Deep learning in pulmonary nodule detection and segmentation: a systematic review

OBJECTIVES

The accurate detection and precise segmentation of lung nodules on computed tomography are key prerequisites for early diagnosis and appropriate treatment of lung cancer. This study was designed to compare detection and segmentation methods for pulmonary nodules using deep-learning techniques to fill methodological gaps and biases in the existing literature.

METHODS

This study utilized a systematic review with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses guidelines, searching PubMed, Embase, Web of Science Core Collection, and the Cochrane Library databases up to May 10, 2023. The Quality Assessment of Diagnostic Accuracy Studies 2 criteria was used to assess the risk of bias and was adjusted with the Checklist for Artificial Intelligence in Medical Imaging. The study analyzed and extracted model performance, data sources, and task-focus information.

RESULTS

After screening, we included nine studies meeting our inclusion criteria. These studies were published between 2019 and 2023 and predominantly used public datasets, with the Lung Image Database Consortium Image Collection and Image Database Resource Initiative and Lung Nodule Analysis 2016 being the most common. The studies focused on detection, segmentation, and other tasks, primarily utilizing Convolutional Neural Networks for model development. Performance evaluation covered multiple metrics, including sensitivity and the Dice coefficient.

CONCLUSIONS

This study highlights the potential power of deep learning in lung nodule detection and segmentation. It underscores the importance of standardized data processing, code and data sharing, the value of external test datasets, and the need to balance model complexity and efficiency in future research.

CLINICAL RELEVANCE STATEMENT

Deep learning demonstrates significant promise in autonomously detecting and segmenting pulmonary nodules. Future research should address methodological shortcomings and variability to enhance its clinical utility.

KEY POINTS

Deep learning shows potential in the detection and segmentation of pulmonary nodules. There are methodological gaps and biases present in the existing literature. Factors such as external validation and transparency affect the clinical application.

A CT-based deep learning model: visceral pleural invasion and survival prediction in clinical stage IA lung adenocarcinoma

Pathologic visceral pleural invasion (VPI) in patients with early-stage lung cancer can result in the upstaging of T1 to T2, in addition to having implications for surgical resection and prognostic outcomes. This study was designed with the goal of establishing and validating a CT-based deep learning (DL) model capable of predicting VPI status and stratifying patients based on their prognostic outcomes. In total, 2077 patients from three centers with pathologically confirmed clinical stage IA lung adenocarcinoma were enrolled. DL signatures were extracted with a 3D residual neural network. DL model was able to effectively predict VPI status. VPI predicted by the DL models, as well as pathologic VPI, was associated with shorter disease-free survival. The established deep learning signature provides a tool capable of aiding the accurate prediction of VPI in patients with clinical stage IA lung adenocarcinoma, thus enabling prognostic stratification.

New trend in artificial intelligence-based assistive technology for thoracic imaging

Although there is no solid agreement for artificial intelligence (AI), it refers to a computer system with intelligence similar to that of humans. Deep learning appeared in 2006, and more than 10 years have passed since the third AI boom was triggered by improvements in computing power, algorithm development, and the use of big data. In recent years, the application and development of AI technology in the medical field have intensified internationally. There is no doubt that AI will be used in clinical practice to assist in diagnostic imaging in the future. In qualitative diagnosis, it is desirable to develop an explainable AI that at least represents the basis of the diagnostic process. However, it must be kept in mind that AI is a physician-assistant system, and the final decision should be made by the physician while understanding the limitations of AI. The aim of this article is to review the application of AI technology in diagnostic imaging from PubMed database while particularly focusing on diagnostic imaging in thorax such as lesion detection and qualitative diagnosis in order to help radiologists and clinicians to become more familiar with AI in thorax.

Development and validation of convolutional neural network-based model to predict the risk of sentinel or non-sentinel lymph node metastasis in patients with breast cancer: a machine learning study

Background

For patients with sentinel lymph node (SLN) metastasis and low risk of residual non-SLN (NSLN) metastasis, axillary lymph node (ALN) dissection could lead to overtreatment. This study aimed to develop and validate an automated preoperative deep learning-based tool to predict the risk of SLN and NSLN metastasis in patients with breast cancer (BC) using dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) images.

