False Positive Reduction Research Articles

Identifying normal mammograms in a large screening population using artificial intelligence

ObjectivesTo evaluate the potential of artificial intelligence (AI) to identify normal mammograms in a screening population.MethodsIn this retrospective study, 9581 double-read mammography screening exams including 68 screen-detected cancers and 187 false positives, a subcohort of the prospective population-based Malmö Breast Tomosynthesis Screening Trial, were analysed with a deep learning–based AI system. The AI system categorises mammograms with a cancer risk score increasing from 1 to 10. The effect on cancer detection and false positives of excluding mammograms below different AI risk thresholds from reading by radiologists was investigated. A panel of three breast radiologists assessed the radiographic appearance, type, and visibility of screen-detected cancers assigned low-risk scores (≤ 5). The reduction of normal exams, cancers, and false positives for the different thresholds was presented with 95% confidence intervals (CI).ResultsIf mammograms scored 1 and 2 were excluded from screen-reading, 1829 (19.1%; 95% CI 18.3–19.9) exams could be removed, including 10 (5.3%; 95% CI 2.1–8.6) false positives but no cancers. In total, 5082 (53.0%; 95% CI 52.0–54.0) exams, including 7 (10.3%; 95% CI 3.1–17.5) cancers and 52 (27.8%; 95% CI 21.4–34.2) false positives, had low-risk scores. All, except one, of the seven screen-detected cancers with low-risk scores were judged to be clearly visible.ConclusionsThe evaluated AI system can correctly identify a proportion of a screening population as cancer-free and also reduce false positives. Thus, AI has the potential to improve mammography screening efficiency.Key Points• Retrospective study showed that AI can identify a proportion of mammograms as normal in a screening population.• Excluding normal exams from screening using AI can reduce false positives.

Reference Class-Based Improvement of Object Detection Accuracy

To date, the Frames Per Second (FPS) and accuracy of object detection based on deep learning have made rapid progress. However, the accuracy is limited by issues such as false positive (FP) cases. FP cases can trigger malfunctions in applications requiring high accuracy, such as in autonomous vehicles, where it is essential to ensure driver safety when malfunctions occur. To reduce the occurrences of FP cases, we conducted an experiment to derive the association by separately detecting a highly relevant element called a reference class, in addition to the target class to be detected. To measure the association, we obtained the integrated association by first finding the associations between the bounding boxes of the target and reference classes. Then we generated a reference class-based model by applying the integrated association to a trained model. The reference class-based model achieved approximately 15% higher accuracy than the trained model at iteration 1,000. Besides, the proposed model reduced the FP cases to approximately half of the 18.964% in the conventional method; the FP reduction through an increase in iteration was only 11.008%. The reference class can be applied in various fields, such as security and autonomous vehicle technology. It can be used to reduce the FP cases and improve the accuracy performance limits in object detection. Furthermore, it is possible to reduce the cost of reinforcing the training dataset and using high-performance hardware, and the time cost of increasing training numbers.

An automated detection system for colonoscopy images using a dual encoder-decoder model

Conventional computer-aided detection systems (CADs) for colonoscopic images utilize shape, texture, or temporal information to detect polyps, so they have limited sensitivity and specificity. This study proposes a method to extract possible polyp features automatically using convolutional neural networks (CNNs). The objective of this work aims at building up a light-weight dual encoder-decoder model structure for polyp detection in colonoscopy Images. This proposed model, though with a relatively shallow structure, is expected to have the capability of a similar performance to the methods with much deeper structures. The proposed CAD model consists of two sequential encoder-decoder networks that consist of several CNN layers and full connection layers. The front end of the model is a hetero-associator (also known as hetero-encoder) that uses backpropagation learning to generate a set of reliably corrupted labeled images with a certain degree of similarity to a ground truth image, which eliminates the need for a large amount of training data that is usually required for medical images tasks. This dual CNN architecture generates a set of noisy images that are similar to the labeled data to train its counterpart, the auto-associator (also known as auto-encoder), in order to increase the successor’s discriminative power in classification. The auto-encoder is also equipped with CNNs to simultaneously capture the features of the labeled images that contain noise. The proposed method uses features that are learned from open medical datasets and the dataset of Zhejiang University (ZJU), which contains around one thousand images. The performance of the proposed architecture is compared with a state-of-the-art detection model in terms of the metrics of the Jaccard index, the DICE similarity score, and two other geometric measures. The improvements in the performance of the proposed model are attributed to the effective reduction in false positives in the auto-encoder and the generation of noisy candidate images by the hetero-encoder.

