Artificial intelligence in COPD CT images: identification, staging, and quantitation.

Chronic obstructive pulmonary disease (COPD) stands as a significant global health challenge, with its intricate pathophysiological manifestations often demanding advanced diagnostic strategies. The recent applications of artificial intelligence (AI) within the realm of medical imaging, especially in computed tomography, present a promising avenue for transformative changes in COPD diagnosis and management. This review delves deep into the capabilities and advancements of AI, particularly focusing on machine learning and deep learning, and their applications in COPD identification, staging, and imaging phenotypes. Emphasis is laid on the AI-powered insights into emphysema, airway dynamics, and vascular structures. The challenges linked with data intricacies and the integration of AI in the clinical landscape are discussed. Lastly, the review casts a forward-looking perspective, highlighting emerging innovations in AI for COPD imaging and the potential of interdisciplinary collaborations, hinting at a future where AI doesn't just support but pioneers breakthroughs in COPD care. Through this review, we aim to provide a comprehensive understanding of the current state and future potential of AI in shaping the landscape of COPD diagnosis and management.

Differentiation of granulomatous nodules with lobulation and spiculation signs from solid lung adenocarcinomas using a CT deep learning model.

BACKGROUND: The diagnosis of solitary pulmonary nodules has always been a difficult and important point in clinical research, especially granulomatous nodules (GNs) with lobulation and spiculation signs, which are easily misdiagnosed as malignant tumors. Therefore, in this study, we utilised a CT deep learning (DL) model to distinguish GNs with lobulation and spiculation signs from solid lung adenocarcinomas (LADCs), to improve the diagnostic accuracy of preoperative diagnosis. METHODS: 420 patients with pathologically confirmed GNs and LADCs from three medical institutions were retrospectively enrolled. The regions of interest in non-enhanced CT (NECT) and venous contrast-enhanced CT (VECT) were identified and labeled, and self-supervised labels were constructed. Cases from institution 1 were randomly divided into a training set (TS) and an internal validation set (IVS), and cases from institutions 2 and 3 were treated as an external validation set (EVS). Training and validation were performed using self-supervised transfer learning, and the results were compared with the radiologists' diagnoses. RESULTS: The DL model achieved good performance in distinguishing GNs and LADCs, with area under curve (AUC) values of 0.917, 0.876, and 0.896 in the IVS and 0.889, 0.879, and 0.881 in the EVS for NECT, VECT, and non-enhanced with venous contrast-enhanced CT (NEVECT) images, respectively. The AUCs of radiologists 1, 2, 3, and 4 were, respectively, 0.739, 0.783, 0.883, and 0.901 in the (IVS) and 0.760, 0.760, 0.841, and 0.844 in the EVS. CONCLUSIONS: A CT DL model showed great value for preoperative differentiation of GNs with lobulation and spiculation signs from solid LADCs, and its predictive performance was higher than that of radiologists.

Learning and depicting lobe-based radiomics feature for COPD Severity staging in low-dose CT images.

BACKGROUND: Chronic obstructive pulmonary disease (COPD) is a prevalent and debilitating respiratory condition that imposes a significant healthcare burden worldwide. Accurate staging of COPD severity is crucial for patient management and treatment planning. METHODS: The retrospective study included 530 hospital patients. A lobe-based radiomics method was proposed to classify COPD severity using computed tomography (CT) images. First, we segmented the lung lobes with a convolutional neural network model. Secondly, the radiomic features of each lung lobe are extracted from CT images, the features of the five lung lobes are merged, and the selection of features is accomplished through the utilization of a variance threshold, t-Test, least absolute shrinkage and selection operator (LASSO). Finally, the COPD severity was classified by a support vector machine (SVM) classifier. RESULTS: 104 features were selected for staging COPD according to the Global initiative for chronic Obstructive Lung Disease (GOLD). The SVM classifier showed remarkable performance with an accuracy of 0.63. Moreover, an additional set of 132 features were selected to distinguish between milder (GOLD I + GOLD II) and more severe instances (GOLD III + GOLD IV) of COPD. The accuracy for SVM stood at 0.87. CONCLUSIONS: The proposed method proved that the novel lobe-based radiomics method can significantly contribute to the refinement of COPD severity staging. By combining radiomic features from each lung lobe, it can obtain a more comprehensive and rich set of features and better capture the CT radiomic features of the lung than simply observing the lung as a whole.

