In tuberculosis (TB), chest radiography (CXR) patterns are highly variable, mimicking pneumonia and many other diseases. This study aims to evaluate the efficacy of Google teachable machine, a deep neural network-based image classification tool, to develop algorithm for predicting TB probability of CXRs. The training dataset included 348 TB CXRs and 3806 normal CXRs for training TB detection. We also collected 1150 abnormal CXRs and 627 normal CXRs for training abnormality detection. For external validation, we collected 250 CXRs from our hospital. We also compared the accuracy of the algorithm to five pulmonologists and radiological reports. In external validation, the AI algorithm showed areas under the curve (AUC) of 0.951 and 0.975 in validation dataset 1 and 2. The accuracy of the pulmonologists on validation dataset 2 showed AUC range of 0.936-0.995. When abnormal CXRs other than TB were added, AUC decreased in both human readers (0.843-0.888) and AI algorithm (0.828). When combine human readers with AI algorithm, the AUC further increased to 0.862-0.885. The TB CXR AI algorithm developed by using Google teachable machine in this study is effective, with the accuracy close to experienced clinical physicians, and may be helpful for detecting tuberculosis by CXR.

暂无摘要。

Countries that are high burden for TB must reverse the COVID-19 pandemic\'s devastating effects to accelerate progress toward ending TB. Vietnam\'s Double X (2X) strategy uses chest radiography (CXR) and GeneXpert (Xpert) rapid diagnostic testing to improve early detection of TB disease. Household contacts and vulnerable populations (e.g., individuals aged 60 years and older, smokers, diabetics, those with alcohol use disorders, and those previously treated for TB) with and without TB symptoms were screened in community campaigns using CXRs, followed by Xpert for those with a positive screen. In public non-TB district facilities, diabetics, respiratory outpatients, inpatients with lung disease, and other vulnerable populations underwent 2X evaluation. During COVID-19 restrictions in Vietnam, the 2X strategy improved access to TB services by decentralization to commune health stations, the lowest level of the health system, and enabling self-screening using a quick response mobile application. The number needed to screen (NNS) with CXRs to diagnose 1 person with TB disease was calculated for all 2X models and showed the highest yield among self-screeners (11 NNS with CXR), high yield for vulnerable populations in communities (60 NNS) and facilities (19 NNS), and moderately high yield for household contacts in community campaigns (154 NNS). Computer-aided diagnosis for CXRs was incorporated into community and facility implementation and improved physicians\' CXR interpretations and Xpert referral decisions. Integration of TB infection and TB disease evaluation increased eligibility for TB preventive treatment among household contacts, a major challenge during implementation. The 2X strategy increased the rational use of Xpert, employing a health system-wide approach that reached vulnerable populations with and without TB symptoms in communities and facilities for early detection of TB disease. This strategy was effectively adapted to different levels of the health system during COVID-19 restrictions and contributed to post-pandemic TB recovery in Vietnam.

Generalization of deep learning (DL) algorithms is critical for the secure implementation of computer-aided diagnosis systems in clinical practice. However, broad generalization remains to be a challenge in machine learning. This research aims to identify and study potential factors that can affect the internal validation and generalization of DL networks, namely the institution where the images come from, the image processing applied by the X-ray device, and the type of response function of the X-ray device. For these purposes, a pre-trained convolutional neural network (CNN) (VGG16) was trained three times for classifying COVID-19 and control chest radiographs with the same hyperparameters, but using different combinations of data acquired in two institutions by three different X-ray device manufacturers. Regarding internal validation, the addition of images from an external institution to the training set did not modify the algorithm\'s internal performance, however, the inclusion of images acquired by a device from a different manufacturer decreased the performance up to 8% (p < 0.05). In contrast, generalization across institutions and X-ray devices with the same type of response function was achieved. Nonetheless, generalization was not observed across devices with different types of response function. This factor was the key impediment to achieving broad generalization in our research, followed by the device\'s image-processing and the inter-institutional differences, which both reduced generalization performance to 18.9% (p < 0.05), and 9.8% (p < 0.05), respectively. Finally, clustering analysis with features extracted by the CNN was performed, revealing a substantial dependence of feature values extracted by the pre-trained CNN on the X-ray device which acquired the images.

Background Diagnosing osteoporosis is challenging due to its often asymptomatic presentation, which highlights the importance of providing screening for high-risk populations. Purpose To evaluate the effectiveness of dual-energy x-ray absorptiometry (DXA) screening in high-risk patients with osteoporosis identified by an artificial intelligence (AI) model using chest radiographs. Materials and Methods This randomized controlled trial conducted at an academic medical center included participants 40 years of age or older who had undergone chest radiography between January and December 2022 without a history of DXA examination. High-risk participants identified with the AI-enabled chest radiographs were randomly allocated to either a screening group, which was offered fully reimbursed DXA examinations between January and June 2023, or a control group, which received usual care, defined as DXA examination by a physician or patient on their own initiative without AI intervention. A logistic regression was used to test the difference in the primary outcome, new-onset osteoporosis, between the screening and control groups. Results Of the 40 658 enrolled participants, 4912 (12.1%) were identified by the AI model as high risk, with 2456 assigned to the screening group (mean age, 71.8 years ± 11.5 [SD]; 1909 female) and 2456 assigned to the control group (mean age, 72.1 years ± 11.8; 1872 female). A total of 315 of 2456 (12.8%) participants in the screening group underwent fully reimbursed DXA, and 237 of 315 (75.2%) were identified with new-onset osteoporosis. After including DXA results by means of usual care in both screening and control groups, the screening group exhibited higher rates of osteoporosis detection (272 of 2456 [11.1%] vs 27 of 2456 [1.1%]; odds ratio [OR], 11.2 [95% CI: 7.5, 16.7]; P < .001) compared with the control group. The ORs of osteoporosis diagnosis were increased in screening group participants who did not meet formalized criteria for DXA compared with those who did (OR, 23.2 [95% CI: 10.2, 53.1] vs OR, 8.0 [95% CI: 5.0, 12.6]; interactive P = .03). Conclusion Providing DXA screening to a high-risk group identified with AI-enabled chest radiographs can effectively diagnose more patients with osteoporosis. Clinical trial registration no. NCT05721157 © RSNA, 2024 Supplemental material is available for this article. See also the editorial by Smith and Rothenberg in this issue.