Surgery of Benign Ovarian Masses by a Gynecological Cancer Surgeon: A Cohort Study in a Tertiary Cancer Centre.

ObjectivesThis study aimed to evaluate diagnostic performance in characterising ovarian masses by our gynaecological oncology multidisciplinary team meeting (MDM). Surgical outcome and overall impact on patients and healthcare service were also assessed.MethodsThis was a prospective cohort study of all women with adnexal masses presenting to the gynaecological oncology MDM at a central London tertiary cancer centre between February 2017 and February 2018. The multidisciplinary team (MDT) outcome, imaging details, subjective opinion, tumour markers, surgical details, and final histological diagnosis were collected. Diagnostic performance was also determined.ResultsThere were 200 eligible patients in the study period. MDM imaging review demonstrated a sensitivity of 98.4% (95% CI: 94.3% to 99.8%) and a specificity of 52% (95% CI: 40.2% to 63.7%). Thirty-five cases were false positive, either presumed invasive cancers (51%) or borderline tumours (49%). The most common histological types were serous (37%) and mucinous (31%) cystadenomas. A retrospective application of the International Ovarian Tumor Analysis (IOTA) Assessment of Different NEoplasias in the adneXa (ADNEX) model suggests a potential reduction in false-positive rates (17%). Among the false-positive cases, there was no postoperative (90 days) mortality and postoperative morbidity was 14% with only grade 2 (CD2) complications according to Clavien and Dindo's CD classification.ConclusionAn MDT has high sensitivity but low specificity when characterising ovarian masses referred with possible ovarian cancer to the tertiary centre. False-positive values in ovarian cancers are an important indicator of over-treatment. More research is required to assess other methods, such as the IOTA ADNEX model, to reduce the false-positive value.

Kidney segmentation from computed tomography images using deep neural network

Background:The precise segmentation of kidneys and kidney tumors can help medical specialists to diagnose diseases and improve treatment planning, which is highly required in clinical practice. Manual segmentation of the kidneys is extremely time-consuming and prone to variability between different specialists due to their heterogeneity. Because of this hard work, computational techniques, such as deep convolutional neural networks, have become popular in kidney segmentation tasks to assist in the early diagnosis of kidney tumors. In this study, we propose an automatic method to delimit the kidneys in computed tomography (CT) images using image processing techniques and deep convolutional neural networks (CNNs) to minimize false positives. Methods:The proposed method has four main steps: (1) acquisition of the KiTS19 dataset, (2) scope reduction using AlexNet, (3) initial segmentation using U-Net 2D, and (4) false positive reduction using image processing to maintain the largest elements (kidneys). Results:The proposed method was evaluated in 210 CTs from the KiTS19 database and obtained the best result with an average Dice coefficient of 96.33%, an average Jaccard index of 93.02%, an average sensitivity of 97.42%, an average specificity of 99.94% and an average accuracy of 99.92%. In the KiTS19 challenge, it presented an average Dice coefficient of 93.03%. Conclusion:In our method, we demonstrated that the kidney segmentation problem in CT can be solved efficiently using deep neural networks to define the scope of the problem and segment the kidneys with high precision and with the use of image processing techniques to reduce false positives.

Pulmonary nodule detection using hybrid two-stage 3D CNNs.

Early detection of pulmonary nodules is an effective way to improve patients' chances of survival. In this work, we propose a novel and efficient way to build a computer-aided detection (CAD) system for pulmonary nodules based on computed tomography (CT) scans. The system can be roughly divided into two steps: nodule candidate detection and false positive reduction. Considering the three-dimensional (3D) nature of nodules, the CAD system adopts 3D convolutional neural networks (CNNs) in both stages. Specifically, in the first stage, a segmentation-based 3D CNN with a hybrid loss is designed to segment nodules. According to the probability maps produced by the segmentation network, a threshold method and connected component analysis are applied to generate nodule candidates. In the second stage, we employ three classification-based 3D CNNs with different types of inputs to reduce false positives. In addition to simple raw data input, we also introduce hybrid inputs to make better use of the output of the previous segmentation network. In experiments, we use data augmentation and batch normalization to avoid overfitting. We evaluate the system on 888 CT scans from the publicly available LIDC-IDRI dataset, and our method achieves the best performance by comparing with the state-of-the-art methods, which has a high detection sensitivity of 97.5% with an average of only one false positive per scan. An additional evaluation on 115 CT scans from local hospitals is also performed. Experimental results demonstrate that our method is highly suited for the detection of pulmonary nodules.