Methods

In this machine learning study, we retrospectively enrolled 988 women with BC from three hospitals in Zhejiang, China between June 1, 2013 to December 31, 2021, June 1, 2017 to December 31, 2021, and January 1, 2019 to June 30, 2023, respectively. Patients were divided into the training set (n = 519), internal validation set (n = 129), external test set 1 (n = 296), and external test set 2 (n = 44). A convolutional neural network (CNN) model was proposed to predict the SLN and NSLN metastasis and was compared with clinical and radiomics approaches. The performance of different models to detect ALN metastasis was measured by the area under the curve (AUC), accuracy, sensitivity, and specificity. This study is registered at ChiCTR, ChiCTR2300070740.

Findings

For SLN prediction, the top-performing model (i.e., the CNN algorithm) achieved encouraging predictive performance in the internal validation set (AUC 0.899, 95% CI, 0.887-0.911), external test set 1 (AUC 0.885, 95% CI, 0.867-0.903), and external test set 2 (AUC 0.768, 95% CI, 0.738-0.798). For NSLN prediction, the CNN-based model also exhibited satisfactory performance in the internal validation set (AUC 0.800, 95% CI, 0.783-0.817), external test set 1 (AUC 0.763, 95% CI, 0.732-0.794), and external test set 2 (AUC 0.728, 95% CI, 0.719-0.738). Based on the subgroup analysis, the CNN model performed well in tumour group smaller than 2.0 cm, with the AUC of 0.801 (internal validation set) and 0.823 (external test set 1). Of 469 patients with BC, the false positive rate of SLN prediction declined from 77.9% to 32.9% using CNN model.

Interpretation

The CNN model can predict the SLN status of any detectable lesion size and condition of NSLN in patients with BC. Overall, the CNN model, employing ready DCE-MRI images could serve as a potential technique to assist surgeons in the personalized axillary treatment of in patients with BC non-invasively.

Funding

National Key Research and Development projects intergovernmental cooperation in science and technology of China, National Natural Science Foundation of China, Natural Science Foundation of Zhejiang Province, and Zhejiang Medical and Health Science Project.

Effects of Expert-Determined Reference Standards in Evaluating the Diagnostic Performance of a Deep Learning Model: A Malignant Lung Nodule Detection Task on Chest Radiographs

OBJECTIVE

Little is known about the effects of using different expert-determined reference standards when evaluating the performance of deep learning-based automatic detection (DLAD) models and their added value to radiologists. We assessed the concordance of expert-determined standards with a clinical gold standard (herein, pathological confirmation) and the effects of different expert-determined reference standards on the estimates of radiologists' diagnostic performance to detect malignant pulmonary nodules on chest radiographs with and without the assistance of a DLAD model.

MATERIALS AND METHODS

This study included chest radiographs from 50 patients with pathologically proven lung cancer and 50 controls. Five expert-determined standards were constructed using the interpretations of 10 experts: individual judgment by the most experienced expert, majority vote, consensus judgments of two and three experts, and a latent class analysis (LCA) model. In separate reader tests, additional 10 radiologists independently interpreted the radiographs and then assisted with the DLAD model. Their diagnostic performance was estimated using the clinical gold standard and various expert-determined standards as the reference standard, and the results were compared using the t test with Bonferroni correction.

RESULTS

The LCA model (sensitivity, 72.6%; specificity, 100%) was most similar to the clinical gold standard. When expert-determined standards were used, the sensitivities of radiologists and DLAD model alone were overestimated, and their specificities were underestimated (all p-values < 0.05). DLAD assistance diminished the overestimation of sensitivity but exaggerated the underestimation of specificity (all p-values < 0.001). The DLAD model improved sensitivity and specificity to a greater extent when using the clinical gold standard than when using the expert-determined standards (all p-values < 0.001), except for sensitivity with the LCA model (p = 0.094).