暂无摘要。

BACKGROUND: In medical imaging, the integration of deep-learning-based semantic segmentation algorithms with preprocessing techniques can reduce the need for human annotation and advance disease classification. Among established preprocessing techniques, Contrast Limited Adaptive Histogram Equalization (CLAHE) has demonstrated efficacy in improving segmentation algorithms across various modalities, such as X-rays and CT. However, there remains a demand for improved contrast enhancement methods considering the heterogeneity of datasets and the various contrasts across different anatomic structures.
METHODS: This study proposes a novel preprocessing technique, ps-KDE, to investigate its impact on deep learning algorithms to segment major organs in posterior-anterior chest X-rays. Ps-KDE augments image contrast by substituting pixel values based on their normalized frequency across all images. We evaluate our approach on a U-Net architecture with ResNet34 backbone pre-trained on ImageNet. Five separate models are trained to segment the heart, left lung, right lung, left clavicle, and right clavicle.
RESULTS: The model trained to segment the left lung using ps-KDE achieved a Dice score of 0.780 (SD = 0.13), while that of trained on CLAHE achieved a Dice score of 0.717 (SD = 0.19), p<0.01. ps-KDE also appears to be more robust as CLAHE-based models misclassified right lungs in select test images for the left lung model. The algorithm for performing ps-KDE is available at https://github.com/wyc79/ps-KDE.
CONCLUSIONS: Our results suggest that ps-KDE offers advantages over current preprocessing techniques when segmenting certain lung regions. This could be beneficial in subsequent analyses such as disease classification and risk stratification.

暂无摘要。

This study aimed to develop a new simple and effective prognostic model using artificial intelligence (AI)-based chest radiograph (CXR) results to predict the outcomes of pneumonia. Patients aged > 18 years, admitted the treatment of pneumonia between March 2020 and August 2021 were included. We developed prognostic models, including an AI-based consolidation score in addition to the conventional CURB-65 (confusion, urea, respiratory rate, blood pressure, and age ≥ 65) and pneumonia severity index (PSI) for predicting pneumonia outcomes, defined as 30-day mortality during admission. A total of 489 patients, including 310 and 179 patients in training and test sets, were included. In the training set, the AI-based consolidation score on CXR was a significant variable for predicting the outcome (hazard ratio 1.016, 95% confidence interval [CI] 1.001-1.031). The model that combined CURB-65, initial O2 requirement, intubation, and the AI-based consolidation score showed a significantly high C-index of 0.692 (95% CI 0.628-0.757) compared to other models. In the test set, this model also demonstrated a significantly high C-index of 0.726 (95% CI 0.644-0.809) compared to the conventional CURB-65 and PSI (p < 0.001 and 0.017, respectively). Therefore, a new prognostic model incorporating AI-based CXR results along with traditional pneumonia severity score could be a simple and useful tool for predicting pneumonia outcomes in clinical practice.

The diagnosis of pneumoconiosis is complex and subjective, leading to inevitable variability in readings. This is especially true for inexperienced doctors. To improve accuracy, a computer-assisted diagnosis system is used for more effective pneumoconiosis diagnoses. Three models (Resnet50, Resnet101, and DenseNet) were used for pneumoconiosis classification based on 1250 chest X-ray images. Three experienced and highly qualified physicians read the collected digital radiography images and classified them from category 0 to category III in a double-blinded manner. The results of the 3 physicians in agreement were considered the relative gold standards. Subsequently, 3 models were used to train and test these images and their performance was evaluated using multi-class classification metrics. We used kappa values and accuracy to evaluate the consistency and reliability of the optimal model with clinical typing. The results showed that ResNet101 was the optimal model among the 3 convolutional neural networks. The AUC of ResNet101 was 1.0, 0.9, 0.89, and 0.94 for detecting pneumoconiosis categories 0, I, II, and III, respectively. The micro-average and macro-average mean AUC values were 0.93 and 0.94, respectively. The accuracy and Kappa values of ResNet101 were 0.72 and 0.7111 for quadruple classification and 0.98 and 0.955 for dichotomous classification, respectively, compared with the relative standard classification of the clinic. This study develops a deep learning based model for screening and staging of pneumoconiosis is using chest radiographs. The ResNet101 model performed relatively better in classifying pneumoconiosis than radiologists. The dichotomous classification displayed outstanding performance, thereby indicating the feasibility of deep learning techniques in pneumoconiosis screening.