MDL-IWS: Multi-view Deep Learning with Iterative Watershed for Pulmonary Fissure Segmentation.

Pulmonary fissure segmentation is important for localization of lung lesions which include nodules at respective lobar territories. This can be very useful for diagnosis as well as treatment planning. In this paper, we propose a novel coarse-to-fine fissure segmentation approach by proposing a Multi-View Deep Learning driven Iterative WaterShed Algorithm (MDL-IWS). Coarse fissure segmentation obtained from multi-view deep learning yields incomplete fissure volume of interest (VOI) with additional false positives. An iterative watershed algorithm (IWS) is presented to achieve fine segmentation of fissure surfaces. As a part of the IWS algorithm, surface fitting is used to generate a more accurate fissure VOI with substantial reduction in false positives. Additionally, a weight map is used to reduce the over-segmentation of watershed in subsequent iterations. Experiments on the publicly available LOLA11 dataset clearly reveal that our method outperforms several state-of-the-art competitors.

A Two Cascaded Network Integrating Regional-based YOLO and 3D-CNN for Cerebral Microbleeds Detection.

Cerebral Microbleeds (CMBs) are small chronic brain hemorrhages, which have been considered as diagnostic indicators for different cerebrovascular diseases including stroke, dysfunction, dementia, and cognitive impairment. In this paper, we propose a fully automated two-stage integrated deep learning approach for efficient CMBs detection, which combines a regional-based You Only Look Once (YOLO) stage for potential CMBs candidate detection and three-dimensional convolutional neural networks (3D-CNN) stage for false positives reduction. Both stages are conducted using the 3D contextual information of microbleeds from the MR susceptibility-weighted imaging (SWI) and phase images. However, we average the adjacent slices of SWI and complement the phase images independently and utilize them as a two- channel input for the regional-based YOLO method. The results in the first stage show that the proposed regional-based YOLO efficiently detected the CMBs with an overall sensitivity of 93.62% and an average number of false positives per subject (FPavg) of 52.18 throughout the five-folds cross-validation. The 3D-CNN based second stage further improved the detection performance by reducing the FPavg to 1.42. The outcomes of this work might provide useful guidelines towards applying deep learning algorithms for automatic CMBs detection.

Deep learning-based pulmonary nodule detection: Effect of slab thickness in maximum intensity projections at the nodule candidate detection stage

Background and ObjectiveTo investigate the effect of the slab thickness in maximum intensity projections (MIPs) on the candidate detection performance of a deep learning-based computer-aided detection (DL-CAD) system for pulmonary nodule detection in CT scans. MethodsThe public LUNA16 dataset includes 888 CT scans with 1186 nodules annotated by four radiologists. From those scans, MIP images were reconstructed with slab thicknesses of 5 to 50 mm (at 5 mm intervals) and 3 to 13 mm (at 2 mm intervals). The architecture in the nodule candidate detection part of the DL-CAD system was trained separately using MIP images with various slab thicknesses. Based on ten-fold cross-validation, the sensitivity and the F2 score were determined to evaluate the performance of using each slab thickness at the nodule candidate detection stage. The free-response receiver operating characteristic (FROC) curve was used to assess the performance of the whole DL-CAD system that took the results combined from 16 MIP slab thickness settings. ResultsAt the nodule candidate detection stage, the combination of results from 16 MIP slab thickness settings showed a high sensitivity of 98.0% with 46 false positives (FPs) per scan. Regarding a single MIP slab thickness of 10 mm, the highest sensitivity of 90.0% with 8 FPs/scan was reached before false positive reduction. The sensitivity increased (82.8% to 90.0%) for slab thickness of 1 to 10 mm and decreased (88.7% to 76.6%) for slab thickness of 15–50 mm. The number of FPs was decreasing with increasing slab thickness, but was stable at 5 FPs/scan at a slab thickness of 30 mm or more. After false positive reduction, the DL-CAD system, utilizing 16 MIP slab thickness settings, had the sensitivity of 94.4% with 1 FP/scan. ConclusionsThe utilization of multi-MIP images could improve the performance at the nodule candidate detection stage, even for the whole DL-CAD system. For a single slab thickness of 10 mm, the highest sensitivity for pulmonary nodule detection was reached at the nodule candidate detection stage, similar to the slab thickness usually applied by radiologists.

3D Context-Aware Convolutional Neural Network for False Positive Reduction in Clustered Microcalcifications Detection.