CONCLUSION

The LCA model was most similar to the clinical gold standard for malignant pulmonary nodule detection on chest radiographs. Expert-determined standards caused bias in measuring the diagnostic performance of the artificial intelligence model.

Diagnostic test accuracy of artificial intelligence-based imaging for lung cancer screening: A systematic review and meta-analysis

BACKGROUND

Lung cancer is the principal cause of cancer-related deaths worldwide. Early detection of lung cancer with screening is indispensable to reduce the high morbidity and mortality rates. Artificial intelligence (AI) is widely utilised in healthcare, including in the assessment of medical images. A growing number of reviews studied the application of AI in lung cancer screening, but no overarching meta-analysis has examined the diagnostic test accuracy (DTA) of AI-based imaging for lung cancer screening.

OBJECTIVE

To systematically review the DTA of AI-based imaging for lung cancer screening.

METHODS

PubMed, EMBASE, Cochrane Library, CINAHL, IEEE Xplore, Web of Science, ACM Digital Library, Scopus, PsycINFO, and ProQuest Dissertations and Theses were searched from inception to date. Studies that were published in English and that evaluated the performance of AI-based imaging for lung cancer screening were included. Two independent reviewers screened titles and abstracts and used the Quality Assessment of Diagnostic Accuracy Studies-2 tool to appraise the quality of selected studies. Grading of Recommendations Assessment, Development, and Evaluation to diagnostic tests was used to assess the certainty of evidence.

RESULTS

Twenty-six studies with 150,721 imaging data were included. Hierarchical summary receiver-operating characteristic model used for meta-analysis demonstrated that the pooled sensitivity for AI-based imaging for lung cancer screening was 94.6 % (95 % CI: 91.4 % to 96.7 %) and specificity was 93.6 % (95 % CI: 88.5 % to 96.6 %). Subgroup analyses revealed that similar results were found among different types of AI, region, data source, and year of publication, but the overall quality of evidence was very low.

CONCLUSION

AI-based imaging could effectively detect lung cancer and be incorporated into lung cancer screening programs. Further high-quality DTA studies on large lung cancer screening populations are required to validate AI's role in early lung cancer detection.

Comparisons of deep learning and machine learning while using text mining methods to identify suicide attempts of patients with mood disorders

BACKGROUND

Suicide attempt is one of the most severe consequences for patients with mood disorders. This study aimed to perform deep learning and machine learning while using text mining to identify patients with suicide attempts and to compare their effectiveness.

METHODS

A total of 13,100 patients with mood disorders were selected. Two traditional text mining methods, logistic regression and Support vector machine (SVM), and one deep learning model (Convolutional neural network, CNN) were adopted to perform overall analysis and gender-specific subgroup analysis of patients to identify suicide attempts. The classification effectiveness of these models was evaluated by accuracy, F1-value, precision, recall, and the area under Receiver operator characteristic curve (ROC).

RESULTS

CNN's results were greater than the other two for all indicators except recall which was slightly smaller than SVM in male subgroup analysis. The accuracy values of the CNN were 98.4 %, 98.2 %, and 98.5 % in the overall analysis and the subgroup analysis for males and females, respectively. The results of McNemar's test showed that CNN and SVM models' predictions were statistically different from the logistic regression model's predictions in the overall analysis and the subgroup analysis for females (P < 0.050).

LIMITATIONS

A fixed number of features were selected based on document frequency to train models; this was a single-site study.

CONCLUSIONS

CNN model was a better way to detect suicide attempts in patients with mood disorders prior to hospital admission, saving time and resources in recognizing high-risk patients and preventing suicide.

