Computer-aided detection of pulmonary nodules in computed tomography: analysis and review of the literature

Lung Nodule Detection using Convolutional Neural Networks with Transfer Learning on CT Images

AIM AND OBJECTIVE

Lung nodule detection is critical in improving the five-year survival rate and reducing mortality for patients with lung cancer. Numerous methods based on Convolutional Neural Networks (CNNs) have been proposed for lung nodule detection in Computed Tomography (CT) images. With the collaborative development of computer hardware technology, the detection accuracy and efficiency can still be improved.

MATERIALS AND METHODS

In this study, an automatic lung nodule detection method using CNNs with transfer learning is presented. We first compared three of the state-of-the-art convolutional neural network (CNN) models, namely, VGG16, VGG19 and ResNet50, to determine the most suitable model for lung nodule detection. We then utilized two different training strategies, namely, freezing layers and fine-tuning, to illustrate the effectiveness of transfer learning. Furthermore, the hyper-parameters of the CNN model such as optimizer, batch size and epoch were optimized.

RESULTS

Evaluated on the Lung Nodule Analysis 2016 (LUNA16) challenge, promising results with an accuracy of 96.86%, a precision of 91.10%, a sensitivity of 90.78%, a specificity of 98.13%, and an AUC of 99.37% were achieved.

CONCLUSION

Compared with other works, state-of-the-art specificity is obtained, which demonstrates that the proposed method is effective and applicable to lung nodule detection.

The Added Value of Computer-aided Detection of Small Pulmonary Nodules and Missed Lung Cancers

Lung cancer at its earliest stage is typically manifested on computed tomography as a pulmonary nodule, which could be detected by low-dose multidetector computed tomography technology and the use of thinner collimation. Within the last 2 decades, computer-aided detection (CAD) of pulmonary nodules has been developed to meet the increasing demand for lung cancer screening computed tomography with a larger set of images per scan. This review introduced the basic techniques and then summarized the up-to-date applications of CAD systems in clinical and research programs and in the low-dose lung cancer screening trials, especially in the detection of small pulmonary nodules and missed lung cancers. Many studies have already shown that the CAD systems could increase the sensitivity and reduce the false-positive rate in the diagnosis of pulmonary nodules, especially for the small and isolated nodules. Further improvements to the current CAD schemes are needed to detect nodules accurately, particularly for subsolid nodules.

Lung Nodule Detection via Deep Reinforcement Learning

Lung cancer is the most common cause of cancer-related death globally. As a preventive measure, the United States Preventive Services Task Force (USPSTF) recommends annual screening of high risk individuals with low-dose computed tomography (CT). The resulting volume of CT scans from millions of people will pose a significant challenge for radiologists to interpret. To fill this gap, computer-aided detection (CAD) algorithms may prove to be the most promising solution. A crucial first step in the analysis of lung cancer screening results using CAD is the detection of pulmonary nodules, which may represent early-stage lung cancer. The objective of this work is to develop and validate a reinforcement learning model based on deep artificial neural networks for early detection of lung nodules in thoracic CT images. Inspired by the AlphaGo system, our deep learning algorithm takes a raw CT image as input and views it as a collection of states, and output a classification of whether a nodule is present or not. The dataset used to train our model is the LIDC/IDRI database hosted by the lung nodule analysis (LUNA) challenge. In total, there are 888 CT scans with annotations based on agreement from at least three out of four radiologists. As a result, there are 590 individuals having one or more nodules, and 298 having none. Our training results yielded an overall accuracy of 99.1% [sensitivity 99.2%, specificity 99.1%, positive predictive value (PPV) 99.1%, negative predictive value (NPV) 99.2%]. In our test, the results yielded an overall accuracy of 64.4% (sensitivity 58.9%, specificity 55.3%, PPV 54.2%, and NPV 60.0%). These early results show promise in solving the major issue of false positives in CT screening of lung nodules, and may help to save unnecessary follow-up tests and expenditures.

Effect of radiation dose reduction and iterative reconstruction on computer-aided detection of pulmonary nodules: Intra-individual comparison

OBJECTIVE

To evaluate the effect of radiation dose reduction and iterative reconstruction (IR) on the performance of computer-aided detection (CAD) for pulmonary nodules.