Using CT radiomic features based on machine learning models to subtype adrenal adenoma.

BACKGROUND: Functioning and non-functioning adrenocortical adenoma are two subtypes of benign adrenal adenoma, and their differential diagnosis is crucial. Current diagnostic procedures use an invasive method, adrenal venous sampling, for endocrinologic assessment. METHODS: This study proposes establishing an accurate differential model for subtyping adrenal adenoma using computed tomography (CT) radiomic features and machine learning (ML) methods. Dataset 1 (289 patients with adrenal adenoma) was collected to develop the models, and Dataset 2 (54 patients) was utilized for external validation. Cuboids containing the lesion were cropped from the non-contrast, arterial, and venous phase CT images, and 1,967 features were extracted from each cuboid. Ten discriminative features were selected from each phase or the combined phases. Random forest, support vector machine, logistic regression (LR), Gradient Boosting Machine, and eXtreme Gradient Boosting were used to establish prediction models. RESULTS: The highest accuracies were 72.7%, 72.7%, and 76.1% in the arterial, venous, and non-contrast phases, respectively, when using radiomic features alone with the ML classifier of LR. When features from the three CT phases were combined, the accuracy of LR reached 83.0%. After adding clinical information, the area under the receiver operating characteristic curve increased for all the machine learning methods except for LR. In Dataset 2, the accuracy of LR was the highest, reaching 77.8%. CONCLUSION: The radiomic features of the lesion in three-phase CT images can potentially suggest the functioning or non-functioning nature of adrenal adenoma. The resulting radiomic models can be a non-invasive, low-cost, and rapid method of minimizing unnecessary testing in asymptomatic patients with incidentally discovered adrenal adenoma.

A radiomics based method for prediction of prostate cancer Gleason score using enlarged region of interest.

BACKGROUND: Prostate cancer (PCa) is one of the most common cancers in men worldwide, and its timely diagnosis and treatment are becoming increasingly important. MRI is in increasing use to diagnose cancer and to distinguish between non-clinically significant and clinically significant PCa, leading to more precise diagnosis and treatment. The purpose of this study is to present a radiomics-based method for determining the Gleason score (GS) for PCa using tumour heterogeneity on multiparametric MRI (mp-MRI). METHODS: Twenty-six patients with biopsy-proven PCa were included in this study. The quantitative T2 values, apparent diffusion coefficient (ADC) and signal enhancement rates (α) were calculated using multi-echo T2 images, diffusion-weighted imaging (DWI) and dynamic contrast-enhanced MRI (DCE-MRI), for the annotated region of interests (ROI). After texture feature analysis, ROI range expansion and feature filtering was performed. Then obtained data were put into support vector machine (SVM), K-Nearest Neighbor (KNN) and other classifiers for binary classification. RESULTS: The highest classification accuracy was 73.96% for distinguishing between clinically significant (Gleason 3 + 4 and above) and non-significant cancers (Gleason 3 + 3) and 83.72% for distinguishing between Gleason 3 + 4 from Gleason 4 + 3 and above, which was achieved using initial ROIs drawn by the radiologists. The accuracy improved when using expanded ROIs to 80.67% using SVM and 88.42% using Bayesian classification for distinguishing between clinically significant and non-significant cancers and Gleason 3 + 4 from Gleason 4 + 3 and above, respectively. CONCLUSIONS: Our results indicate the research significance and value of this study for determining the GS for prostate cancer using the expansion of the ROI region.

