Application of deep learning to automated diagnosis of computed tomography images of pulmonary infections

Application of deep learning to automated diagnosis of computed tomography images of pulmonary infections | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Application of deep learning to automated diagnosis of computed tomography images of pulmonary infections Beibei Meng, Changyu Lei, Qingjia Chi, Zhenhong Hu, Haichao Liu, and 3 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6323914/v1 This work is licensed under a CC BY 4.0 License Status: Under Revision Version 1 posted 12 You are reading this latest preprint version Abstract Background: Currently, clinical pathogen diagnosis primarily consists of two methods: high-throughput sequencing and traditional culture. However, considering the evident limitations in terms of time and cost associated with the diagnostic process, there is an urgent need for a highly sensitive diagnostic and detection system to complement traditional methods. Therefore, this study proposes a Mix-inResNet model based on Convolutional Neural Network (CNN) aiming to enhance the accuracy of pulmonary infection diagnosis and effectively detect various types of lung diseases. Purpose : This study aimed to develop a deep learning model, Mix-inResNet, based on computed tomography (CT) scannings for the preclassification of pneumonia-related pathologies. Methods : A total of 273 cases of patients with pulmonary infections were included in this study, and both metagenomic next-generation sequencing (mNGS) and traditional culture methods were employed to comfirm the pathology. We possess a substantial collection of 2,858 imaging datasets with notable levels of infection were included to compare the diagnostic efficacy of the Mix-inResNet model against other models (ResNet 101, Inception V3, and VGG 16), as well as clinicians and imaging specialists. Results :The Mix-inResNet model achieved precision, accuracy, F1 score, and Gwet's kappa values of 94.61%, 94.41%, 94.22%, and 93.09%, respectively. For bacterial, fungal, viral, and tuberculous pneumonia, as well as normal chest CT tests, the sensitivities were 88.14%, 92.11%, 91.94%, 98.00%, and 100%, respectively. The specificities were 98.68%, 100.00%, 97.77%, 97.46%, and 99.04%, respectively, with 95% confidence intervals ranging from 0.88 to 0.97. Compared to other models or manual identification, the Mix-inResNet model demonstrated the highest area under the curve (AUC) values and achieved more accurate classification. Confusion matrix results indicate that the Mix-inResNet model provides an excellent understanding and accurate identification of pneumonia. Additionally, Grad-CAM technology was utilized to localize the site of pneumonia, offering more intuitive and clear results. Conclusions : The Mix-inResNet model performs at a level comparable to imaging specialists in distinguishing pneumonia classifications. It effectively reduces the misdiagnosis risk and can promptly deliver relevant pathogenic information. This provides clinicians with rapid and accurate diagnostic support, aiding the implementation of early therapeutic interventions. pulmonary infection CT scan deep learning artificial intelligence accuracy Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 Figure 9 1. Introduction Pulmonary infections are increasing in morbidity and mortality worldwide [1] . According to the World Health Organization's (WHO) 2019 report on the world's leading causes of death, pulmonary infections were ranked fourth, causing 2.6 million deaths and affecting 489 million people globally. Additionally, it is in the top 10 causes of death from infectious diseases in countries of different income levels [2, 3] . Pulmonary infections can be caused by many different pathogens, each with its clinical characteristics. It is crucial to quickly and accurately identify these pathogens to make an accurate treatment decisions. However, traditional diagnostic methods heavily rely on physician experience and subjective judgment, which can limit the accuracy and efficiency of diagnosis. Previous studies have found that general practitioners had a sensitivity of 29% and a specificity of 99% upon pneumonia [4] . The current diagnostic process for pulmonary infections has significant limitations regarding timeliness and affordability. Therefore, there is an urgent need for a highly sensitive diagnostic and detection system to assist the traditional methods, which are often time-consuming and expensive. Automated disease diagnostics based on Artificial Intelligence (AI) have shown significant growth potential. The demand of sensitive diagnostic and detection system has triggered a boom in AI technology in healthcare and demonstrated a wide range of potential applications [5, 6] . AI technology has profoundly transformed the medical and healthcare sectors, particularly in the areas of drug discovery and medical imaging. Its accuracy in medical imaging highlights its potential to drive further advancements in the field [7] . Machine Learning (ML) and Deep Learning (DL), as two key technologies embodying AI, have been tested in disease diagnosis [8] , lesion detection [9] , prognosis analysis [10] , and other medical fields. In certain automated medical image recognition and detection tasks, deep learning models have shown detection levels comparable to radiologists [7, 11] . Given the impressive performance and established effectiveness of deep learning in image recognition and classification tasks, numerous researchers have explored integrating this technology into CT image analysis to accurately classify and predict various pulmonary diseases [12-14] . CT imaging is valuable for diagnosing pneumonia. With the help of AI, we can create CT image-based tools to accurately classify pneumonia. The 2019 coronavirus disease pandemic has highlighted the need for alternative methods due to the misclassification of similar respiratory disease symptoms. In contrast, there have been AI studies in chest imaging, such as CT image-based classification of new coronavirus and common pneumonia [15, 16] , the classification of pediatric bacterial pneumonia and viral pneumonia based on chest radiographs [17] , and the identification of pneumonic lung cancer and pneumonia based on chest CT images [18] . A limited number of studies are available on identifying pneumonia caused by various pathogen infections using chest CT images. Therefore, this study proposes a convolutional neural network (CNN)-based model for the detection of various types of pulmonary diseases, aiming to enhance the recognition accuracy in pulmonary infection diagnosis. The performance of CT scan datasets is analyzed and compared with conventional neural networks and clinical experts to improve diagnose accuracy, based on pathogen confirmation by mNGS and traditional culture. CT imaging can detect abnormal changes in the chest before clinical manifestation, making early and timely critical diagnosis. 1. Comparison of flowcharts between traditional culture and the Mix-inResNet model. Fig. 1 shows a comparison of traditional traditional culture methods and deep learning methods in imaging pathogen diagnosis and their advantages and disadvantages. 2. Methods and materials 2.1 Data sets 2.1.1 Data distribution In this study, we collected data from 273 patients (152 males and 121 females) whose ages ranged from 42 to 69, and who underwent chest CT scans. Their alveolar lavage fluid (BALF) was sent for metagenomic next generation sequencing (mNGS) and conventional cultures, which indicated single pathogen infections. The data was collected from March 2021 to April 2024 at our hospital. Following manual screening, we compiled an imaging dataset comprising 2858 images featuring obvious infection types, including 589 cases of bacterial pneumonia, 380 cases of fungal pneumonia, 619 cases of viral pneumonia, 503 cases of pulmonary tuberculosis, and 767 images of normal conditions. These images were divided into training, validation, and testing sets in an 8:1:1 ratio, as illustrated in Figure 2. 2.1.2 Diagnosis of pathogens Bronchoalveolar lavage fluid (BALF ） : Before surgery, conduct comprehensive preoperative examinations to exclude significant cardiovascular and cerebrovascular diseases, as well as a history of anticoagulant medication usage. Instruct the patient to observe a water fast for a duration of 4 hours prior to the surgical procedure. Select the specific lung segment for bronchoalveolar lavage based on chest CT imaging, ensuring an adequate sample volume of no less than 5mL. Subsequently, perform both conventional laboratory culture and high-throughput sequencing techniques on the collected samples. Sputum specimen: A sputum specimen was collected and sent to the Laboratory Department of the hospital. It was inoculated into the appropriate medium, commonly used LB agar culture plate and blood agar culture plate, and cultured for 24-72 hours. Positive colonies were selected and tested using the Mérieux biochemical identifier from France. Based on the results of the biochemical identification, a culture identification report was issued. The clinician will then consider the clinical manifestations of the case, exclude the possibility of background or contaminating bacteria, and decide whether the identified pathogen is the cause of the pulmonary infection. Metagenomic next generation sequenc ing (mNGS): A high-throughput sequencing method that operates independently of traditional microbial culture. This approach does not necessitate preconceived assumptions regarding potential pathogenic microorganism. It is characterized by its unbiased nature, broad coverage, rapid detection capabilities, high sensitivity, and accuracy [19] . The interpretation criteria for a positive mNGS test include the following four items [20, 21] (1) mNGS and conventional culture identified the same microorganism, and the reads number of mNGS of a single species is greater than 50; (2) Mycobacterium tuberculosis sequence number at the species or genus level is judged positive when at least 1 sequence number (reads) is consistent with the reference genome; (3) The relative abundance of the bacteria or fungi at the genus level is >30%; (4) The coverage of microorganisms is 10 times higher than any other bacteria and 5 times higher than any other fungi, as illustrated in Figure 3. 