False positives (FPs) reduction is indispensable for clustered microcalcifications (MCs) detection in digital breast tomosynthesis (DBT), since there might be excessive false candidates in the detection stage. Considering that DBT volume has an anisotropic resolution, we proposed a novel 3D context-aware convolutional neural network (CNN) to reduce FPs, which consists of a 2D intra-slices feature extraction branch and a 3D inter-slice features fusion branch. In particular, 3D anisotropic convolutions were designed to learn representations from DBT volumes and inter-slice information fusion isonly performed on the feature map level, which could avoid the influence of anisotropic resolution of DBT volume. The proposed method was evaluated on a large-scale Chinese women population of 877 cases with 1754 DBT volumes and compared with 8 related methods. Experimental results show that the proposed network achieved the best performance with an accuracy of 92.68% for FPs reduction with an AUC of 97.65%, and the FPs are 0.0512 per DBT volume at a sensitivity of 90%. This also proved that making full use of 3D contextual information of DBT volume can improve the performance of the classification algorithm.

Automated lung cancer diagnosis using three-dimensional convolutional neural networks.

Lung cancer is the deadliest cancer worldwide. It has been shown that early detection using low-dose computer tomography (LDCT) scans can reduce deaths caused by this disease. We present a general framework for the detection of lung cancer in chest LDCT images. Our method consists of a nodule detector trained on the LIDC-IDRI dataset followed by a cancer predictor trained on the Kaggle DSB 2017 dataset and evaluated on the IEEE International Symposium on Biomedical Imaging (ISBI) 2018 Lung Nodule Malignancy Prediction test set. Our candidate extraction approach is effective to produce accurate candidates with a recall of 99.6%. In addition, our false positive reduction stage classifies successfully the candidates and increases precision by a factor of 2000. Our cancer predictor obtained a ROC AUC of 0.913 and was ranked 1st place at the ISBI 2018 Lung Nodule Malignancy Prediction challenge. Graphical abstract.

Pulmonary nodule detection on chest radiographs using balanced convolutional neural network and classic candidate detection.

Computer-aided detection (CADe) systems play a crucial role in pulmonary nodule detection via chest radiographs (CXRs). A two-stage CADe scheme usually includes nodule candidate detection and false positive reduction. A pure deep learning model, such as faster region convolutional neural network (faster R-CNN), has been successfully applied for nodule candidate detection via computed tomography (CT). The model is yet to achieve a satisfactory performance in CXR, because the size of the CXR is relatively large and the nodule in CXR has been obscured by structures such as ribs. In contrast, the CNN has proved effective for false positive reduction compared to the shallow method. In this paper, we developed a CADe scheme using the balanced CNN with classic candidate detection. First, the scheme applied a multi-segment active shape model to accurately segment pulmonary parenchyma. The grayscale morphological enhancement technique was then used to improve the conspicuity of the nodule structure. Based on the nodule enhancement image, 200 nodule candidates were selected and a region of interest (ROI) was cropped for each. Nodules in CXR exhibit a large variation in density, and rib crossing and vessel tissue usually present similar features to the nodule. Compared to the original ROI image, the nodule enhancement ROI image has potential discriminative features from false positive reduction. In this study, the nodule enhancement ROI image, corresponding segmentation result, and original ROI image were encoded into a red-green-blue (RGB) color image instead of the duplicated original ROI image as input of the CNN (GoogLeNet) for false positive reduction. With the Japanese Society of Radiological Technology database, the CADe scheme achieved high performance of the published literatures (a sensitivity of 91.4 % and 97.1 %, with 2.0 false positives per image (FPs/image) and 5.0 FPs/image, respectively) for nodule cases.

Can grating interferometry-based mammography discriminate benign from malignant microcalcifications in fresh biopsy samples?

PurposeIn addition to absorption imaging, grating interferometry-based mammography (GIM) is capable of detecting differential-phase and scattering signals. In particular, the scattering signal can enable a quantifiable characterization of breast lesions. The purpose of this study was to determine if suspicious microcalcifications associated with benign or malignant lesions can be discriminated based on their absorption and scattering properties. Materials and MethodsIn this prospective, ethically approved study, 62 patients (mean age 60 y, range 39–89) with suspicious microcalcifications, who underwent stereotactic biopsies, were included. Biopsies were measured with an experimental GIM device and the ratios of the scattering and absorption signal (R-value) for microcalcifications were calculated. The mean R-values for benign and malignant lesions associated with microcalcifications were compared with the final histopathological diagnosis using a t-test. ResultsTwenty of the 62 participants had microcalcifications associated with malignancy. Comparing the two largest histopathological sub-groups of fibrosis (n = 23) vs. ductal carcinoma in situ (n = 15) resulted in an average R-value of 4.08 for benign and 2.80 for malignant lesions; p = 0.07. All microcalcifications associated with malignancy had an R-value below 4.71. Excluding microcalcifications with an R-value above this threshold would result in an 11 % reduction of false positives. ConclusionThe novel GIM modality has the potential to non-invasively characterize microcalcifications and might aid in the discrimination of benign from malignant lesions in fresh biopsy samples.