A narrative review of deep learning applications in lung cancer research: from screening to prognostication

Background and Objective

Deep learning (DL) algorithms have been developed for various tasks, including lung nodule detection on chest radiographs or lung cancer computed tomography screening, potential candidate selection in lung cancer screening, malignancy prediction for indeterminate pulmonary nodules, lung cancer staging, treatment response prediction, prognostication, and prediction of genetic mutations in lung cancer. Furthermore, these DL algorithms have been applied in various clinical settings in order for them to be generalized in real-world clinical practice. Multiple DL algorithms have been corroborated to be on par with experts or current clinical prediction models for several specific tasks. However, no article has yet comprehensively reviewed DL algorithms dedicated to lung cancer research. This narrative review presents an overview of the literature dealing with DL techniques applied in lung cancer research and briefly summarizes the results according to the DL algorithms’ clinical use cases.

Methods

we performed a narrative review by searching the Embase and OVID-MEDLINE databases for articles published in English from October, 2016 until September, 2021 and reviewing the bibliographies of key references to identify important literature related to DL in lung cancer research. The background, development, results, and clinical implications of each DL algorithm are briefly discussed. Lastly, we end this review article by highlighting future directions in lung cancer research using DL techniques.

Key Content and Findings

DL algorithms have been introduced to show comparable or higher performance than human experts in various clinical settings. Specifically, they have been actively applied to detect lung nodules in chest radiographs or computed tomography (CT) examinations, optimize candidate selection for lung cancer screening (LCS), predict the malignancy of lung nodules, stage lung cancer, and predict treatment response, patients’ prognoses, and genetic mutations in lung cancers.

Conclusions

DL algorithms have corroborated their potential value for various tasks, ranging from lung cancer screening to prognostication of lung cancer patients. Future research is warranted for the clinical application of these algorithms in daily clinical practice and verification of their real-world clinical usefulness.

Convolutional Neural Network-Based Diagnostic Model for a Solid, Indeterminate Solitary Pulmonary Nodule or Mass on Computed Tomography

PURPOSE

To establish a non-invasive diagnostic model based on convolutional neural networks (CNNs) to distinguish benign from malignant lesions manifesting as a solid, indeterminate solitary pulmonary nodule (SPN) or mass (SPM) on computed tomography (CT).

METHOD

A total of 459 patients with solid indeterminate SPNs/SPMs on CT were ultimately included in this retrospective study and assigned to the train (n=366), validation (n=46), and test (n=47) sets. Histopathologic analysis was available for each patient. An end-to-end CNN model was proposed to predict the natural history of solid indeterminate SPN/SPMs on CT. Receiver operating characteristic curves were plotted to evaluate the predictive performance of the proposed CNN model. The accuracy, sensitivity, and specificity of diagnoses by radiologists alone were compared with those of diagnoses by radiologists by using the CNN model to assess its clinical utility.

RESULTS

For the CNN model, the AUC was 91% (95% confidence interval [CI]: 0.83-0.99) in the test set. The diagnostic accuracy of radiologists with the CNN model was significantly higher than that without the model (89 vs. 66%, P<0.01; 87 vs. 61%, P<0.01; 85 vs. 66%, P=0.03, in the train, validation, and test sets, respectively). In addition, while there was a slight increase in sensitivity, the specificity improved significantly by an average of 42% (the corresponding improvements in the three sets ranged from 43, 33, and 42% to 82, 78, and 84%, respectively; P<0.01 for all).

CONCLUSION

The CNN model could be a valuable tool in non-invasively differentiating benign from malignant lesions manifesting as solid, indeterminate SPNs/SPMs on CT.

Automatic Detection of Small Intestinal Hookworms in Capsule Endoscopy Images Based on a Convolutional Neural Network