2.1.3 Exclusion criteria: Inclusion criteria: 1. Patients diagnosed with pulmonary infection and confirmed by BLAF examination with mNGS or traditional culture showing single pathogen infection; 2. Undergo chest CT examination during hospitalization. Exclusion criteria: 1. Poor image quality, heavy artifacts affecting the observation results; 2. Patients with severe pulmonary comorbidities; 3. History of chest surgery; 4. Incomplete clinical data. 2.2 Image annotation method : Image identification and grouping were conducted by a trained and qualified clinician specialized in image recognition, along with the deputy chief physician of the imaging department. In contrast, the process was carried out without considering the patient's clinical information and other imaging data. The anonymized dataset was taken with RadiAnt DICOM Viewer 2023.1.0 software, WL: -550 HU, WW: 1350 HU, image size of 512*512, resolution of 96 dpi, and the save format was selected as DICOM mode, and the anonymized dataset was categorized separately: bacterial, fungal, viral, tuberculous pneumonia, and normal CT image. 2.3 CT Scanning Parameters and Typical Chest CTs 2.3.1 CT parameters CT examination was performed using a Toshiba Aquilion 16-row spiral CT machine. Scanning parameters: tube voltage of 120 kV, tube current of 250 mA, scanning collimation of 1.0 mm × 16, pitch of 1.3, rotation time of 0.5 s/r, field of view of 500 mm × 500 mm, acquisition matrix of 512 × 512, CT scanning was completed, and the images were uploaded to the CT workstation to complete the reconstruction of the 2-mm thin-layer pulmonary CT images using volumetric data. Image Acquisition: The patient was placed in the supine position. The patient was placed in the supine position and was instructed to hold his/her breath for 10 s. All scans were performed at the end of deep inhalation, covering the range from the rib-diaphragm angle at the base of the lungs to the thoracic inlet. Standard lung window (window position -550 HU, window width 1350 HU) and mediastinal window (window position 40 HU, window width 350 HU) were used for observation. 2.3.2 Typical Chest CT Bacterial pneumonia often presents with thickening and blurring of the lung texture, patchy consolidation, aerated bronchial signs, associated cavities, and ground-glass shadows [22] . Cavitation, air crescent signs, haloing, and ring-like appearance often characterize fungal pneumonia [23] . Typical signs of viral pneumonia are increased lung texture, ground-glass shadows, small patchy or extensive infiltrates, and solid changes. In severe cases, diffuse nodular infiltrates in both lungs, but lobar solid changes and pleural effusions are uncommon [24] . Common indicators of pulmonary tuberculosis include lesions in the upper lobe's postapical segment, the lower lobe's dorsal segment, and the posterior basal segment. These lesions display various forms, such as infiltration, proliferation, and the presence of caseous and fibrocalcified lesions, with the features of uneven density, clearer margins, and chronic changes. They also tend to form cavities and spread [25] , as illustrated in Figure 4. Note: Red arrows point to foci of infection. 2.4 Overview of the model structure Mix-inResNet comprises a backbone network, a basic convolution module (BlockA-C), a dimensionality reduction module (ReductionA-C), a self-attention module, an average pooling layer, a dropout layer, and a softmax layer. The detailed design structure is shown in the attached material. In the backbone network, hybrid convolution is used to fuse features processed by depth separable convolution, dilation convolution and standard convolution. The aim is to achieve a deeper, broader, and more abstract level by integrating features of different scopes and levels of abstraction. The network design utilizes the Inception network's features to simultaneously perform convolution and pooling operations at multiple scales and then combine the results in parallel. This multi-scale parallel strategy aims to capture and integrate image features at different scales and orientations. In Block A-C, a ResNet feature is introduced that uses "residual concatenation" to address the gradient vanishing problem and enable the training of deeper hierarchical structures, thus improving performance. Please refer to the related paper for more information about Inception and ResNet. [26-28] . 2.5 Image preprocessing The images were classified using various neural networks (Mix-inResNet, ResNet 101, Inception V3, and VGG 16), and their classification results were compared and evaluated. All input images were normalized to 256 × 256 pixels. To improve the generalization ability and robustness of the network training, sophisticated data augmentation techniques such as specular reflection, random rotation, random cropping, and gaussian noise injection were applied to the input images for all classification networks. The hidden layer weights were randomly initialized at the beginning of training and optimized using the SDG optimizer with a custom learning rate scheduler. The initial learning rate was set to 1e-4, and momentum was set to 0.9 and 0.0001, respectively. The network was trained for 300 epochs with a batch size of 8. All networks were implemented using PyTorch, and the code was run on a server equipped with a RTX 4080 GPU. All network operations were performed using PyTorch ( https://pytorch.org ). 2.6 Model training In model training, the loss function used is Cross-Entropy Loss (CEL). An Auxiliary Classifier is also incorporated to provide extra gradient signals, enhancing feature extraction capability [29] . The modeling process is also improved by introducing an Auxiliary Classifier to provide additional gradient signals to enhance feature extraction. This study describes the loss function in the model and its formulas as follows. Where: in Equation 1, N is the number of samples, and is the actual label. is the probability distribution predicted by the model. In Equation 2 is the cross-entropy loss of the primary classifier, and is the cross-entropy loss of the auxiliary classifier, both of which use the same cross-entropy loss formula in Equation 1, and is the weight coefficient of the auxiliary loss. We have included the loss function curves to further demonstrate the performance of the models. As depicted in Figure 5, during the 0th-20th Epoch, the model rapidly converges and learns the features. The model gradually decreases from the 20th to the 170th Epoch, indicating a progressive fit to the features. After 170 Epochs, the model stabilizes, suggesting it is approaching the minimum loss point. The Mix-inResNe model, represented by the yellow line, demonstrates the best convergence performance and the fastest learning results among these models. 3. Statistical methods We conducted statistical and deep learning analysis using Python (version 3.8.0). We utilized the Gradient-weighted Class Activation Mapping (Grad-CAM) technique to visually identify the infection sites of different pathogens. Data processing was performed using SPSS 26.0 software. Normally distributed measures were presented as mean ± standard deviation, while non-normally distributed data were described as median (M) and interquartile range (P25, P75). To evaluate the diagnostic performance of the Mix-inResNet model in manual classification, we utilized quantitative assessment metrics, including accuracy, precision, sensitivity, specificity, F1 score, confusion matrix, and ROC. Additionally, we assessed the inter-reader and clinical reliability using Gwet's kappa values and their 95% confidence intervals. 4. Results 4.1 Comparison of the model with physicians in terms of sensitivity, specificity and 95% confidence intervals Based on the findings in Table 1 , the Mix-inResNet model demonstrated high sensitivity in detecting bacterial, fungal, viral, tuberculous pneumonia, and normal chest CT, with percentages of 88.14%, 92.11%, 91.94%, 98.00%, and 100.00%, respectively. Additionally, the model exhibited high specificity, with percentages of 98.68%, 100.00%, 97.77%, 97.46%, and 99.04%, respectively. Compared to other models, the Mix-inResNet model showed significant advantages in sensitivity and specificity. Notably, it achieved 100% specificity in fungal detection, indicating a low misdiagnosis rate and enhancing its clinical applicability. The results indicate that the Mix-inResNet model outperformed other models (such as Inception V3, ResNet101, VGG16) and even demonstrated comparable performance to radiologist, with confidence intervals of 0.88-0.97. Table 1 Comparison of sensitivity, specificity and 95% confidence intervals among models for different pneumonia types and normal chest CT scans. Sensitivity Bacterial pneumonia Fungal pneumonia Viral pneumonia Pulmonary tuberculosis Normal Mix-inResNet 88.14% 92.11% 91.94% 98.00% 100.00% Inception V3 83.05% 63.16% 90.32% 86.00% 100.00% ResNet101 77.97% 86.84% 93.55% 98.00% 100.00% VGG16 79.66% 60.53% 83.87% 82.00% 100.00% Clinician 88.14% 73.68% 91.94% 92.00% 100.00% Radiologist 89.83% 92.11% 93.55% 98.00% 100.00% Specificity Mix-inResNet 98.68% 100.00% 97.77% 97.46% 99.04% Inception V3 92.51% 99.19% 97.32% 96.61% 98.09% ResNet101 98.68% 97.98% 97.32% 96.61% 99.52% VGG16 95.15% 96.77% 92.86% 95.76% 99.52% Clinician 94.27% 100.00% 99.11% 96.61% 98.56% Radiologist 98.68% 100.00% 97.77% 9 7.88% 99.52% 95% Confidence Interval Mix-inResNet 0.88-0.97 Inception V3 0.80-0.88 ResNet101 0.86-0.95 VGG16 0.76-0.85 Clinician 0.84-0.93 Radiologist 0.89-0.98 4.2 Comparison of precision, accuracy, F1 score and Gwet's kappa values In this study, the Mix-inResNet model achieved a precision of 94.61%, accuracy of 94.41%, F1 score of 94.22%, and Gwet's kappa value of 93.09%. In comparison, the radiologist results were 95.23% precision, 95.10% accuracy, 94.89% F1 score, and 93.94% Gwet's kappa value. Figure 6 illustrates that the model performed excellently in classifying various types of pneumonia, with results comparable to those of the associate chief physician of the radiologist in terms of precision, accuracy, F1 scores, and Gwet's kappa values. The model and the radiologist outperformed the other models and the clinician. The classification consistency of the Mix-inResNet model (kappa = 93.09%) was similar to that of the radiologist (kappa = 93.94%), demonstrating good agreement in classification and diagnostic efficiency. Additionally, Figure 7 displays the heat map results generated by the model. 