CLSSL-ResNet: Predicting malignancy of solitary pulmonary nodules from CT images by chimeric label with self-supervised learning.

BACKGROUND: Pulmonary granulomatous nodules (GN) with spiculation or lobulation have a similar morphological appearance to solid lung adenocarcinoma (SADC) under computed tomography (CT). However, these two kinds of solid pulmonary nodules (SPN) have different malignancies and are sometimes misdiagnosed. OBJECTIVE: This study aims to predict malignancies of SPNs by a deep learning model automatically. METHODS: A chimeric label with self-supervised learning (CLSSL) is proposed to pre-train a ResNet-based network (CLSSL-ResNet) for distinguishing isolated atypical GN from SADC in CT images. The malignancy, rotation, and morphology labels are integrated into a chimeric label and utilized to pre-train a ResNet50. The pre-trained ResNet50 is then transferred and fine-tuned to predict the malignancy of SPN. Two image datasets of 428 subjects (Dataset1, 307; Dataset2, 121) from different hospitals are collected. Dataset1 is divided into training, validation, and test data by a ratio of 7:1:2 to develop the model. Dataset2 is utilized as an external validation dataset. RESULTS: CLSSL-ResNet achieves an area under the ROC curve (AUC) of 0.944 and an accuracy (ACC) of 91.3%, which was much higher than that of the consensus of two experienced chest radiologists (77.3%). CLSSL-ResNet also outperforms other self-supervised learning models and many counterparts of other backbone networks. In Dataset2, AUC and ACC of CLSSL-ResNet are 0.923 and 89.3%, respectively. Additionally, the ablation experiment result indicates higher efficiency of the chimeric label. CONCLUSION: CLSSL with morphology labels can increase the ability of feature representation by deep networks. As a non-invasive method, CLSSL-ResNet can distinguish GN from SADC via CT images and may support clinical diagnoses after further validation.

Alterations of functional connectivity of the lateral habenula in subclinical depression and major depressive disorder.

BACKGROUND: Major depressive disorder (MDD) is a common cause of disability and morbidity, affecting about 10% of the population worldwide. Subclinical depression (SD) can be understood as a precursor of MDD, and therefore provides an MDD risk indicator. The pathogenesis of MDD and SD in humans is still unclear, and the current diagnosis lacks accurate biomarkers and gold standards. METHODS: A total of 40 MDD, 34 SD, and 40 healthy control (HC) participants matched by age, gender, and education were included in this study. Resting-state functional magnetic resonance images (rs-fMRI) were used to analyze the functional connectivity (FC) of the posterior parietal thalamus (PPtha), which includes the lateral habenula, as the region of interest. Analysis of variance with the post hoc t-test test was performed to find significant differences in FC and clarify the variations in FC among the HC, SD, and MDD groups. RESULTS: Increased FC was observed between PPtha and the left inferior temporal gyrus (ITG) for MDD versus SD, and between PPtha and the right ITG for SD versus HC. Conversely, decreased FC was observed between PPtha and the right middle temporal gyrus (MTG) for MDD versus SD and MDD versus HC. The FC between PPtha and the middle frontal gyrus (MFG) in SD was higher than that in MDD and HC. Compared with the HC group, the FC of PPtha-ITG (left and right) increased in both the SD and MDD groups, PPtha-MTG (right) decreased in both the SD and MDD groups and PPtha-MFG (right) increased in the SD group and decreased in the MDD group. CONCLUSION: Through analysis of FC measured by rs-fMRI, the altered FC between PPtha and several brain regions (right and left ITG, right MTG, and right MFG) has been identified in participants with SD and MDD. Different alterations in FC between PPtha and these regions were identified for patients with depression. These findings might provide insights into the potential pathophysiological mechanisms of SD and MDD, especially related to PPtha and the lateral habenula.

Texture Analysis of Native T1 Images as a Novel Method for Noninvasive Assessment of Uremic Cardiomyopathy.

