with likelihood of malignancy, BioMedical Engineering OnLine 15 (2016).

19.

Javaid

, Javid

, Rehman

M.Z.U.

, Shah

S.I.A.

, A novel approach to CAD system for the detection of lung nodules in CT images, Computer Methods and Programs in Biomedicine 135 (2016), 125–139.

20.

Abbas

, Segmentation of differential structures on computed tomography images for diagnosis lung-related diseases, Biomedical Signal Processing and Control 33 (2017), 325–334.

21.

Liao

, Liang

, Li

, Hu

, Song

, Evaluate the malignancy of pulmonary nodules using the 3-d deep leaky noisy-or network, IEEE Transactions on Neural Networks and Learning Systems (2019), 1–12.

22.

Jiang

, Ma

, Qian

, et al., An automatic detection system of lung nodule based on multigroup patch-based deep learning network, IEEE Journal of Biomedical and Health Informatics 22 (2018), 1227–1237.

23.

Gong

, Liu

, Wang

, et al., Automatic detection of pulmonary nodules in CT images by incorporating 3d tensor filtering with local image feature analysis, Physica Medica 46 (2018), 124–133.

24.

Gupta

, Saar

, Martens

, Moullec

Y.L.

, Automatic detection of multisize pulmonary nodules in CT images: Large-scale validation of the false-positive reduction step, Medical Physics 45 (2018), 1135–1149.

25.

Zhang

, Wang

, Li

, Chen

, 3d skeletonization feature based computer-aided detection system for pulmonary nodules in CT datasets, Computers in Biology and Medicine 92 (2018), 64–72.

26.

Filho

P.R.

, Cortez

P.C.

, da Silva Barros

A.C.

, et al., Novel and powerful 3d adaptive crisp active contour method applied in the segmentation of CT lung images, Medical Image Analysis 35 (2017), 503–516.

27.

Shaukat

, Raja

, Gooya

, Frangi

A.F.

, Fully automatic detection of lung nodules in CT images using a hybrid feature set, Medical Physics 44 (2017), 3615–3629.

28.

Gong

, Liu

, Wang

, et al., Computer-aided detection of pulmonary nodules using dynamic self-adaptive template matching and a FLDA classifier, Physica Medica 32 (2016), 1502–1509.

29.

Jin

, Li

, Tong

, Lin

, A deep 3d residual CNN for false-positive reduction in pulmonary nodule detection, Medical Physics 45 (2018), 2097–2107.

30.

Tan

, Huo

, Liang

, Li

, Expert knowledge-infused deep learning for automatic lung nodule detection, Journal of X-Ray Science and Technology 27 (2019), 17–35.

31.

Tang

, Kim

D.R.

, Xie

, Automated pulmonary nodule detection using 3d deep convolutional neural networks, in 2018 IEEE 15th International Symposium on Biomedical Imaging (ISBI 2018), 2018.

32.

S.B.

, Freedman

M.T.

, Gillis

L.B.

, et al., Journal club: Computer-aided detection of lung nodules on CT with a computerized pulmonary vessel suppressed function, American Journal of Roentgenology 210 (2018), 480–488.

33.

Ali

, Hart

G.R.

, Gunabushanam

, et al., Lung nodule detection via deep reinforcement learning, Frontiers in Oncology 8 (2018), 108.

34.

Thawani

, McLane

, Beig

, et al., Radiomics and radiogenomics in lung cancer: A review for the clinician, Lung Cancer 115 (2018), 34–41.

35.

Alahmari

S.S.

, Cherezov

, Goldgof

D.B.

, et al., Delta radiomics improves pulmonary nodule malignancy prediction in lung cancer screening, IEEE Access 6 (2018), 77796–77806.

36.

Xie

, Xia

, Zhang

, et al., Knowledge-based collaborative deep learning for benign-malignant lung nodule classification on chest CT, IEEE Transactions on Medical Imaging 38 (2019), 991–1004.

37.

Paul

, Hall

, Goldgof

, Schabath

, Gillies

, Predicting nodule malignancy using a CNN ensemble approach, in 2018 International Joint Conference on Neural Networks (IJCNN) IEEE, 2018.

38.

Kaya

, Cascaded classifiers and stacking methods for classification of pulmonary nodule characteristics, Computer Methods and Programs in Biomedicine 166 (2018), 77–89.

39.

Dey

, Lu

, Hong

, Diagnostic classification of lung nodules using 3d neural networks, in 2018 IEEE 15th International Symposium on Biomedical Imaging (ISBI 2018), 2018.

40.

, Lu

, Zhang

, et al., Automatic lung nodule detection using multi-scale dot nodule-enhancement filter and weighted support vector machines in chest computed tomography, Plos One, 14 (2019), e0210551.