4.3 Comparison of ROC curves of four models The area under the curve (AUC) reflects the value of a diagnostic test, with a larger area closer to 1.0 indicating higher diagnostic accuracy. In the Mix-inResNet model, the AUC for bacterial, fungal, viral, tuberculous pneumonia, and normal chest CT was 0.99, 1.00, 0.99, 1.00, and 1.00, respectively (Figure 8) . The Mix-inResNet model outperformed other models by achieving the highest AUC values for fungal pneumonia, pulmonary tuberculosis, and normal chest CT samples. This indicates a superior level of accuracy in judgment and precision in classification. In the VGG16 model, the AUC for bacterial, fungal, viral, tuberculous pneumonia, and normal chest CT was 0.93, 0.91, 0.95, 0.94, and 1.00, respectively, showing lower classification recognition accuracy than all other models. Based on the figure, the AUC values of the four models mentioned are in descending order: Mix-inResNet > ResNet101 > Inception V3 > VGG16. 4.4 Comparison of confusion matrices among four models The confusion matrix compares the model's predictions with the actual classifications, providing a more intuitive way to assess the performance of four models in predicting accuracy for different categories of pneumonia. The Mix-inResNet model demonstrates a strong ability to distinguish between different types of pneumonia. It accurately identified 52 out of 59 cases of bacterial pneumonia, 35 out of 38 cases of fungal pneumonia, 57 out of 62 cases of viral pneumonia, 49 out of 50 cases of pulmonary tuberculosis, and all 77 cases of normal pneumonia. This highlights the model's vital role in accurate identification and the importance of the matrix in assessing its diagnostic capabilities ( Figure 9A) . The Inception V3 model had a few instances of misclassification among the five categories of pneumonia. Still, it provides valuable insights into the diagnostic accuracy of the model and enhances the prediction accuracy for certain pneumonia categories ( Figure 9B) . The ResNet101 model exhibited occasional misclassification among five pneumonia categories but accurately classified pulmonary tuberculosis, with only one instance misidentified as viral pneumonia (Figure 9C) . Comparatively, the VGG16 model performed poorly in classification accuracy, correctly classifying only 23 samples as fungal pneumonia and misclassifying 15 samples as other types of pneumonia (Figure 9D) . Discussion Pulmonary infections are the most common infectious diseases in respiratory medicine. They progress rapidly and result in high mortality, posing a severe threat to people's lives and health [3] . The current methods for detecting pathogens in pulmonary infections have several limitations, including low sensitivity and accuracy, time-consuming, high labor or economic costs, and the abuse of antibiotic [30, 31] . It is crucial for computed tomography (CT) in diagnosing pneumonia. Chest CT scans offer a clear view of chest structures and are highly sensitive to chest lesions. This imaging method is able to identify small and hidden lesions by observing density, morphology, and margins. While biases in interpreting imaging results and delays in reporting them by radiologists and physicians can hinder the timely identification of various pathogens in clinical practice. This often leads to delays in accurate anti-infection treatment and can worsen the patient's condition. Artificial intelligence is a complex and interdisciplinary field that has demonstrated significant potential, particularly in the realm of medical imaging within the field of medicine [7, 11] . Image datasets are commonly used to develop deep-learning diagnostic models to differentiate between various diseases in visual imaging diagnostics. The process of feature extraction was seamlessly integrated [32] . Pham [33] et al. conducted a classification study of COVID-19 pneumonia and other pneumonias using an extensive public database with 16 pre-trained Convolutional Neural Networks (CNNs) and achieved an accuracy of 0.917. Polsinelli [34] et al. proposed a lightweight SqueezeNet model to efficiently distinguish COVID-19 CT images from other community-acquired pneumonia and healthy individuals, achieving an accuracy of 0.850. While many studies have shown promising results in diagnosing pneumonia, most have been limited to COVID-19 and non-COVID-19 cases. This study presents the Mix-inResNet model, an innovative convolutional neural network (CNN) that utilizes CT scan images for multi-category image classification. This approach aims to support clinicians in simultaneously screening various types of pulmonary inflammation across large-scale samples. Our deep learning system demonstrates exceptional proficiency in differentiating among various categories of pneumonia. Additionally, our findings indicate a higher accuracy level than prior studies that employed artificial intelligence techniques for CT diagnosis [35] . We compared the model with physicians to assess sensitivity, specificity, and 95% confidence intervals. The sensitivity of the Mix-inResNet model for detecting bacterial, fungal, viral, tuberculous pneumonia, and normal CT was 88.14%, 92.11%, 91.94%, 98.00%, and 100%, respectively. The specificities were 98.68%, 100.00%, 97.77%, 97.46%, and 99.04%, respectively. These results indicate that the model can accurately distinguish between different types of pneumonia. Compared to radiologist, sensitivity and specificity were validated, demonstrating the potential for clinical application. The Mix-inResNet model demonstrates 100% specificity in diagnosing fungal pneumonia, ensuring that negative results are not misclassified. The model's advantages are further validated by a 95% confidence interval ranging from 0.88 to 0.97. Additionally, the Mix-inResNet model shows performance of 94.61% precision, 94.41% accuracy, 94.22% F1 score, and 93.09% Gwet's kappa value, compared to radiologist performance of 95.23%, 95.10%, 94.89%, and 93.94% respectively. Compared to radiologist, the AI system utilizing CT images can rapidly identify potentially pathogenic infectious pneumonias. This technology significantly enhances diagnostic accuracy, offering timely guidance on clinical treatment options and ultimately maximizing patient benefits. The artificial intelligence system exhibits real-time capabilities and high accuracy, playing a pivotal role in facilitating early treatment decisions, thereby reducing hospitalization duration and curtailing treatment expenses. Simultaneously, it effectively mitigates the risk of complications arising from delayed diagnosis and treatment due to waiting for pathogen detection [36, 37] . Experience and excellent qualifications are the main reasons for the high accuracy of imaging physicians, consistent with previous studies [36] . Additionally, we utilized the Grad-CAM technique to localize the site of pneumonia, providing more intuitive and clear results [38] . The loss function curve (yellow line) of the Mix-inResNe model exhibits the best convergence performance and has the fastest learning effect in efficiently acquiring knowledge from the training data. This model significantly dominates in terms of learning when compared to other models. Zhang [39] et al. categorized pneumonia into four types: bacterial pneumonia, fungal pneumonia, common viral pneumonia, and novel coronavirus pneumonia, based on deep learning, with AUCs of 0.989, 0.996, 0.994, and 0.997, respectively. Their findings align with our study, however, the sample size is limited, particularly in regards to fungal pneumonia (n=4), which may potentially impact the model's performance when compared to other forms of pneumonia within the test group.We included 38 cases of fungal pneumonia and demonstrated the excellent classification ability of Mix-inResNet in distinguishing between bacterial, fungal, viral, tuberculosis pneumonia, and routine chest CT scans. The area under the ROC curve (AUC) for each category was 0.99, 1.00, 0.99, 1.00, and 1.00, respectively. The Mix-inResNet model outperformed the other models with the highest AUC value, demonstrating excellent classification performance. We expanded the scope of our study by employing CT imaging to differentiate between bacterial, fungal, viral, and tuberculous pneumonia, which are not commonly found in existing literature. We conducted a comparative analysis of the confusion matrices and observed that the Mix-inResNet model exhibited remarkable proficiency in distinguishing among various pneumonia categories. This model achieved significantly fewer misclassifications when compared to the other three models evaluated. The Mix-inResNet model demonstrated strong performance in accurately identifying different pneumonia types, significantly reducing the risk of misdiagnosis. Additionally, it provides timely information on pathogens that can inform clinical decisions regarding the treatment of infections. The model's diagnostic efficacy was further confirmed by validating key performance metrics such as precision, accuracy, sensitivity, specificity, recall, and F1 score, which are crucial for reducing misdiagnosis in healthcare settings. Conclusion The Mix-inResNet model demonstrated superior accuracy and reliability, showing great potential for improving the diagnostic process and patient outcomes. The Mix-inResNet model has demonstrated significant potential in clinical settings, enhancing the accuracy and reliability of lung disease diagnostic procedures. This study indicates that the Mix-inResNet model has the potential to revolutionize AI-driven disease diagnosis, particularly in pulmonary infections. Its high accuracy and robust performance across various class configurations demonstrate its potential as a valuable tool in medical diagnosis. Declarations Acknowledgement We would appreciate all participants for their contributions to this study and thank the sequencing technical support from the Beijing Genomics Institute (Beijing, China). Author contributions BM: Data curation, Manuscript preparation. CL: Data curation, Investigation, Writing – review & editing. QC: Data curation, Resources, Sample collection. ZH: Funding acquisition, Writing – review & editing. HL: Data curation，Investigation & editing. CM: Data curation, Writing - review & editing. TL: Data curation, Formal analysis, Writing – review & editing. JF: Methodology, Resources, Writing - review & editing. AG: Project administration, Writing – review & editing. YF: Funding acquisition, Project administration, Writing – review & editing. Funding The study was supported by Natural Science Foundation of Hubei Province (Project No. JCZRYB202500657), Young Talent Project of General Hospital of Center Theater (ZZ20231693). Data availability The data used to support the findings of this study are available from the corresponding author