BACKGROUND: Noncontrast cardiac T1 times are increased in dialysis patients which might indicate fibrotic alterations in uremic cardiomyopathy. PURPOSE: To explore the application of the texture analysis (TA) of T1 images in the assessment of myocardial alterations in dialysis patients. STUDY TYPE: Case-control study. POPULATION: A total of 117 subjects, including 22 on hemodialysis, 44 on peritoneal dialysis, and 51 healthy controls. FIELD STRENGTH: A 3 T, steady-state free precession (SSFP) sequence, modified Look-Locker imaging (MOLLI). ASSESSMENT: Two independent, blinded researchers manually delineated endocardial and epicardial borders of the left ventricle (LV) on midventricular T1 maps for TA. STATISTICAL TESTS: Texture feature selection was performed, incorporating reproducibility verification, machine learning, and collinearity analysis. Multivariate linear regressions were performed to examine the independent associations between the selected texture features and left ventricular function in dialysis patients. Texture features' performance in discrimination was evaluated by sensitivity and specificity. Reproducibility was estimated by the intraclass correlation coefficient (ICC). RESULTS: Dialysis patients had greater T1 values than normal (P < 0.05). Five texture features were filtered out through feature selection, and four showed a statistically significant difference between dialysis patients and healthy controls. Among the four features, vertical run-length nonuniformity (VRLN) had the most remarkable difference among the control and dialysis groups (144 ± 40 vs. 257 ± 74, P < 0.05), which overlap was much smaller than Global T1 times (1268 ± 38 vs. 1308 ± 46 msec, P < 0.05). The VRLN values were notably elevated (cutoff = 170) in dialysis patients, with a specificity of 97% and a sensitivity of 88%, compared with T1 times (specificity = 76%, sensitivity = 60%). In dialysis patients, VRLN was significantly and independently associated with left ventricular ejection fraction (P < 0.05), global longitudinal strain (P < 0.05), radial strain (P < 0.05), and circumferential strain (P < 0.05); however, T1 was not. DATA CONCLUSION: The texture features obtained by TA of T1 images and VRLN may be a better parameter for assessing myocardial alterations than T1 times. LEVEL OF EVIDENCE: 4 TECHNICAL EFFICACY: Stage 3.

Altered spontaneous neural activity in the precuneus, middle and superior frontal gyri, and hippocampus in college students with subclinical depression.

BACKGROUND: Subclinical depression (ScD) is a prevalent condition associated with relatively mild depressive states, and it poses a high risk of developing into major depressive disorder (MDD). However, the neural pathology of ScD is still largely unknown. Identifying the spontaneous neural activity involved in ScD may help clarify risk factors for MDD and explore treatment strategies for mild stages of depression. METHODS: A total of 34 ScD subjects and 40 age-, sex-, and education-matched healthy controls were screened from 1105 college students. The amplitude of low-frequency fluctuation (ALFF) and regional homogeneity (ReHo) of resting-state fMRI were calculated to reveal neural activity. Strict statistical strategies, including Gaussian random field (GRF), false discovery rate (FDR), and permutation test (PT) with threshold-free cluster enhancement (TFCE), were conducted. Based on the altered ALFF and ReHo, resting-state functional connectivity (RSFC) was further analyzed using a seed-based approach. RESULTS: The right precuneus and left middle frontal gyrus (MFG) both showed significantly increased ALFF and ReHo in ScD subjects. Moreover, the left hippocampus and superior frontal gyrus (SFG) showed decreased ALFF and increased ReHo, respectively. In addition, ScD subjects showed increased RSFC between MFG and hippocampus compared to healthy controls, and significant positive correlation was found between the Beck Depression Inventory-II (BDI-II) score and RSFC from MFG to hippocampus in ScD group. CONCLUSION: Spontaneous neural activities in the right precuneus, left MFG, SFG, and hippocampus were altered in ScD subjects. Functional alterations in these dorsolateral prefrontal cortex and default mode network regions are largely related to abnormal emotional processing in ScD, and indicate strong associations with brain impairments in MDD, which provide insight into potential pathophysiology mechanisms of subclinical depression.