ElixirNet: Relation-Aware Network Architecture Adaptation for Medical Lesion Detection

Most advances in medical lesion detection network are limited to subtle modification on the conventional detection network designed for natural images. However, there exists a vast domain gap between medical images and natural images where the medical image detection often suffers from several domain-specific challenges, such as high lesion/background similarity, dominant tiny lesions, and severe class imbalance. Is a hand-crafted detection network tailored for natural image undoubtedly good enough over a discrepant medical lesion domain? Is there more powerful operations, filters, and sub-networks that better fit the medical lesion detection problem to be discovered? In this paper, we introduce a novel ElixirNet that includes three components: 1) TruncatedRPN balances positive and negative data for false positive reduction; 2) Auto-lesion Block is automatically customized for medical images to incorporates relation-aware operations among region proposals, and leads to more suitable and efficient classification and localization. 3) Relation transfer module incorporates the semantic relationship and transfers the relevant contextual information with an interpretable graph, thus alleviates the problem of lack of annotations for all types of lesions. Experiments on DeepLesion and Kits19 prove the effectiveness of ElixirNet, achieving improvement of both sensitivity and precision over FPN with fewer parameters.

Recurrent attention network for false positive reduction in the detection of pulmonary nodules in thoracic CT scans.

Multiview two-dimensional (2D) convolutional neural networks (CNNs) and three-dimensional (3D) CNNs have been successfully used for analyzing volumetric data in many state-of-the-art medical imaging applications. We propose an alternative modular framework that analyzes volumetric data with an approach that is analogous to radiologists' interpretation, and apply the framework to reduce false positives that are generated in computer-aided detection (CADe) systems for pulmonary nodules in thoracic computed tomography (CT) scans. In our approach, a deep network consisting of 2D CNNs first processes slices individually. The features extracted in this stage are then passed to a recurrent neural network (RNN), thereby modeling consecutive slices as a sequence of temporal data and capturing the contextual information across all three dimensions in the volume of interest. Outputs of the RNN layer are weighed before the final fully connected layer, enabling the network to scale the importance of different slices within a volume of interest in an end-to-end training framework. We validated the proposed architecture on the false positive reduction track of the lung nodule analysis (LUNA) challenge for pulmonary nodule detection in chest CT scans, and obtained competitive results compared to 3D CNNs. Our results show that the proposed approach can encode the 3D information in volumetric data effectively by achieving a sensitivity >0.8 with just 1/8 false positives per scan. Our experimental results demonstrate the effectiveness of temporal analysis of volumetric images for the application of false positive reduction in chest CT scans and show that state-of-the-art 2D architectures from the literature can be directly applied to analyzing volumetric medical data. As newer and better 2D architectures are being developed at a much faster rate compared to 3D architectures, our approach makes it easy to obtain state-of-the-art performance on volumetric data using new 2D architectures.

Lipid Annotator: Towards Accurate Annotation in Non-Targeted Liquid Chromatography High-Resolution Tandem Mass Spectrometry (LC-HRMS/MS) Lipidomics Using A Rapid and User-Friendly Software.

Lipidomics has great promise in various applications; however, a major bottleneck in lipidomics is the accurate and comprehensive annotation of high-resolution tandem mass spectral data. While the number of available lipidomics software has drastically increased over the past five years, the reduction of false positives and the realization of obtaining structurally accurate annotations remains a significant challenge. We introduce Lipid Annotator, which is a user-friendly software for lipidomic analysis of data collected by liquid chromatography high-resolution tandem mass spectrometry (LC-HRMS/MS). We validate annotation accuracy against lipid standards and other lipidomics software. Lipid Annotator was integrated into a workflow applying an iterative exclusion MS/MS acquisition strategy to National Institute of Standards and Technology (NIST) SRM 1950 Metabolites in Frozen Human Plasma using reverse phase LC-HRMS/MS. Lipid Annotator, LipidMatch, and MS-DIAL produced consensus annotations at the level of lipid class for 98% and 96% of features detected in positive and negative mode, respectively. Lipid Annotator provides percentages of fatty acyl constituent species and employs scoring algorithms based on probability theory, which is less subjective than the tolerance and weighted match scores commonly used by available software. Lipid Annotator enables analysis of large sample cohorts and improves data-processing throughput as compared to previous lipidomics software.