upon request. Ethics approval and consent to participate The studies involving humans were approved by the Animal Ethical and Experimental Committee of the General Hospital of Center Theater of PLA (Wuhan, Hubei province. Permit No. 2024085-01 in accordance with their rules and regulations. The participants provided their written informed consent to participate in this study. Consent for publication Not applicable. Competing interests The authors declare that they have no competing interests. References Lei J, Wang L, Li Q, Gao L, Zhang J, Tan Y. Identification of RAGE and OSM as New Prognosis Biomarkers of Severe Pneumonia[J]. Can Respir J, 2022,2022:3854191. Global burden of 369 diseases and injuries in 204 countries and territories, 1990-2019: a systematic analysis for the Global Burden of Disease Study 2019[J]. Lancet, 2020,396(10258):1204-1222. Magill S S, Edwards J R, Bamberg W, et al. Multistate point-prevalence survey of health care-associated infections[J]. N Engl J Med, 2014,370(13):1198-1208. van Vugt S F, Verheij T J, de Jong P A, et al. Diagnosing pneumonia in patients with acute cough: clinical judgment compared to chest radiography[J]. Eur Respir J, 2013,42(4):1076-1082. Cui Y, Wang X, Lin F, et al. MiR-29a-3p Improves Acute Lung Injury by Reducing Alveolar Epithelial Cell PANoptosis[J]. Aging Dis, 2022,13(3):899-909. Hao S, Xin Q, Xiaomin Z, et al. Group membership modulates the hold-up problem: an event-related potentials and oscillations study[J]. Soc Cogn Affect Neurosci, 2023,18(1). Jiang B, Guo N, Ge Y, Zhang L, Oudkerk M, Xie X. Development and application of artificial intelligence in cardiac imaging[J]. Br J Radiol, 2020,93(1113):20190812. Harmon S A, Sanford T H, Xu S, et al. Artificial intelligence for the detection of COVID-19 pneumonia on chest CT using multinational datasets[J]. Nat Commun, 2020,11(1):4080. Ozturk T, Talo M, Yildirim E A, Baloglu UB, Yildirim O, Rajendra Acharya U. Automated detection of COVID-19 cases using deep neural networks with X-ray images[J]. Comput Biol Med, 2020,121:103792. Ji G W, Zhu F P, Xu Q, et al. Machine-learning analysis of contrast-enhanced CT radiomics predicts recurrence of hepatocellular carcinoma after resection: A multi-institutional study[J]. EBioMedicine, 2019,50:156-165. Nazir M, Shakil S, Khurshid K. Role of deep learning in brain tumor detection and classification (2015 to 2020): A review[J]. Comput Med Imaging Graph, 2021,91:101940. Ibrahim D M, Elshennawy N M, Sarhan A M. Deep-chest: Multi-classification deep learning model for diagnosing COVID-19, pneumonia, and lung cancer chest diseases[J]. Comput Biol Med, 2021,132:104348. Sun W, Zheng B, Qian W. Automatic feature learning using multichannel ROI based on deep structured algorithms for computerized lung cancer diagnosis[J]. Comput Biol Med, 2017,89:530-539. Hao S, Jiali P, Xiaomin Z, et al. Group identity modulates bidding behavior in repeated lottery contest: neural signatures from event-related potentials and electroencephalography oscillations[J]. Front Neurosci, 2023,17:1184601. Li L, Qin L, Xu Z, et al. Using Artificial Intelligence to Detect COVID-19 and Community-acquired Pneumonia Based on Pulmonary CT: Evaluation of the Diagnostic Accuracy[J]. Radiology, 2020,296(2):E65-E71. Zhang K, Liu X, Shen J, et al. Clinically Applicable AI System for Accurate Diagnosis, Quantitative Measurements, and Prognosis of COVID-19 Pneumonia Using Computed Tomography[J]. Cell, 2020,181(6):1423-1433. Kermany D S, Goldbaum M, Cai W, et al. Identifying Medical Diagnoses and Treatable Diseases by Image-Based Deep Learning[J]. Cell, 2018,172(5):1122-1131. Du W, Luo X, Chen M. A Practical Deep Learning Model in Differentiating Pneumonia-Type Lung Carcinoma from Pneumonia on CT Images: ResNet Added with Attention Mechanism[J]. J Oncol, 2022,2022:8906259. Huang J, Jiang E, Yang D, et al. Metagenomic Next-Generation Sequencing versus Traditional Pathogen Detection in the Diagnosis of Peripheral Pulmonary Infectious Lesions[J]. Infect Drug Resist, 2020,13:567-576. Li H, Gao H, Meng H, et al. Detection of Pulmonary Infectious Pathogens From Lung Biopsy Tissues by Metagenomic Next-Generation Sequencing[J]. Front Cell Infect Microbiol, 2018,8:205. Miao Q, Ma Y, Wang Q, et al. Microbiological Diagnostic Performance of Metagenomic Next-generation Sequencing When Applied to Clinical Practice[J]. Clin Infect Dis, 2018,67(suppl_2):S231-S240. De Pauw B, Walsh T J, Donnelly J P, et al. Revised definitions of invasive fungal disease from the European Organization for Research and Treatment of Cancer/Invasive Fungal Infections Cooperative Group and the National Institute of Allergy and Infectious Diseases Mycoses Study Group (EORTC/MSG) Consensus Group[J]. Clin Infect Dis, 2008,46(12):1813-1821. Kunihiro Y, Tanaka N, Kawano R, et al. Differential diagnosis of pulmonary infections in immunocompromised patients using high-resolution computed tomography[J]. Eur Radiol, 2019,29(11):6089-6099. Cilloniz C, Martin-Loeches I, Garcia-Vidal C, San Jose A, Torres A. Microbial Etiology of Pneumonia: Epidemiology, Diagnosis and Resistance Patterns[J]. Int J Mol Sci, 2016,17(12). Rea G, Sperandeo M, Lieto R, et al. Chest Imaging in the Diagnosis and Management of Pulmonary Tuberculosis: The Complementary Role of Thoraci Ultrasound[J]. Front Med (Lausanne), 2021,8:753821. Szegedy C, Liu W, Jia Y, et al. Going deeper with convolutions[C], Boston, MA, 2015. He K, Zhang X, Ren S, et al. Deep Residual Learning for Image Recognition[C], Las Vegas, 2016. Szegedy C, Ioffe S, Vanhoucke V, et al. Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning[C], San Francisco, 2017. Szegedy C, Vanhoucke V, Ioffe S, et al. Rethinking the Inception Architecture for Computer Vision[C], Las Vegas, 2016. Dellinger R P, Levy M M, Rhodes A, et al. Surviving Sepsis Campaign: international guidelines for management of severe sepsis and septic shock, 2012[J]. Intensive Care Med, 2013,39(2):165-228. Chaudhuri A, Martinez-Martin P, Kennedy P G, et al. EFNS guideline on the management of community-acquired bacterial meningitis: report of an EFNS Task Force on acute bacterial meningitis in older children and adults[J]. Eur J Neurol, 2008,15(7):649-659. Li H, Wang C. MIS-Net: A deep learning-based multi-class segmentation model for CT images[J]. PLoS One, 2024,19(3):e0299970. Pham T D. A comprehensive study on classification of COVID-19 on computed tomography with pretrained convolutional neural networks[J]. Sci Rep, 2020,10(1):16942. Polsinelli M, Cinque L, Placidi G. A light CNN for detecting COVID-19 from CT scans of the chest[J]. Pattern Recognit Lett, 2020,140:95-100. Chumbita M, Cilloniz C, Puerta-Alcalde P, et al. Can Artificial Intelligence Improve the Management of Pneumonia[J]. J Clin Med, 2020,9(1). Wang F, Li X, Wen R, et al. Pneumonia-Plus: a deep learning model for the classification of bacterial, fungal, and viral pneumonia based on CT tomography[J]. Eur Radiol, 2023,33(12):8869-8878. Chung M, Bernheim A, Mei X, et al. CT Imaging Features of 2019 Novel Coronavirus (2019-nCoV)[J]. Radiology, 2020,295(1):202-207. Selvaraju R R, Cogswell M, Das A, et al. Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization[C], Venice, 2017. Zhang Y H, Hu X F, Ma J C, et al. Clinical Applicable AI System Based on Deep Learning Algorithm for Differentiation of Pulmonary Infectious Disease[J]. Front Med (Lausanne), 2021,8:753055. Additional Declarations No competing interests reported. Supplementary Files FIgureS1.jpg supplementarymaterials.docx Cite Share Download PDF Status: Under Revision Version 1 posted Editorial decision: Revision requested 13 May, 2026 Reviews received at journal 08 May, 2026 Reviews received at journal 05 May, 2026 Reviewers agreed at journal 29 Apr, 2026 Reviewers agreed at journal 27 Apr, 2026 Reviews received at journal 26 May, 2025 Reviewers agreed at journal 26 May, 2025 Reviewers invited by journal 05 May, 2025 Editor invited by journal 09 Apr, 2025 Editor assigned by journal 08 Apr, 2025 Submission checks completed at journal 08 Apr, 2025 First submitted to journal 27 Mar, 2025 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-6323914","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":452241358,"identity":"40a6b88c-808e-4594-adf8-c5da63b66224","order_by":0,"name":"Beibei Meng","email":"","orcid":"","institution":"Department of Respiratory and Critical Care Medicine, General Hospital of Center Theater of PLA, Wuhan 430070, China","correspondingAuthor":false,"prefix":"","firstName":"Beibei","middleName":"","lastName":"Meng","suffix":""},{"id":452241359,"identity":"f17a8518-3e19-499a-8993-ad51969f5f3d","order_by":1,"name":"Changyu Lei","email":"","orcid":"","institution":"Department of Orthopedics, General Hospital of Center Theater of PLA, Wuhan 430070, China","correspondingAuthor":false,"prefix":"","firstName":"Changyu","middleName":"","lastName":"Lei","suffix":""},{"id":452241360,"identity":"a836d241-443c-46b8-a216-e6c89fb7401c","order_by":2,"name":"Qingjia Chi","email":"","orcid":"","institution":"School of Physics and Mechanics, Wuhan University of Technology, Wuhan, Hubei, 430070, China","correspondingAuthor":false,"prefix":"","firstName":"Qingjia","middleName":"","lastName":"Chi","suffix":""},{"id":452241361,"identity":"15f95e46-68a8-4246-933c-3f321ce2a11b","order_by":3,"name":"Zhenhong Hu","email":"","orcid":"","institution":"Department of Respiratory and Critical Care Medicine, General Hospital of Center Theater of PLA, Wuhan 430070, China","correspondingAuthor":false,"prefix":"","firstName":"Zhenhong","middleName":"","lastName":"Hu","suffix":""},{"id":452241362,"identity":"c6f0a128-005c-4428-8d89-14b9f5601dd4","order_by":4,"name":"Haichao Liu","email":"","orcid":"","institution":"Department of Respiratory and Critical Care Medicine, General Hospital of Center Theater of PLA, Wuhan 430070, China","correspondingAuthor":false,"prefix":"","firstName":"Haichao","middleName":"","lastName":"Liu","suffix":""},{"id":452241363,"identity":"94210567-ae7d-44bb-b07f-e93c444add97","order_by":5,"name":"Congzheng Mao","email":"","orcid":"","institution":"Department of Respiratory and Critical Care Medicine, General Hospital of Center Theater of PLA, Wuhan 430070, China","correspondingAuthor":false,"prefix":"","firstName":"Congzheng","middleName":"","lastName":"Mao","suffix":""},{"id":452241364,"identity":"c5c7fa8f-3e5d-4e47-ab5e-89c23c0072ae","order_by":6,"name":"Ai Guoping","email":"","orcid":"","institution":"Department of Radiology, General Hospital of Center Theater of PLA, Wuhan 