Deep symmetric three-dimensional convolutional neural networks for identifying acute ischemic stroke via diffusion-weighted images.

BACKGROUND: Acute ischemic stroke (AIS) results in high morbidity, disability, and mortality. Early and automatic diagnosis of AIS can help clinicians administer the appropriate interventions. OBJECTIVE: To develop a deep symmetric 3D convolutional neural network (DeepSym-3D-CNN) for automated AIS diagnosis via diffusion-weighted imaging (DWI) images. METHODS: This study includes 190 study subjects (97 AIS and 93 Non-AIS) by collecting both DWI and Apparent Diffusion Coefficient (ADC) images. 3D DWI brain images are split into left and right hemispheres and input into two paths. A map with 125×253×14×12 features is extracted by each path of Inception Modules. After the features computed from two paths are subtracted through L-2 normalization, four multi-scale convolution layers produce the final predation. Three comparative models using DWI images including MedicalNet with transfer learning, Simple DeepSym-3D-CNN (each 3D Inception Module is replaced by a simple 3D-CNN layer), and L-1 DeepSym-3D-CNN (L-2 normalization is replaced by L-1 normalization) are constructed. Moreover, using ADC images and the combination of DWI and ADC images as inputs, the performance of DeepSym-3D-CNN is also investigated. Performance levels of all three models are evaluated by 5-fold cross-validation and the values of area under ROC curve (AUC) are compared by DeLong's test. RESULTS: DeepSym-3D-CNN achieves an accuracy of 0.850 and an AUC of 0.864. DeLong's test of AUC values demonstrates that DeepSym-3D-CNN significantly outperforms other comparative models (p < 0.05). The highlighted regions in the feature maps of DeepSym-3D-CNN spatially match with AIS lesions. Meanwhile, DeepSym-3D-CNN using DWI images presents the significant higher AUC than that either using ADC images or using DWI-ADC images based on DeLong's test (p < 0.05). CONCLUSIONS: DeepSym-3D-CNN is a potential method for automatically identifying AIS via DWI images and can be extended to other diseases with asymmetric lesions.

The capability of detecting small vessels beyond the conventional MRI sensitivity using iron-based contrast agent enhanced susceptibility weighted imaging.

Imaging brain microvasculature is important in cerebrovascular diseases. However, there is still a lack of non-invasive, non-radiation, and whole-body imaging techniques to investigate them. The aim of this study is to develop an ultra-small superparamagnetic iron oxide (USPIO) enhanced susceptibility weighted imaging (SWI) method for imaging micro-vasculature in both animal (~10 µm in rat) and human brain. We hypothesized that the USPIO-SWI technique could improve the detection sensitivity of the diameter of small subpixel vessels 10-fold compared with conventional MRI methods. Computer simulations were first performed with a double-cylinder digital model to investigate the theoretical basis for this hypothesis. The theoretical results were verified using in vitro phantom studies and in vivo rat MRI studies (n = 6) with corresponding ex vivo histological examinations. Additionally, in vivo human studies (n = 3) were carried out to demonstrate the translational power of the USPIO-SWI method. By directly comparing the small vessel diameters of an in vivo rat using USPIO-SWI with the small vessel diameters of the corresponding histological slide using laser scanning confocal microscopy, 13.3-fold and 19.9-fold increases in SWI apparent diameter were obtained with 5.6 mg Fe/kg and 16.8 mg Fe/kg ferumoxytol, respectively. The USPIO-SWI method exhibited its excellent ability to detect small vessels down to about 10 µm diameter in rat brain. The in vivo human study unveiled hidden arterioles and venules and demonstrated its potential in clinical practice. Theoretical modeling simulations and in vitro phantom studies also confirmed a more than 10-fold increase in the USPIO-SWI apparent diameter compared with the actual small vessel diameter size. It is feasible to use SWI blooming effects induced by USPIO to detect small vessels (down to 10 µm in diameter for rat brain), well beyond the spatial resolution limit of conventional MRI methods. The USPIO-SWI method demonstrates higher potential in cerebrovascular disease investigations.