METHODS

In this prospective study twenty-five patients were included who were scanned for pulmonary nodule follow-up. Image acquisition was performed at routine dose and three reduced dose levels in a single session by decreasing mAs-values with 45%, 60% and 75%. Tube voltage was fixed at 120 kVp for patients ≥ 80 kg and 100 kVp for patients < 80 kg. Data were reconstructed with filtered back projection (FBP), iDose(4) (levels 1,4,6) and IMR (levels 1-3). All noncalcified solid pulmonary nodules ≥ 4 mm identified by two radiologists in consensus served as the reference standard. Subsequently, nodule volume was measured with CAD software and compared to the reference consensus. The numbers of true-positives, false-positives and missed pulmonary nodules were evaluated as well as the sensitivity.

RESULTS

Median effective radiation dose was 2.2 mSv at routine dose and 1.2, 0.9 and 0.6 mSv at respectively 45%, 60% and 75% reduced dose. A total of 28 pulmonary nodules were included. With FBP at routine dose, 89% (25/28) of the nodules were correctly identified by CAD. This was similar at reduced dose levels with FBP, iDose(4) and IMR. CAD resulted in a median number of false-positives findings of 11 per scan with FBP at routine dose (93% of the CAD marks) increasing to 15 per scan with iDose(4) (95% of the CAD marks) and 26 per scan (96% of the CAD marks) with IMR at the lowest dose level.

CONCLUSION

CAD can identify pulmonary nodules at submillisievert dose levels with FBP, hybrid and model-based IR. However, the number of false-positive findings increased using hybrid and especially model-based IR at submillisievert dose while dose reduction did not affect the number of false-positives with FBP.

The Solitary Pulmonary Nodule

Due to the high etiological diversity and the potential for malignancy, pulmonary nodules represent a clinical challenge, becoming increasingly frequent as the number of CT examinations rises. The topic gains even more importance as clear evidence for the effectiveness of CT screening was provided by the National Lung Screening Trial (NLST). Yet, the results were tempered by the high false-positive rate and the requirement of performing further diagnostic procedures. The management of those detected solitary pulmonary nodules is currently based on the individuals' risk of developing lung cancer, the pulmonary nodule characteristics and the capability of diagnostic and therapeutic approaches.

Screening for lung cancer with low-dose computed tomography: a review of current status

Screening using low-dose computed tomography (CT) represents an exciting new development in the struggle to improve outcomes for people with lung cancer. Randomised controlled evidence demonstrating a 20% relative lung cancer mortality benefit has led to endorsement of screening by several expert bodies in the US and funding by healthcare providers. Despite this pivotal result, many questions remain regarding technical and logistical aspects of screening, cost-effectiveness and generalizability to other settings. This review discusses the rationale behind screening, the results of on-going trials, potential harms of screening and current knowledge gaps.

Cognitive and system factors contributing to diagnostic errors in radiology

OBJECTIVE

In this article, we describe some of the cognitive and system-based sources of detection and interpretation errors in diagnostic radiology and discuss potential approaches to help reduce misdiagnoses.

CONCLUSION

Every radiologist worries about missing a diagnosis or giving a false-positive reading. The retrospective error rate among radiologic examinations is approximately 30%, with real-time errors in daily radiology practice averaging 3-5%. Nearly 75% of all medical malpractice claims against radiologists are related to diagnostic errors. As medical reimbursement trends downward, radiologists attempt to compensate by undertaking additional responsibilities to increase productivity. The increased workload, rising quality expectations, cognitive biases, and poor system factors all contribute to diagnostic errors in radiology. Diagnostic errors are underrecognized and underappreciated in radiology practice. This is due to the inability to obtain reliable national estimates of the impact, the difficulty in evaluating effectiveness of potential interventions, and the poor response to systemwide solutions. Most of our clinical work is executed through type 1 processes to minimize cost, anxiety, and delay; however, type 1 processes are also vulnerable to errors. Instead of trying to completely eliminate cognitive shortcuts that serve us well most of the time, becoming aware of common biases and using metacognitive strategies to mitigate the effects have the potential to create sustainable improvement in diagnostic errors.