430070, China","correspondingAuthor":false,"prefix":"","firstName":"Ai","middleName":"","lastName":"Guoping","suffix":""},{"id":452241365,"identity":"2c2fb244-d5fb-4521-847f-6f59afb9a589","order_by":7,"name":"Yao Fang","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAtklEQVRIiWNgGAWjYBACPgbmhg9AWo6BmVgtbAyMjTOAtDHpWhIbiHYYG3tjY8PPHbXp89t5D35gqLlDhBaeg42NvWeO5244zJcswXDsGRFaJBLbH/C2HcvdwMxjIMHYcJgoLY2Nf9uOpcs38xj/IFpLM29bTQLDYR4zIm0B+qVZtu2A4QagFouEY0Ro4WdvPtj4tq1OXr7/jPGNDzVEaIECqMoEojUwMNSRoHYUjIJRMApGHAAAlpA5fRwjJo8AAAAASUVORK5CYII=","orcid":"","institution":"Department of Respiratory and Critical Care Medicine, General Hospital of Center Theater of PLA, Wuhan 430070, China","correspondingAuthor":true,"prefix":"","firstName":"Yao","middleName":"","lastName":"Fang","suffix":""}],"badges":[],"createdAt":"2025-03-28 01:38:17","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-6323914/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6323914/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":82561045,"identity":"2bb61058-c108-4ad1-bce5-12d729d00f1d","added_by":"auto","created_at":"2025-05-13 01:38:02","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":3668912,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eComparison of two detection methods\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"Figure1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6323914/v1/aaa032f5052889f286b6ad99.jpg"},{"id":82561043,"identity":"ac58960a-31ed-4308-8428-61c809352e64","added_by":"auto","created_at":"2025-05-13 01:38:02","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":23988,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eCharacterization of data distribution in training, validation and test sets\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"Figure2.png","url":"https://assets-eu.researchsquare.com/files/rs-6323914/v1/5ba66546c5c98c213c4b75fd.png"},{"id":82561042,"identity":"e77fa0b4-42bc-4cc7-bdaa-014d366f8a10","added_by":"auto","created_at":"2025-05-13 01:38:02","extension":"jpg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":3291112,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eThe bioinformatics analysis workflow of mNGS\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"Figure3.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6323914/v1/251d6fbf87bae7a3ac752fdc.jpg"},{"id":82559787,"identity":"b7acdfe4-dec0-4bc6-97b2-b192e964e7e2","added_by":"auto","created_at":"2025-05-13 01:30:02","extension":"jpg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":1187206,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eChest CT of four typical pneumonias\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNote: Red arrows point to foci of infection.\u003c/p\u003e","description":"","filename":"Figure4.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6323914/v1/333b8c331f7b095d5bbd7cb2.jpg"},{"id":82559809,"identity":"e078d942-223c-4b71-8aba-e182a8c0515e","added_by":"auto","created_at":"2025-05-13 01:30:03","extension":"jpg","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":843557,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eComparison of loss function among four models\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"Figure5.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6323914/v1/afe4082e50953f8cf62b3dcc.jpg"},{"id":82559772,"identity":"5a5dcf57-d905-4c5f-9f2a-508016716723","added_by":"auto","created_at":"2025-05-13 01:30:02","extension":"jpg","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":1156328,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eComparison of different models with physician precision, F1 score, accuracy, and consistency\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"Figure6.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6323914/v1/c4f23d813963def1a984e778.jpg"},{"id":82559773,"identity":"1e1a67d0-3fac-4460-bab4-2da921214723","added_by":"auto","created_at":"2025-05-13 01:30:02","extension":"jpg","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":2052350,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eGrad-CAM technique for detecting pneumonia Heatmap\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNote: The color bar on the right panel indicates the strength of attention, with darker red indicating the strongest attention and darker blue indicating weaker attention\u003c/p\u003e","description":"","filename":"Figurre7.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6323914/v1/c5263596d2ac1e523b6af71b.jpg"},{"id":82561050,"identity":"2e1ce9c3-6ba3-4e2f-8db6-38110a48a4c3","added_by":"auto","created_at":"2025-05-13 01:38:03","extension":"jpg","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":1314719,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eROC curve\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"Figure8.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6323914/v1/d8139494a45b147eee641e6f.jpg"},{"id":82561048,"identity":"664a303b-b252-49fb-8fe6-c2077be1dda4","added_by":"auto","created_at":"2025-05-13 01:38:03","extension":"jpg","order_by":9,"title":"Figure 9","display":"","copyAsset":false,"role":"figure","size":1389035,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eConfusion matrix\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNote: The discriminative abilities of Mix-inResNet, Inception V3, ResNet101, and VGG16 models in distinguishing between four different types of pneumonia and normal chest CT scans are denoted by A, B, C, and D respectively.\u003c/p\u003e","description":"","filename":"Figure9.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6323914/v1/2b2da24ff9b6976808692ba9.jpg"},{"id":82562581,"identity":"ffe4a3e0-d199-455d-909a-dbeaea909a39","added_by":"auto","created_at":"2025-05-13 01:46:12","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":16074831,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6323914/v1/02868086-d5a4-45c4-b6e2-8df39db05312.pdf"},{"id":82559777,"identity":"c149685a-9a27-4cf0-af8d-8498f424b835","added_by":"auto","created_at":"2025-05-13 01:30:02","extension":"jpg","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":3318095,"visible":true,"origin":"","legend":"","description":"","filename":"FIgureS1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6323914/v1/04f4fff7d1cfde3013190451.jpg"},{"id":82559775,"identity":"59fd7a09-a2f6-4771-9ec0-18218e78d262","added_by":"auto","created_at":"2025-05-13 01:30:02","extension":"docx","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":16498,"visible":true,"origin":"","legend":"","description":"","filename":"supplementarymaterials.docx","url":"https://assets-eu.researchsquare.com/files/rs-6323914/v1/cb9bdff94a986395b632675b.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Application of deep learning to automated diagnosis of computed tomography images of pulmonary infections","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003ePulmonary infections are increasing in morbidity and mortality worldwide\u003csup\u003e[1]\u003c/sup\u003e. According to the World Health Organization\u0026apos;s (WHO) 2019 report on the world\u0026apos;s leading causes of death, pulmonary infections were ranked fourth, causing 2.6 million deaths and affecting 489 million people globally. Additionally, it is in the top 10 causes of death from infectious diseases in countries of different income levels\u003csup\u003e[2, 3]\u003c/sup\u003e. Pulmonary infections can be caused by many different pathogens, each with its clinical characteristics. It is crucial to quickly and accurately identify these pathogens to make an accurate treatment decisions. However, traditional diagnostic methods heavily rely on physician experience and subjective judgment, which can limit the accuracy and efficiency of diagnosis. Previous studies have found that general practitioners had a sensitivity of 29% and a specificity of 99% upon pneumonia\u003csup\u003e[4]\u003c/sup\u003e. The current diagnostic process for pulmonary infections has significant limitations regarding timeliness and affordability. Therefore, there is an urgent need for a highly sensitive diagnostic and detection system to assist the traditional methods, which are often time-consuming and expensive.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eAutomated disease diagnostics based on Artificial Intelligence (AI) have shown significant growth potential. The demand of sensitive diagnostic and detection system has triggered a boom in AI technology in healthcare and demonstrated a wide range of potential applications\u003csup\u003e[5, 6]\u003c/sup\u003e. AI technology has profoundly transformed the medical and healthcare sectors, particularly in the areas of drug discovery and medical imaging. Its accuracy in medical imaging highlights its potential to drive further advancements in the field\u003csup\u003e[7]\u003c/sup\u003e. Machine Learning (ML) and Deep Learning (DL), as two key technologies embodying AI, have been tested in disease diagnosis\u003csup\u003e[8]\u003c/sup\u003e, lesion detection\u003csup\u003e[9]\u003c/sup\u003e, prognosis analysis\u003csup\u003e[10]\u003c/sup\u003e, and other medical fields. In certain automated medical image recognition and detection tasks, deep learning models have shown detection levels comparable to radiologists\u003csup\u003e[7, 11]\u003c/sup\u003e. Given the impressive performance and established effectiveness of deep learning in image recognition and classification tasks, numerous researchers have explored integrating this technology into CT image analysis to accurately classify and predict various\u0026nbsp;pulmonary\u0026nbsp;diseases\u003csup\u003e[12-14]\u003c/sup\u003e. CT imaging is valuable\u0026nbsp;for diagnosing pneumonia. With the help of AI, we can create CT image-based tools to accurately classify pneumonia. The 2019 coronavirus disease pandemic has highlighted the need for alternative methods due to the misclassification of similar respiratory disease symptoms. In contrast, there have been AI studies in chest imaging, such as CT image-based classification of new coronavirus and common pneumonia\u003csup\u003e[15, 16]\u003c/sup\u003e, the\u0026nbsp;classification of pediatric bacterial pneumonia and viral pneumonia based on chest radiographs\u003csup\u003e[17]\u003c/sup\u003e, and the identification of pneumonic lung cancer and pneumonia based on chest CT images\u003csup\u003e[18]\u003c/sup\u003e. A limited number of studies are available on identifying pneumonia caused by various pathogen infections using chest CT images. Therefore, this study proposes a convolutional neural network (CNN)-based model for the detection of various types of pulmonary diseases, aiming to enhance the recognition accuracy in pulmonary infection diagnosis. The performance of CT scan datasets is analyzed and compared with conventional neural networks and clinical experts to improve diagnose accuracy, based on pathogen confirmation by mNGS and traditional culture. CT imaging can detect abnormal changes in the chest before clinical manifestation, making early and timely critical diagnosis.