The impact of metabolic syndrome (MetS) on surgical outcome for patients with mostly HBV-related hepatocellular carcinoma (HCC) underwent hepatectomy.

BACKGROUND AND OBJECTIVES: The impact of metabolic syndrome (MetS) on surgical outcome, mostly in patients with HBV-related hepatocellular carcinoma (HCC) who underwent hepatectomy. METHODS: A propensity score matching analysis was conducted. Patients were categorized into two groups MetS-related hepatocellular carcinoma (MetS-HCC) and 1:1 matched non-MetS-related HCC (non-MetS-HCC). Surgical outcomes were compared between the two groups. RESULTS: Seventy-four MetS-HCC patients and 74 propensity score-matched non-MetS-HCC patients were selected for analysis. The incidence of surgical site infection was higher in the MetS-HCC group than in the non-MetS-HCC group (12.16% vs 0%, P < .005). There was no difference in the recurrence-free survival and overall survival between patients in the MetS-HCC group and in non-MetS-HCC group (P > .05). Microvascular invasion and severe postoperative complications were independent risk factors for recurrence-free survival and overall survival. CONCLUSIONS: Hepatectomy for patients with mostly HBV-related HCC in the presence of MetS can result in a higher rate of postoperative surgical site infection compared with those in the absence of MetS, but long-term survival rates are comparable.

Predicting Unnecessary Nodule Biopsies from a Small, Unbalanced, and Pathologically Proven Dataset by Transfer Learning.

This study explores an automatic diagnosis method to predict unnecessary nodule biopsy from a small, unbalanced, and pathologically proven database. The automatic diagnosis method is based on a convolutional neural network (CNN) model. Because of the small and unbalanced samples, the presented method aims to improve the transfer learning capability via the VGG16 architecture and optimize the related transfer learning parameters. For comparison purpose, a traditional machine learning method is implemented, which extracts the texture features and classifies the features by support vector machine (SVM). The database includes 68 biopsied nodules, 16 are pathologically proven benign and the remaining 52 are malignant. To consider the volumetric data by the CNN model, each image slice from each nodule volume is selected randomly until all image slices of each nodule are utilized. The leave-one-out and 10-folder cross validations are applied to train and test the randomly selected 68 image slices (one image slice from one nodule) in each experiment, respectively. The averages over all the experimental outcomes are the final results. The experiments revealed that the features from both the medical and the natural images share the similarity of focusing on simpler and less-abstract objects, leading to the conclusion that not the more the transfer convolutional layers, the better the classification results. Transfer learning from other larger datasets can supply additional information to small and unbalanced datasets to improve the classification performance. The presented method has shown the potential to adapt CNN architecture to improve the prediction of unnecessary nodule biopsy from small, unbalanced, and pathologically proven volumetric dataset.

Identification of COVID-19 samples from chest X-Ray images using deep learning: A comparison of transfer learning approaches.

BACKGROUND: The novel coronavirus disease 2019 (COVID-19) constitutes a public health emergency globally. The number of infected people and deaths are proliferating every day, which is putting tremendous pressure on our social and healthcare system. Rapid detection of COVID-19 cases is a significant step to fight against this virus as well as release pressure off the healthcare system. OBJECTIVE: One of the critical factors behind the rapid spread of COVID-19 pandemic is a lengthy clinical testing time. The imaging tool, such as Chest X-ray (CXR), can speed up the identification process. Therefore, our objective is to develop an automated CAD system for the detection of COVID-19 samples from healthy and pneumonia cases using CXR images. METHODS: Due to the scarcity of the COVID-19 benchmark dataset, we have employed deep transfer learning techniques, where we examined 15 different pre-trained CNN models to find the most suitable one for this task. RESULTS: A total of 860 images (260 COVID-19 cases, 300 healthy and 300 pneumonia cases) have been employed to investigate the performance of the proposed algorithm, where 70% images of each class are accepted for training, 15% is used for validation, and rest is for testing. It is observed that the VGG19 obtains the highest classification accuracy of 89.3% with an average precision, recall, and F1 score of 0.90, 0.89, 0.90, respectively. CONCLUSION: This study demonstrates the effectiveness of deep transfer learning techniques for the identification of COVID-19 cases using CXR images.