Ancylostomiasis is a fairly common small bowel parasite disease identified by capsule endoscopy (CE) for which a computer-aided clinical detection method has not been established. We sought to develop an artificial intelligence system with a convolutional neural network (CNN) to automatically detect hookworms in CE images. We trained a deep CNN system based on a YOLO-V4 (You Look Only Once-Version4) detector using 11236 CE images of hookworms. We assessed its performance by calculating the area under the receiver operating characteristic curve and its sensitivity, specificity, and accuracy using an independent test set of 10,529 small-bowel images including 531 images of hookworms. The trained CNN system required 403 seconds to evaluate 10,529 test images. The area under the curve for the detection of hookworms was 0.972 (95% confidence interval (CI), 0.967-0.978). The sensitivity, specificity, and accuracy of the CNN system were 92.2%, 91.1%, and 91.2%, respectively, at a probability score cut-off of 0.485. We developed and validated a CNN-based system for detecting hookworms in CE images. By combining this high-accuracy, high-speed, and oversight-preventing system with other CNN systems, we hope it will become an important supplement for detecting intestinal abnormalities in CE images. This trial is registered with ChiCTR2000034546 (a clinical research of artificial-intelligence-aided diagnosis for hookworms in small intestine by capsule endoscope images).

Extended application of a CT-based artificial intelligence prognostication model in patients with primary lung cancer undergoing stereotactic ablative radiotherapy

BACKGROUND AND PURPOSE

To validate a computed tomography (CT)-based deep learning prognostication model, originally developed for a surgical cohort, in patients with primary lung cancer undergoing stereotactic ablative radiotherapy (SABR).

MATERIALS AND METHODS

This retrospective study identified patients with clinical stage T1-2N0M0 lung cancer treated with SABR between 2013 and 2018. The outcomes were local recurrence-free survival (LRFS), disease-free survival (DFS), and overall survival (OS). The discrimination performance of the model, which extracted a quantitative score of cumulative risk for an adverse event up to 900 days, was evaluated using time-dependent receiver operating characteristic curve analysis. Multivariable Cox regression was performed to investigate the independent prognostic value of the model output adjusting for clinical factors including age, sex, smoking status, and clinical T category.

RESULTS

In total, 135 patients (median age, 78 years; 101 men; 78 [57.8%] adenocarcinomas and 57 [42.2%] squamous cell carcinomas) were evaluated. Most patients (117/135) were treated with 48-60 Gy in four fractions. Median biologically effective dose was 150.0 Gy (interquartile range, 126.9, 150.0 Gy). For LRFS, the area under the curve (AUC) was 0.72 (95% confidence interval [CI]: 0.58, 0.87). The AUCs were 0.70 (95% CI: 0.60, 0.81) for DFS and 0.66 (95% CI: 0.54, 0.77) for OS. Model output was associated with LRFS (adjusted hazard ratio [HR], 1.043; 95% CI: 1.003, 1.085; P = 0.04), DFS (adjusted HR, 1.03; 95% CI: 1.01, 1.05; P = 0.008), and OS (adjusted HR, 1.025; 95% CI: 1.002, 1.047; P = 0.03).

CONCLUSION

This study showed external validity and transportability of the CT-based deep learning prognostication model for radiotherapy candidates.

Machine learning for lung texture analysis on thin-section CT: Capability for assessments of disease severity and therapeutic effect for connective tissue disease patients in comparison with expert panel evaluations

BACKGROUND

The need for quantitative assessment of interstitial lung involvement on thin-section computed tomography (CT) has arisen in interstitial lung diseases including connective tissue disease (CTD).

PURPOSE

To evaluate the capability of machine learning (ML)-based CT texture analysis for disease severity and treatment response assessments in comparison with qualitatively assessed thin-section CT for patients with CTD.

MATERIAL AND METHODS

A total of 149 patients with CTD-related ILD (CTD-ILD) underwent initial and follow-up CT scans (total 364 paired serial CT examinations), pulmonary function tests, and serum KL-6 level tests. Based on all follow-up examination results, all paired serial CT examinations were assessed as "Stable" (n = 188), "Worse" (n = 98) and "Improved" (n = 78). Next, quantitative index changes were determined by software, and qualitative disease severity scores were assessed by consensus of two radiologists. To evaluate differences in each quantitative index as well as in disease severity score between paired serial CT examinations, Tukey's honestly significant difference (HSD) test was performed among the three statuses. Stepwise regression analyses were performed to determine changes in each pulmonary functional parameter and all quantitative indexes between paired serial CT scans.