41.

Sakamoto

, Nakano

, Zhao

, Sekiyama

, Lung nodule classification by the combination of fusion classifier and cascaded convolutional neural networks, in 2018 IEEE 15th International Symposium on Biomedical Imaging (ISBI 2018), 2018.

42.

Yoganand

, Virendra

, Yuhua

, et al., Test-retest reproducibility analysis of lung CT image features, Journal of Digital Imaging 27 (2014), 805–823.

43.

Balagurunathan

, Gu

, Wang

, et al., Reproducibility and prognosis of quantitative features extracted from CT images, Translational Oncology 7 (2014), 72–87.

44.

Chen

C.H.

, Chang

C.K.

, Tu

C.Y.

, et al., Radiomic features analysis in computed tomography images of lung nodule classification. PLos One 13 (2018), e0192002.

45.

Oliveira

D.A.B.

, Viana

M.P.

, An efficient multi-scale data representation method for lung nodule false positive reduction using convolutional neural networks, in 2018 IEEE 15th International Symposium on Biomedical Imaging (ISBI 2018), 2018.

46.

Hussein

, Cao

, Song

, Bagci

, Risk stratification of lung nodules using 3d CNN-based multi-task learning, in Lecture Notes in Computer Science, Springer International Publishing, 2017, pp. 249–260.

47.

Khosravan

, Bagci

, Semi-supervised multi-task learning for lung cancer diagnosis, in 2018 40th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), 2018.

48.

, Zhou

, Wang

, Joint learning for pulmonary nodule segmentation, attributes and malignancy prediction, in 2018 IEEE 15th International Symposium on Biomedical Imaging (ISBI 2018), 2018.

49.

Heuvelmans

M.A.

, Walter

J.E.

, Vliegenthart

, et al., Disagreement of diameter and volume measurements for pulmonary nodule size estimation in CT lung cancer screening, Thorax 73 (2017), 779–781.

50.

Walter

J.E.

, Heuvelmans

M.A.

, de Bock

G.H.

, et al., Characteristics of new solid nodules detected in incidence screening rounds of low-dose CT lung cancer screening: The NELSON study, Thorax 73 (2018), 741–747.

51.

, Balagurunathan

, Liu

, et al., Comparison between radiological semantic features and lung-RADS in predicting malignancy of screen-detected lung nodules in the national lung screening trial, Clinical Lung Cancer 19 (2018), 148–156.

52.

Mets

O.M.

, Chung

, Scholten

E.T.

, et al., Incidental perifissural nodules on routine chest computed tomography: Lung cancer or not? European Radiology 28 (2017), 1095–1101.

53.

Silva

, Prokop

, Jacobs

, et al., Long-term active surveillance of screening detected subsolid nodules is a safe strategy to reduce overtreatment, Journal of Thoracic Oncology 13 (2018), 1454–1463.

54.

Charbonnier

J.P.

, Chung

, Scholten

E.T.

, et al., Automatic segmentation of the solid core and enclosed vessels in subsolid pulmonary nodules,”. Scientific Reports 8 (2018), 646.

55.

Revel

M.P.

, Mannes

, Benzakoun

, et al., “Subsolid lung nodule classification: A CT criterion for improving interobserver agreement,”. Radiology 286 (2018), 316–325.

56.

Jacobson

F.L.

, Jaklitsch

M.T.

, Computed tomography scanning for early detection of lung cancer, Annual Review of Medicine 69 (2018), 235–245.

57.

Lastwika

K.J.

, Kargl

, Zhang

, et al., Tumor-derived autoantibodies identify malignant pulmonary nodules, American Journal of Respiratory and Critical Care Medicine 199 (2019), 1257–1266.

58.

Ohno

, Aoyagi

, Chen

, et al., Comparison of computer-aided detection (CADe) capability for pulmonary nodules among standard-, reduced- and ultra-low-dose CTs with and without hybrid type iterative reconstruction technique, European Journal of Radiology 100 (2018), 49–57.

59.

Balamoutoff

, Serrano

, Hugonnet

, et al., Added value of a single fast 20-second deep-inspiration breath-hold acquisition in FDG PET/CT in the assessment of lung nodules, Radiology 286 (2018), 260–270.

60.

, HU

, H

, JB

and van Beek

EJR

, MRI for solitary pulmonary nodule and mass assessment: Current state of the art, J Magn Reson Imaging 47 (2018), 1437–1458.

61.

Thomas

, Pattanayak

, Szabo

, Pinsky

, Characteristics and outcomes of small cell lung cancer detected by CT screening, Chest 154 (2018), 1284–1290.