Non-contrast MRI for breast screening: preliminary study on detectability of benign and malignant lesions in women with dense breasts.

PURPOSE: The importance of breast cancer screening has long been known. Unfortunately, there is no imaging modality for screening women with dense breasts that is both sensitive and without concerns regarding potential side effects. The purpose of this study is to explore the possibility of combined diffusion-weighted imaging and turbo inversion recovery magnitude MRI (DWI + TIRM) to overcome the difficulty of detection sensitivity and safety. METHODS: One hundred and seventy-six breast lesions from 166 women with dense breasts were retrospectively evaluated. The lesion visibility, area under the curve (AUC), sensitivity and specificity of cancer detection by MG, DWI + TIRM, and clinical MRI were evaluated and compared. MG plus clinical MRI served as the gold standard for lesion detection and pathology served as the gold standard for cancer detection. RESULTS: Lesion visibility of DWI + TIRM (96.6%) was significantly superior to MG (67.6%) in women with dense breasts (p < 0.001). There was no significant difference compared with clinical MRI. DWI + TIRM showed higher accuracy (AUC = 0.935) and sensitivity (93.68%) for breast cancer detection than MG (AUC = 0.783, sensitivity = 46.32%), but was comparable to clinical MRI (AUC = 0.944, sensitivity = 93.68%). The specificity of DWI + TIRM (83.95%) was lower than MG (98.77%), but higher than clinical MRI (77.78%). CONCLUSIONS: DWI combined with TIRM could be a safe, sensitive, and practical alternative for screening women with dense breasts.

Correction to: Non-contrast MRI for breast screening: preliminary study on detectability of benign and malignant lesions in women with dense breasts.

In the original version of the article, the image of Figure 2 was erroneously duplicated as Figure 4. The correct version of Figure 4 is given below. The original article has been corrected.

Segmentation of lung parenchyma in CT images using CNN trained with the clustering algorithm generated dataset.

BACKGROUND: Lung segmentation constitutes a critical procedure for any clinical-decision supporting system aimed to improve the early diagnosis and treatment of lung diseases. Abnormal lungs mainly include lung parenchyma with commonalities on CT images across subjects, diseases and CT scanners, and lung lesions presenting various appearances. Segmentation of lung parenchyma can help locate and analyze the neighboring lesions, but is not well studied in the framework of machine learning. METHODS: We proposed to segment lung parenchyma using a convolutional neural network (CNN) model. To reduce the workload of manually preparing the dataset for training the CNN, one clustering algorithm based method is proposed firstly. Specifically, after splitting CT slices into image patches, the k-means clustering algorithm with two categories is performed twice using the mean and minimum intensity of image patch, respectively. A cross-shaped verification, a volume intersection, a connected component analysis and a patch expansion are followed to generate final dataset. Secondly, we design a CNN architecture consisting of only one convolutional layer with six kernels, followed by one maximum pooling layer and two fully connected layers. Using the generated dataset, a variety of CNN models are trained and optimized, and their performances are evaluated by eightfold cross-validation. A separate validation experiment is further conducted using a dataset of 201 subjects (4.62 billion patches) with lung cancer or chronic obstructive pulmonary disease, scanned by CT or PET/CT. The segmentation results by our method are compared with those yielded by manual segmentation and some available methods. RESULTS: A total of 121,728 patches are generated to train and validate the CNN models. After the parameter optimization, our CNN model achieves an average F-score of 0.9917 and an area of curve up to 0.9991 for classification of lung parenchyma and non-lung-parenchyma. The obtain model can segment the lung parenchyma accurately for 201 subjects with heterogeneous lung diseases and CT scanners. The overlap ratio between the manual segmentation and the one by our method reaches 0.96. CONCLUSIONS: The results demonstrated that the proposed clustering algorithm based method can generate the training dataset for CNN models. The obtained CNN model can segment lung parenchyma with very satisfactory performance and have the potential to locate and analyze lung lesions.