Technical parameters and interpretive issues in screening computed tomography scans for lung cancer

Lung cancer screening computed tomographies (CTs) differ from traditional chest CT scans in that they are performed at very low radiation doses, which allow the detection of small nodules but which have a much higher noise content than would be acceptable in a diagnostic chest CT. The technical parameters require a great deal of attention on the part of the user, because inappropriate settings could result in either excess radiation dose to the large population of screened individuals or in low-quality images with impaired nodule detectability. Both situations undermine the main goal of the screening program, which is to detect lung nodules using as low a radiation dose as can reasonably be achieved. Once an image has been obtained, there are unique interpretive issues that must be addressed mainly because of the very high noise content of the images and the high prevalence of incidental findings in the chest unrelated to the sought-after pulmonary nodules.

A novel computer-aided lung nodule detection system for CT images

PURPOSE

The paper presents a complete computer-aided detection (CAD) system for the detection of lung nodules in computed tomography images. A new mixed feature selection and classification methodology is applied for the first time on a difficult medical image analysis problem.

METHODS

The CAD system was trained and tested on images from the publicly available Lung Image Database Consortium (LIDC) on the National Cancer Institute website. The detection stage of the system consists of a nodule segmentation method based on nodule and vessel enhancement filters and a computed divergence feature to locate the centers of the nodule clusters. In the subsequent classification stage, invariant features, defined on a gauge coordinates system, are used to differentiate between real nodules and some forms of blood vessels that are easily generating false positive detections. The performance of the novel feature-selective classifier based on genetic algorithms and artificial neural networks (ANNs) is compared with that of two other established classifiers, namely, support vector machines (SVMs) and fixed-topology neural networks. A set of 235 randomly selected cases from the LIDC database was used to train the CAD system. The system has been tested on 125 independent cases from the LIDC database.

RESULTS

The overall performance of the fixed-topology ANN classifier slightly exceeds that of the other classifiers, provided the number of internal ANN nodes is chosen well. Making educated guesses about the number of internal ANN nodes is not needed in the new feature-selective classifier, and therefore this classifier remains interesting due to its flexibility and adaptability to the complexity of the classification problem to be solved. Our fixed-topology ANN classifier with 11 hidden nodes reaches a detection sensitivity of 87.5% with an average of four false positives per scan, for nodules with diameter greater than or equal to 3 mm. Analysis of the false positive items reveals that a considerable proportion (18%) of them are smaller nodules, less than 3 mm in diameter.

CONCLUSIONS

A complete CAD system incorporating novel features is presented, and its performance with three separate classifiers is compared and analyzed. The overall performance of our CAD system equipped with any of the three classifiers is well with respect to other methods described in literature.

Assessing operating characteristics of CAD algorithms in the absence of a gold standard

PURPOSE

The authors examine potential bias when using a reference reader panel as "gold standard" for estimating operating characteristics of CAD algorithms for detecting lesions. As an alternative, the authors propose latent class analysis (LCA), which does not require an external gold standard to evaluate diagnostic accuracy.

METHODS

A binomial model for multiple reader detections using different diagnostic protocols was constructed, assuming conditional independence of readings given true lesion status. Operating characteristics of all protocols were estimated by maximum likelihood LCA. Reader panel and LCA based estimates were compared using data simulated from the binomial model for a range of operating characteristics. LCA was applied to 36 thin section thoracic computed tomography data sets from the Lung Image Database Consortium (LIDC): Free search markings of four radiologists were compared to markings from four different CAD assisted radiologists. For real data, bootstrap-based resampling methods, which accommodate dependence in reader detections, are proposed to test of hypotheses of differences between detection protocols.

RESULTS

In simulation studies, reader panel based sensitivity estimates had an average relative bias (ARB) of -23% to -27%, significantly higher (p-value < 0.0001) than LCA (ARB--2% to -6%). Specificity was well estimated by both reader panel (ARB -0.6% to -0.5%) and LCA (ARB 1.4%-0.5%). Among 1145 lesion candidates LIDC considered, LCA estimated sensitivity of reference readers (55%) was significantly lower (p-value 0.006) than CAD assisted readers' (68%). Average false positives per patient for reference readers (0.95) was not significantly lower (p-value 0.28) than CAD assisted readers' (1.27).

CONCLUSIONS

Whereas a gold standard based on a consensus of readers may substantially bias sensitivity estimates, LCA may be a significantly more accurate and consistent means for evaluating diagnostic accuracy.

Computer-aided detection of lung nodules: influence of the image reconstruction kernel for computer-aided detection performance

OBJECTIVE

To evaluate the relationship between a computed tomographic reconstruction kernel and the sensitivity of a computer-aided detection (CAD) system for lung nodule detection.