RESULTS

Δ% normal lung, Δ% consolidation, Δ% ground glass opacity, Δ% reticulation, and Δdisease severity score showed significant differences among the three statuses (P < 0.05). All differences in pulmonary functional parameters were significantly affected by Δ% normal lung, Δ% reticulation, and Δ% honeycomb (0.16 ≤r² ≤0.42; P < 0.05).

CONCLUSION

ML-based CT texture analysis has better potential than qualitatively assessed thin-section CT for disease severity assessment and treatment response evaluation for CTD-ILD.

Lung Cancer and Granuloma Identification Using a Deep Learning Model to Extract 3-Dimensional Radiomics Features in CT Imaging

BACKGROUND

We aimed to evaluate a deep learning (DL) model combining perinodular and intranodular radiomics features and clinical features for preoperative differentiation of solitary granuloma nodules (GNs) from solid lung cancer nodules in patients with spiculation, lobulation, or pleural indentation on CT.

PATIENTS AND METHODS

We retrospectively recruited 915 patients with solitary solid pulmonary nodules and suspicious signs of malignancy. Data including clinical characteristics and subjective CT findings were obtained. A 3-dimensional U-Net-based DL model was used for tumor segmentation and extraction of 3-dimensional radiomics features. We used the Maximum Relevance and Minimum Redundancy (mRMR) algorithm and the eXtreme Gradient Boosting (XGBoost) algorithm to select the intranodular, perinodular, and gross nodular radiomics features. We propose a medical image DL (IDL) model, a clinical image DL (CIDL) model, a radiomics DL (RDL) model, and a clinical image radiomics DL (CIRDL) model to preoperatively differentiate GNs from solid lung cancer. Five-fold cross-validation was used to select and evaluate the models. The prediction performance of the models was evaluated using receiver operating characteristic and calibration curves.

RESULTS

The CIRDL model achieved the best performance in differentiating between GNs and solid lung cancer (area under the curve [AUC] = 0.9069), which was significantly higher compared with the IDL (AUC = 0.8322), CIDL (AUC = 0.8652), intra-RDL (AUC = 0.8583), peri-RDL (AUC = 0.8259), and gross-RDL (AUC = 0.8705) models.

CONCLUSION

The proposed CIRDL model is a noninvasive diagnostic tool to differentiate between granuloma nodules and solid lung cancer nodules and reduce the need for invasive diagnostic and surgical procedures.

Machine learning for lung CT texture analysis: Improvement of inter-observer agreement for radiological finding classification in patients with pulmonary diseases

PURPOSE

To evaluate the capability ML-based CT texture analysis for improving interobserver agreement and accuracy of radiological finding assessment in patients with COPD, interstitial lung diseases or infectious diseases.

MATERIALS AND METHODS

Training cases (n = 28), validation cases (n = 17) and test cases (n = 89) who underwent thin-section CT at a 320-detector row CT with wide volume scan and two 64-detector row CTs with helical scan were enrolled in this study. From 89 CT data, a total of 350 computationally selected ROI including normal lung, emphysema, nodular lesion, ground-glass opacity, reticulation and honeycomb were evaluated by three radiologists as well as by the software. Inter-observer agreements between consensus reading with and without using the software or software alone and standard references determined by consensus of pulmonologists and chest radiologists were determined using κ statistics. Overall distinguishing accuracies were compared among all methods by McNemar's test.

RESULTS

Agreements for consensus readings obtained with and without the software or the software alone with standard references were determined as significant and substantial or excellent (with the software: κ = 0.91, p < 0.0001; without the software: κ = 0.81, p < 0.0001; the software alone: κ = 0.79, p < 0.0001). Overall differentiation accuracy of consensus reading using the software (94.9 [332/350] %) was significantly higher than that of consensus reading without using the software (84.3 [295/350] %, p < 0.0001) and the software alone (82.3 [288/350] %, p < 0.0001).

CONCLUSION

ML-based CT texture analysis software has potential for improving interobserver agreement and accuracy for radiological finding assessments in patients with COPD, interstitial lung diseases or infectious diseases.