Evaluation of data preprocessings for the comparison of GC–MS chemical profiles of seized cannabis samples

Cannabis is the most frequently used illicit drug in Belgium, where it is mainly cultivated indoor. To improve the fight against this drug, cannabis-profiling methods are required. Cannabis is a natural product and its chemical composition depends on many factors, which cause a high heterogeneity and variability in the secondary metabolites, and make this study challenging. The aim of this study is to combine cannabis profiling with statistical methodology to evaluate the intra (within)- and inter (between)–plantation variabilities with the goal to define a suitable approach linking seized marijuana to given plantations. The data set used contains 46 samples from 9 locations. The chemical profiles, consisting of data from eight cannabinoids, are obtained by gas chromatography - mass spectrometry. The raw data (peak areas) is pretreated with different preprocessing methods. The Pearson correlation coefficients between intra-location profiles were calculated after each pre-treatment, and the 95 and 99 % confidence limits determined. All preprocessed data were then compared with the internal standard normalization reference method with the aim to minimize the overlap between intra- and inter-location results, i.e. to reduce the number of false positives, and to obtain the best discrimination. Furthermore, cross-validation was used to evaluate the model originating from the most efficient data pre-treatment technique. The best results were obtained, when the peak areas were normalized to the internal standard with subsequent calculation of the fourth root. It results in a reduction of false positives for both confidence limits to 11 % and 14 % compared to 21 % and 27 % for the reference method. Cross-validation reveals similar false positive results as for the calibration set. In conclusion, when preprocessing the data, an improved model is obtained resulting in a significant decrease in the number of false positives. After studying the predictive performance of the model, it appears to be representative for the entire plantation information.

Morphological Detection and Extraction of Rail Surface Defects

Rail inspection by means of a visual system has been a subject of a number of publications in recent years. The main requirements with regard to such a system are that it has to be fast, nondestructive, and accurate. This article presents a system for rail defect detection and shape extraction utilizing morphological operations. The proposed system is evaluated using publicly available rail surface discrete defects (RSDDs) data sets of Type-I and Type-II. Several versions of the system are compared, and the best one is reduced to just 17 basic morphological operations implementing defect detection, defect extraction, and false-positive (FP) reduction scheme. Two evaluation methods are considered, which is a pixel-level method and a defect-level method. 70% precision and 79% recall were obtained at the defect level for Type-I RSDD. Similarly, 82% precision and 89% recall were obtained at the defect level for Type-II RSDD. Experiments showed that the FP reduction scheme is effective for type II RSDD when considering defect-level results, and it increases precision by 7%–9% while maintaining the recall almost intact. The execution time of the morphological operations of the detection algorithm together with FP reduction is in the order of 50 ms on a usual laptop.

An Embedded Multi-branch 3D Convolution Neural Network for False Positive Reduction in Lung Nodule Detection.

Numerous lung nodule candidates can be produced through an automated lung nodule detection system. Classifying these candidates to reduce false positives is an important step in the detection process. The objective during this paper is to predict real nodules from a large number of pulmonary nodule candidates. Facing the challenge of the classification task, we propose a novel 3D convolution neural network (CNN) to reduce false positives in lung nodule detection. The novel 3D CNN includes embedded multiple branches in its structure. Each branch processes a feature map from a layer with different depths. All of these branches are cascaded at their ends; thus, features from different depth layers are combined to predict the categories of candidates. The proposed method obtains a competitive score in lung nodule candidate classification on LUNA16 dataset with an accuracy of 0.9783, a sensitivity of 0.8771, a precision of 0.9426, and a specificity of 0.9925. Moreover, a good performance on the competition performance metric (CPM) is also obtained with a score of 0.830. As a 3D CNN, the proposed model can learn complete and three-dimensional discriminative information about nodules and non-nodules to avoid some misidentification problems caused due to lack of spatial correlation information extracted from traditional methods or 2D networks. As an embedded multi-branch structure, the model is also more effective in recognizing the nodules of various shapes and sizes. As a result, the proposed method gains a competitive score on the false positive reduction in lung nodule detection and can be used as a reference for classifying nodule candidates.

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

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. 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. 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. 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.