METHODS

We retrospectively studied 36 consecutive patients with no known pulmonary nodules who underwent low-dose computed tomography for lung cancer screening with 3 different reconstruction kernels (B, C, and L). All series were reviewed with a commercial CAD system for lung nodule detection.

RESULTS

The 36 scans showed 231 uncalcified nodules (170 micronodules and 61 nodules). There was little variation of sensitivities for each series (82%, 88%, and 82% for the nodules of B, C, and L, respectively). When the results of 2 series were combined, sensitivities were boosted (B + C, 89%; B + L, 95%; and C + L, 96% for the nodules).

CONCLUSIONS

Sensitivity of the CAD system was influenced by the selection of the reconstruction kernel. By combining data from 2 different kernels, CAD sensitivity can be elevated without further patient radiation exposure.

Computer-aided detection (CAD) of lung nodules in CT scans: radiologist performance and reading time with incremental CAD assistance

OBJECTIVE

The diagnostic performance of radiologists using incremental CAD assistance for lung nodule detection on CT and their temporal variation in performance during CAD evaluation was assessed.

METHODS

CAD was applied to 20 chest multidetector-row computed tomography (MDCT) scans containing 190 non-calcified > or =3-mm nodules. After free search, three radiologists independently evaluated a maximum of up to 50 CAD detections/patient. Multiple free-response ROC curves were generated for free search and successive CAD evaluation, by incrementally adding CAD detections one at a time to the radiologists' performance.

RESULTS

The sensitivity for free search was 53% (range, 44%-59%) at 1.15 false positives (FP)/patient and increased with CAD to 69% (range, 59-82%) at 1.45 FP/patient. CAD evaluation initially resulted in a sharp rise in sensitivity of 14% with a minimal increase in FP over a time period of 100 s, followed by flattening of the sensitivity increase to only 2%. This transition resulted from a greater prevalence of true positive (TP) versus FP detections at early CAD evaluation and not by a temporal change in readers' performance. The time spent for TP (9.5 s +/- 4.5 s) and false negative (FN) (8.4 s +/- 6.7 s) detections was similar; FP decisions took two- to three-times longer (14.4 s +/- 8.7 s) than true negative (TN) decisions (4.7 s +/- 1.3 s).

CONCLUSIONS

When CAD output is ordered by CAD score, an initial period of rapid performance improvement slows significantly over time because of non-uniformity in the distribution of TP CAD output and not to a changing reader performance over time.

Incidentally detected small pulmonary nodules on CT

The widespread use of multidetector computed tomography for imaging of the chest has lead to a significant increase in the number of incidentally detected pulmonary nodules. The significance of these nodules is often uncertain and further investigations may be required. This article will review the spectrum of imaging appearances of small pulmonary nodules, and highlight the few features that allow confident characterization of a nodule as benign or malignant; current guidelines for the management of incidentally detected nodules will also be discussed.

Multi slice computed tomography in the study of pulmonary metastases

PURPOSE

This study was undertaken to assess the performance of 16-slice computed tomography (MSCT) using Multi-Planar Reformatting (MPR), Maximum Intensity Projection (MIP) and Volume Rendering (VR) reconstructions to study pulmonary metastases.

MATERIALS AND METHODS

CT studies of 32 patients with pulmonary metastases were retrospectively reviewed. Images were assessed for the following parameters: number, size, location, distribution of the nodules and the presence of the "mass-vessel sign". These parameters were evaluated by two observers on axial-source images and on MPR, MIP and VR reconstructions. Sensitivity of each reconstruction and interobserver agreement were calculated.

RESULTS

Two-dimensional (2D) axial images and MIP and VR reconstructions exhibited 100% sensitivity for lesions >10 mm. For nodules 6-10 mm, sensitivity was 49%-55% for the 2D images, 90% for MIP and 80%-85% for VR reconstructions. For metastasis <or= 5 mm, sensitivity was 22% for 2D images, 87%-89% for MIP and 55%-58% for VR reconstructions. Coronal and sagittal MPR, MIP and VR did not improve the detection rate compared with the corresponding axial images. MIP and VR provided overlapping results in detecting the "mass-vessel sign".

CONCLUSIONS

MIP are the most sensitive reconstructions for detecting small pulmonary nodules.