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e1. Comparison of flowcharts between traditional culture and the Mix-inResNet model.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eFig. 1 shows a comparison of traditional traditional culture methods and deep learning methods in imaging pathogen diagnosis and their advantages and disadvantages.\u003c/p\u003e"},{"header":"2. Methods and materials","content":"\u003cp\u003e\u003cstrong\u003e2.1 Data sets\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e2.1.1 Data distribution\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eIn this study, we collected data from 273 patients (152 males and 121 females) whose ages ranged from 42 to 69, and who underwent chest CT scans. Their alveolar lavage fluid (BALF) was sent for metagenomic next generation sequencing (mNGS) and conventional cultures, which indicated single pathogen infections. The data was collected from March 2021 to April 2024 at our hospital.\u0026nbsp;Following manual screening, we compiled an imaging dataset comprising 2858 images featuring obvious infection types, including 589 cases of bacterial pneumonia, 380 cases of fungal pneumonia, 619 cases of viral pneumonia, 503 cases of pulmonary tuberculosis, and 767 images of normal conditions. These images were divided into training, validation, and testing sets in an 8:1:1 ratio, as illustrated in Figure 2.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e2.1.2 Diagnosis of pathogens\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eBronchoalveolar lavage fluid (BALF\u003c/strong\u003e\u003cstrong\u003e）\u003c/strong\u003e\u003cstrong\u003e:\u003c/strong\u003eBefore surgery, conduct comprehensive preoperative examinations to exclude significant cardiovascular and cerebrovascular diseases, as well as a history of anticoagulant medication usage. Instruct the patient to observe a water fast for a duration of 4 hours prior to the surgical procedure. Select the specific lung segment for bronchoalveolar lavage based on chest CT imaging, ensuring an adequate sample volume of no less than 5mL. Subsequently, perform both conventional laboratory culture and high-throughput sequencing techniques on the collected samples.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eSputum specimen:\u0026nbsp;\u003c/strong\u003eA sputum specimen was collected and sent to the Laboratory Department of the hospital. It was inoculated into the appropriate medium, commonly used LB agar culture plate and blood agar culture plate, and cultured for 24-72 hours. Positive colonies were selected and tested using the M\u0026eacute;rieux biochemical identifier from France. Based on the results of the biochemical identification, a culture identification report was issued. The clinician will then consider the clinical manifestations of the case, exclude the possibility of background or contaminating bacteria, and decide whether the identified pathogen is the cause of the pulmonary infection.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMetagenomic next generation sequenc\u003c/strong\u003e\u003cstrong\u003eing (mNGS):\u0026nbsp;\u003c/strong\u003eA high-throughput sequencing method that operates independently of traditional microbial culture. This approach does not necessitate preconceived assumptions regarding potential pathogenic microorganism. It is characterized by its unbiased nature, broad coverage, rapid detection capabilities, high sensitivity, and accuracy\u003csup\u003e[19]\u003c/sup\u003e. The interpretation criteria for a positive mNGS test include the following four items\u003csup\u003e[20, 21]\u003c/sup\u003e (1) mNGS and conventional culture identified the same microorganism, and the reads number of mNGS of a single species is greater than 50; (2) Mycobacterium tuberculosis sequence number at the species or genus level is judged positive when at least 1 sequence number (reads) is consistent with the reference genome; (3) The relative abundance of the bacteria or fungi at the genus level is \u0026gt;30%; (4) The coverage of microorganisms is 10 times higher than any other bacteria and 5 times higher than any other fungi, as illustrated in Figure 3.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e2.1.3 Exclusion criteria:\u0026nbsp;\u003c/strong\u003eInclusion criteria: 1. Patients diagnosed with pulmonary infection and confirmed by BLAF examination with mNGS or traditional culture showing single pathogen infection; 2. Undergo chest CT examination during hospitalization. Exclusion criteria: 1. Poor image quality, heavy artifacts\u0026nbsp;affecting the observation results; 2. Patients with severe pulmonary comorbidities; 3. History of chest surgery; 4. Incomplete clinical data.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e2.2 Image annotation method\u003c/strong\u003e\u003cstrong\u003e:\u003c/strong\u003e Image identification and grouping were conducted by a trained and qualified clinician specialized in image recognition, along with the deputy chief physician of the imaging department. In contrast, the process was carried out without considering the patient\u0026apos;s clinical information and other imaging data. The anonymized dataset was taken with RadiAnt DICOM Viewer 2023.1.0 software, WL: -550 HU, WW: 1350 HU, image size of 512*512, resolution of 96 dpi, and the save format was selected as DICOM mode, and the anonymized dataset was categorized separately: bacterial, fungal, viral, tuberculous pneumonia, and normal CT image.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e2.3\u003c/strong\u003e\u003cstrong\u003e\u0026nbsp;CT Scanning Parameters and Typical Chest CTs\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e2.3.1 CT parameters\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eCT examination was performed using a Toshiba Aquilion 16-row spiral CT machine. Scanning parameters: tube voltage of 120 kV, tube current of 250 mA, scanning collimation of 1.0 mm \u0026times; 16, pitch of 1.3, rotation time of 0.5 s/r, field of view of 500 mm \u0026times; 500 mm, acquisition matrix of 512 \u0026times; 512, CT scanning was completed, and the images were uploaded to the CT workstation to complete the reconstruction of the 2-mm thin-layer\u0026nbsp;pulmonary\u0026nbsp;CT images using volumetric data. Image Acquisition: The patient was placed in the supine position. The patient was placed in the supine position and was instructed to hold his/her breath for 10 s. All scans were performed at the end of deep inhalation, covering the range from the rib-diaphragm angle at the base of the lungs to the thoracic inlet. Standard lung window (window position -550 HU, window width 1350 HU) and mediastinal window (window position 40 HU, window width 350 HU) were used for observation.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e2.3.2 Typical Chest CT\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eBacterial pneumonia often presents with thickening and blurring of the lung texture, patchy consolidation, aerated bronchial signs, associated cavities, and ground-glass shadows\u003csup\u003e[22]\u003c/sup\u003e. Cavitation, air crescent signs, haloing, and\u0026nbsp;ring-like appearance often characterize fungal pneumonia\u003csup\u003e[23]\u003c/sup\u003e. Typical signs of viral pneumonia are increased lung texture, ground-glass shadows, small patchy or extensive infiltrates, and\u0026nbsp;solid changes. In severe cases, diffuse nodular infiltrates in both lungs, but lobar solid changes and pleural effusions are uncommon\u003csup\u003e[24]\u003c/sup\u003e. Common indicators of pulmonary tuberculosis include lesions in the upper lobe\u0026apos;s postapical segment, the lower lobe\u0026apos;s dorsal segment, and the posterior basal segment. These lesions display various forms, such as infiltration, proliferation, and the presence of caseous and fibrocalcified lesions, with the features of uneven density, clearer margins, and chronic changes. They also tend to form cavities and spread\u003csup\u003e[25]\u003c/sup\u003e,\u0026nbsp;as illustrated in Figure 4.\u003c/p\u003e\n\u003cp\u003eNote: Red arrows point to foci of infection.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e2.4 Overview of the model structure\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eMix-inResNet comprises a backbone network, a basic convolution module (BlockA-C), a dimensionality reduction module (ReductionA-C), a self-attention module, an average pooling layer, a dropout layer, and a softmax layer. The detailed design structure is shown in the attached material. In the backbone network, hybrid convolution is used to fuse features processed by depth separable convolution, dilation convolution and standard convolution. The aim is to achieve a deeper, broader, and more abstract level by integrating features of different scopes and levels of abstraction.\u0026nbsp;The network design utilizes the Inception network\u0026apos;s features to simultaneously perform convolution and pooling operations at multiple scales and then combine the results in parallel. This multi-scale parallel strategy aims to capture and integrate image features at different scales and orientations. In Block A-C, a ResNet feature is introduced that uses \u0026quot;residual concatenation\u0026quot; to address the gradient vanishing problem and enable the training of deeper hierarchical structures, thus improving performance. Please refer to the related paper for more information about Inception and ResNet.\u003csup\u003e[26-28]\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e2.5 Image preprocessing\u003c/strong\u003e\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe images were classified using various neural networks (Mix-inResNet, ResNet 101, Inception V3, and VGG 16), and their classification results were compared and evaluated. All input images were normalized to 256 \u0026times; 256 pixels. To improve the generalization ability and robustness of the network training, sophisticated data augmentation techniques such as specular reflection, random rotation, random cropping, and gaussian noise injection were applied to the input images for all classification networks. The hidden layer weights were randomly initialized at the beginning of training and optimized using the SDG optimizer with a custom learning rate scheduler. The initial learning rate was set to 1e-4, and momentum was set to 0.9 and 0.0001, respectively. The network was trained for 300 epochs with a batch size of 8. All networks were implemented using PyTorch, and the code was run on a server equipped with a RTX 4080 GPU. All network operations were performed using PyTorch (\u003cem\u003ehttps://pytorch.org\u003c/em\u003e).