Structural and functional alterations of the tracheobronchial tree after left upper pulmonary lobectomy for lung cancer.

BACKGROUND: Pulmonary lobectomy has been a well-established curative treatment method for localized lung cancer. After left upper pulmonary lobectomy, the upward displacement of remaining lower lobe causes the distortion or kink of bronchus, which is associated with intractable cough and breathless. However, the quantitative study on structural and functional alterations of the tracheobronchial tree after lobectomy has not been reported. We sought to investigate these alterations using CT imaging analysis and computational fluid dynamics (CFD) method. METHODS: Both preoperative and postoperative CT images of 18 patients who underwent left upper pulmonary lobectomy are collected. After the tracheobronchial tree models are extracted, the angles between trachea and bronchi, the surface area and volume of the tree, and the cross-sectional area of left lower lobar bronchus are investigated. CFD method is further used to describe the airflow characteristics by the wall pressure, airflow velocity, lobar flow rate, etc. RESULTS: It is found that the angle between the trachea and the right main bronchus increases after operation, but the angle with the left main bronchus decreases. No significant alteration is observed for the surface area or volume of the tree between pre-operation and post-operation. After left upper pulmonary lobectomy, the cross-sectional area of left lower lobar bronchus is reduced for most of the patients (15/18) by 15-75%, especially for 4 patients by more than 50%. The wall pressure, airflow velocity and pressure drop significantly increase after the operation. The flow rate to the right lung increases significantly by 2-30% (but there is no significant difference between each lobe), and the flow rate to the left lung drops accordingly. Many vortices are found in various places with severe distortions. CONCLUSIONS: The favorable and unfavorable adaptive alterations of tracheobronchial tree will occur after left upper pulmonary lobectomy, and these alterations can be clarified through CT imaging and CFD analysis. The severe distortions at left lower lobar bronchus might exacerbate postoperative shortness of breath.

Deep CNN models for pulmonary nodule classification: Model modification, model integration, and transfer learning.

BACKGROUND: Deep learning has made spectacular achievements in analysing natural images, but it faces challenges for medical applications partly due to inadequate images. OBJECTIVE: Aiming to classify malignant and benign pulmonary nodules using CT images, we explore different strategies to utilize the state-of-the-art deep convolutional neural networks (CNN). METHODS: Experiments are conducted using the Lung Image Database Consortium image collection (LIDC-IDRI), which is a public database containing 1018 cases. Three strategies are implemented including to 1) modify some state-of-the-art CNN architectures, 2) integrate different CNNs and 3) adopt transfer learning. Totally, 11 deep CNN models are compared using the same dataset. RESULTS: Study demonstrates that, for the model modification scheme, a concise CifarNet performs better than the other modified CNNs with more complex architectures, achieving an area under ROC curve of AUC = 0.90. Integrated CNN models do not significantly improve the classification performance, but the model complexity is reduced. Transfer learning outperforms the other two schemes and ResNet with fine-tuning leads to the best performance with an AUC = 0.94, as well as the sensitivity of 91% and an overall accuracy of 88%. CONCLUSIONS: Model modification, model integration, and transfer learning can play important roles to identify and generate optimal deep CNN models in classifying pulmonary nodules based on CT images efficiently. Transfer learning is preferred when applying deep learning to medical imaging applications.