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e2.6 Model training\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eIn model training, the loss function used is Cross-Entropy Loss (CEL). An Auxiliary Classifier is also incorporated to provide extra gradient signals, enhancing feature extraction capability\u003csup\u003e[29]\u003c/sup\u003e. The modeling process is also improved by introducing an Auxiliary Classifier to provide additional gradient signals to enhance feature extraction. This study describes the loss function in the model and its formulas as follows.\u003c/p\u003e\n\u003cp\u003e\u003cimg src=\"data:image/png;base64,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\" width=\"566\" height=\"183\"\u003e\u003c/p\u003e\n\u003cp\u003eWhere: in Equation 1, N is the number of samples, and\u003cimg width=\"11\" height=\"35\" src=\"data:image/png;base64,R0lGODlhCwAjAHcAMSH+GlNvZnR3YXJlOiBNaWNyb3NvZnQgT2ZmaWNlACH5BAEAAAAALAAABgAKAAoAhAAAAAAAAAAAOgAAZgA6kDoAOjoAZjo6OjqQ22YAOmY6kGaQ22a2/5A6AJA6OpA6ZpDbtpDb/7ZmALb/27b//9uQOtv///+2Zv/bkP//tv//2wECAwECAwECAwECAwECAwU7YBMQVhUMFABgRQRoDaOqEgJkijUDF1nJGocrk4DkdqrKwQaoCFwrQ+olVGGY1Mhj8phmGwLZ7BJYhAAAOw==\" alt=\"image\"\u003e\u0026nbsp;is the actual label.\u003cimg width=\"11\" height=\"35\" src=\"data:image/png;base64,R0lGODlhCwAjAHcAMSH+GlNvZnR3YXJlOiBNaWNyb3NvZnQgT2ZmaWNlACH5BAEAAAAALAAAAgAKAA4AhQAAAAAAAAAAOgAAZgA6kDoAOjoAZjo6OjpmtjqQ22YAAGYAOmY6OmY6kGaQ22a2/5A6AJA6OpA6ZpDbtpDb25Db/7ZmALZmZrbb/7b/27b//9uQOtu2Ztv///+2Zv/bkP//tv//2wECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwZOQABIQegAjkcQg7JBII8XzJHjeFqtkEBxExhojp9CBRCCPJ6WhLBhRHq251BkDFpM2NfNQQ3YCMYAHwZfZHNgfIUVEhkShIkQAmduAQ5BADs=\" alt=\"image\"\u003e\u0026nbsp;is the probability distribution predicted by the model. In Equation 2\u003cimg width=\"29\" height=\"35\" src=\"data:image/png;base64,R0lGODlhHQAjAHcAMSH+GlNvZnR3YXJlOiBNaWNyb3NvZnQgT2ZmaWNlACH5BAEAAAAALAAAAwAdAA0AhQAAAAAAAAAAOgAAZgA6OgA6ZgA6kABmkABmtjoAADoAOjpmtjqQ22YAAGYAZmY6AGY6ZmY6kGZmOmZmZmaQkGa222a2/5A6AJA6ZpBmZpCQZpDb/7ZmALZmOraQOracZrbb27bb/7b//9uQOtu2kNv/29v///+2Zv/bkP/btv//tv//2wECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwaGQIAQMAoUNsOkcrk8GUyjBXNK5TCoWKaqYcl6hygF8ut1msheKxqgcoi0kPd6mnoECBXAKvM5XAQaAU9bAQwqESUXfgJIGSEAJHYFIH8bIwgrGCIjFigLHE9hll1TnkQWmSIrFwEIhyamnFSxqJodG56vtFOrARQJAhKCHgHFxRWsEwEDIkEAOw==\" alt=\"image\"\u003e\u0026nbsp;is the cross-entropy loss of the primary classifier, and\u003cimg width=\"45\" height=\"37\" src=\"data:image/png;base64,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\" alt=\"image\"\u003e\u0026nbsp;is the cross-entropy loss of the auxiliary classifier, both of which use the same cross-entropy loss formula in Equation 1, and\u003cimg width=\"11\" height=\"35\" src=\"data:image/png;base64,R0lGODlhCwAjAHcAMSH+GlNvZnR3YXJlOiBNaWNyb3NvZnQgT2ZmaWNlACH5BAEAAAAALAMABgAIAAcAhAAAAAAAAAAAOgA6ZgBmtjoAADoAZjpmZjqQ22YAAGYAZmY6AGZmAGa2/5A6AJA6OpDb/7ZmkLb/27b//9uQOtu2tv+2Zv/bkP//tv//2wECAwECAwECAwECAwECAwECAwUmIIAtwSA5SIYC1KFMlzGJCQHEc/ZAt3wXDVojE2EgKgBLQQBRIkIAOw==\" alt=\"image\"\u003e\u0026nbsp;is the weight coefficient of the auxiliary loss.\u003c/p\u003e\n\u003cp\u003eWe have included the loss function curves to further demonstrate the performance of the models. As depicted in \u003cstrong\u003eFigure 5,\u0026nbsp;\u003c/strong\u003eduring the 0th-20th Epoch, the model rapidly converges and learns the features. The model gradually decreases from the 20th to the 170th Epoch, indicating a progressive fit to the features. After 170 Epochs, the model stabilizes, suggesting it is approaching the minimum loss point. The Mix-inResNe model, represented by the yellow line, demonstrates the best convergence performance and the fastest learning results among these models.\u003c/p\u003e"},{"header":"3. Statistical methods","content":"\u003cp\u003eWe conducted statistical and deep learning analysis using Python (version 3.8.0). We utilized the Gradient-weighted Class Activation Mapping (Grad-CAM) technique to visually identify the infection sites of different pathogens. Data processing was performed using SPSS 26.0 software. Normally distributed measures were presented as mean\u0026thinsp;\u0026plusmn;\u0026thinsp;standard deviation, while non-normally distributed data were described as median (M) and interquartile range (P25, P75). To evaluate the diagnostic performance of the Mix-inResNet model in manual classification, we utilized quantitative assessment metrics, including accuracy, precision, sensitivity, specificity, F1 score, confusion matrix, and ROC. Additionally, we assessed the inter-reader and clinical reliability using Gwet's kappa values and their 95% confidence intervals.\u003c/p\u003e"},{"header":"4. Results","content":"\u003cp\u003e\u003cstrong\u003e4.1 Comparison of the model with physicians in terms of\u0026nbsp;\u003c/strong\u003e\u003cstrong\u003esensitivity, specificity\u0026nbsp;\u003c/strong\u003e\u003cstrong\u003eand 95% confidence intervals\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eBased on the findings in \u003cstrong\u003eTable 1\u003c/strong\u003e, the Mix-inResNet model demonstrated high sensitivity in detecting bacterial, fungal, viral, tuberculous pneumonia, and normal chest CT, with percentages of 88.14%, 92.11%, 91.94%, 98.00%, and 100.00%, respectively. Additionally, the model exhibited high specificity, with percentages of 98.68%, 100.00%, 97.77%, 97.46%, and 99.04%, respectively. Compared to other models, the Mix-inResNet model showed significant advantages in\u0026nbsp;sensitivity and specificity. Notably, it achieved 100% specificity in fungal detection, indicating a low misdiagnosis rate and enhancing its clinical applicability. The results indicate that the Mix-inResNet model outperformed other models (such as Inception V3, ResNet101, VGG16) and even demonstrated comparable performance to radiologist, with confidence intervals of 0.88-0.97.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 1\u0026nbsp;\u003c/strong\u003e\u003cstrong\u003eComparison of\u0026nbsp;\u003c/strong\u003e\u003cstrong\u003esensitivity, specificity\u0026nbsp;\u003c/strong\u003e\u003cstrong\u003eand 95% confidence intervals\u0026nbsp;\u003c/strong\u003e\u003cstrong\u003eamong models for different pneumonia types and normal chest CT scans.\u003c/strong\u003e\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"99%\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 19px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"5\" valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003eSensitivity\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 19px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 16px;\"\u003e\n \u003cp\u003eBacterial pneumonia\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 16px;\"\u003e\n \u003cp\u003eFungal pneumonia\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003eViral pneumonia\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003e\u0026nbsp;Pulmonary tuberculosis\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003eNormal\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 19px;\"\u003e\n \u003cp\u003eMix-inResNet\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 16px;\"\u003e\n \u003cp\u003e88.14%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 16px;\"\u003e\n \u003cp\u003e\u0026nbsp;92.11%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003e91.94%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003e98.00%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003e100.00%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 19px;\"\u003e\n \u003cp\u003eInception V3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 16px;\"\u003e\n \u003cp\u003e83.05%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 16px;\"\u003e\n \u003cp\u003e63.16%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003e90.32%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003e86.00%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003e100.00%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 19px;\"\u003e\n \u003cp\u003eResNet101\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 16px;\"\u003e\n \u003cp\u003e77.97%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 16px;\"\u003e\n \u003cp\u003e86.84%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003e93.55%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003e98.00%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003e100.00%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 19px;\"\u003e\n \u003cp\u003eVGG16\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 16px;\"\u003e\n \u003cp\u003e79.66%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 16px;\"\u003e\n \u003cp\u003e60.53%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003e83.87%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003e82.00%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003e100.00%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 19px;\"\u003e\n \u003cp\u003eClinician\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 16px;\"\u003e\n \u003cp\u003e88.14%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 16px;\"\u003e\n \u003cp\u003e73.68%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003e91.94%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003e92.00%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003e100.00%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 19px;\"\u003e\n \u003cp\u003eRadiologist\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 16px;\"\u003e\n \u003cp\u003e89.83%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 16px;\"\u003e\n \u003cp\u003e92.11%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003e93.55%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003e98.00%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003e100.00%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 19px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"5\" valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003eSpecificity\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 19px;\"\u003e\n \u003cp\u003eMix-inResNet\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 16px;\"\u003e\n \u003cp\u003e98.68%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 16px;\"\u003e\n \u003cp\u003e100.00%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003e97.77%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003e97.46%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003e99.04%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 19px;\"\u003e\n \u003cp\u003eInception V3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 16px;\"\u003e\n \u003cp\u003e92.51%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 16px;\"\u003e\n \u003cp\u003e99.19%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cpre\u003e97.32%\u003c/pre\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003e96.61%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003e98.09%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 19px;\"\u003e\n \u003cp\u003eResNet101\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 16px;\"\u003e\n \u003cp\u003e98.68%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 16px;\"\u003e\n \u003cp\u003e97.98%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003e97.32%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003e96.61%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003e99.52%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 19px;\"\u003e\n \u003cp\u003eVGG16\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 16px;\"\u003e\n \u003cp\u003e95.15%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 16px;\"\u003e\n \u003cp\u003e96.77%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003e92.86%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003e95.76%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003e99.52%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 19px;\"\u003e\n \u003cp\u003eClinician\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 16px;\"\u003e\n \u003cp\u003e94.27%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 16px;\"\u003e\n \u003cp\u003e100.00%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003e99.11%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003e96.61%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003e98.56%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 19px;\"\u003e\n \u003cp\u003eRadiologist\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 16px;\"\u003e\n \u003cp\u003e98.68%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 16px;\"\u003e\n \u003cp\u003e100.00%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003e\u0026nbsp;97.77%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003e9 7.88%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003e99.52%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 19px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"5\" valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003e95% Confidence Interval\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 19px;\"\u003e\n \u003cp\u003eMix-inResNet\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"5\" valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003e0.88-0.97\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 19px;\"\u003e\n \u003cp\u003eInception V3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"5\" valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003e0.80-0.88\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 19px;\"\u003e\n \u003cp\u003eResNet101\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"5\" valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003e0.86-0.95\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 19px;\"\u003e\n \u003cp\u003eVGG16\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"5\" valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003e0.76-0.85\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 19px;\"\u003e\n \u003cp\u003eClinician\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"5\" valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003e0.84-0.93\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 19px;\"\u003e\n \u003cp\u003eRadiologist\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"5\" valign=\"top\" style=\"width: 80px;\"\u003e\n \u003cp\u003e0.89-0.98\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u003cstrong\u003e4.2\u0026nbsp;\u003c/strong\u003e\u003cstrong\u003eComparison of\u0026nbsp;\u003c/strong\u003e\u003cstrong\u003eprecision, accuracy, F1 score and\u0026nbsp;\u003c/strong\u003e\u003cstrong\u003eGwet\u0026apos;s kappa\u0026nbsp;\u003c/strong\u003e\u003cstrong\u003evalues\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eIn this study, the Mix-inResNet model achieved a precision of 94.61%, accuracy of 94.41%, F1 score of 94.22%, and Gwet\u0026apos;s kappa value of 93.09%. In comparison, the radiologist results were 95.23% precision, 95.10% accuracy, 94.89% F1 score, and 93.94% Gwet\u0026apos;s kappa value. \u003cstrong\u003eFigure 6\u003c/strong\u003e illustrates that the model performed excellently in classifying various types of pneumonia, with results comparable to those of the associate chief physician of the radiologist in terms of precision, accuracy, F1 scores, and Gwet\u0026apos;s kappa values. The model and the radiologist outperformed the other models and the clinician. The classification consistency of the Mix-inResNet model (kappa = 93.09%) was similar to that of the radiologist (kappa = 93.94%), demonstrating good agreement in classification and diagnostic efficiency. Additionally, \u003cstrong\u003eFigure 7\u003c/strong\u003e displays the heat map results generated by the model.\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e4.3 Comparison of ROC curves of four models\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe area under the curve (AUC) reflects the value of a diagnostic test, with a larger area closer to 1.0 indicating higher diagnostic accuracy. In the Mix-inResNet model, the AUC for bacterial, fungal, viral, tuberculous pneumonia, and normal chest CT was 0.99, 1.00, 0.99, 1.00, and 1.00, respectively \u003cstrong\u003e(Figure 8)\u003c/strong\u003e. The Mix-inResNet model outperformed other models by achieving the highest AUC values for fungal pneumonia, pulmonary tuberculosis, and normal chest CT samples. This indicates a superior level of accuracy in judgment and precision in classification. In the VGG16 model, the AUC for bacterial, fungal, viral, tuberculous pneumonia, and normal chest CT was 0.93, 0.91, 0.95, 0.94, and 1.00, respectively, showing lower classification recognition accuracy than all other models. Based on the figure, the AUC values of the four models mentioned are in descending order: Mix-inResNet \u0026gt; ResNet101 \u0026gt; Inception V3 \u0026gt; VGG16.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e4.4 Comparison of confusion matrices among four models\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe confusion matrix compares the model\u0026apos;s predictions with the actual classifications, providing a more intuitive way to assess the performance of four models in predicting accuracy for different categories of pneumonia. The Mix-inResNet model demonstrates a strong ability to distinguish between different types of pneumonia. It accurately identified 52 out of 59 cases of bacterial pneumonia, 35 out of 38 cases of fungal pneumonia, 57 out of 62 cases of viral pneumonia, 49 out of 50 cases of pulmonary tuberculosis, and all 77 cases of normal pneumonia. This highlights the model\u0026apos;s vital role in accurate identification and the importance of the matrix in assessing its diagnostic capabilities (\u003cstrong\u003eFigure 9A)\u003c/strong\u003e. The Inception V3 model had a few instances of misclassification among the five categories of pneumonia. Still, it provides valuable insights into the diagnostic accuracy of the model and enhances the prediction accuracy for certain pneumonia categories (\u003cstrong\u003eFigure 9B)\u003c/strong\u003e. The ResNet101 model exhibited occasional misclassification among five pneumonia categories but accurately classified pulmonary tuberculosis, with only one instance misidentified as viral pneumonia\u003cstrong\u003e\u0026nbsp;(Figure 9C)\u003c/strong\u003e. Comparatively, the VGG16 model performed poorly in classification accuracy, correctly classifying only 23 samples as fungal pneumonia and misclassifying 15 samples as other types of pneumonia\u003cstrong\u003e\u0026nbsp;(Figure 9D)\u003c/strong\u003e.\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003ePulmonary infections are the most common infectious diseases in respiratory medicine. They progress rapidly and result in high mortality, posing a severe threat to people\u0026apos;s lives and health\u003csup\u003e[3]\u003c/sup\u003e. The current methods for detecting pathogens in pulmonary infections have several limitations, including low sensitivity and accuracy, time-consuming, high labor or economic costs, and the abuse of antibiotic\u003csup\u003e[30, 31]\u003c/sup\u003e. It is crucial for computed tomography (CT) in diagnosing pneumonia. Chest CT scans offer a clear view of chest structures and are highly sensitive to chest lesions. This imaging method is able to identify small and hidden lesions by observing density, morphology, and margins.\u0026nbsp;While biases in interpreting imaging results and delays in reporting them by radiologists and physicians can hinder the timely identification of various pathogens in clinical practice. This often leads to delays in accurate anti-infection treatment and can worsen the patient\u0026apos;s condition.\u003c/p\u003e\n\u003cp\u003eArtificial intelligence is a complex and interdisciplinary field that has demonstrated significant potential, particularly in the realm of medical imaging within the field of medicine\u003csup\u003e[7, 11]\u003c/sup\u003e. Image datasets are commonly used to develop deep-learning diagnostic models to differentiate between various diseases in visual imaging diagnostics. The process of feature extraction was seamlessly integrated\u003csup\u003e[32]\u003c/sup\u003e. Pham\u003csup\u003e[33]\u003c/sup\u003e et al. conducted a classification study of COVID-19 pneumonia and other pneumonias using an extensive public database with 16 pre-trained Convolutional Neural Networks (CNNs) and achieved an accuracy of 0.917. Polsinelli\u003csup\u003e[34]\u003c/sup\u003e et al. proposed a lightweight SqueezeNet model to efficiently distinguish COVID-19 CT images from other community-acquired pneumonia and healthy individuals, achieving an accuracy of 0.850. While many studies have shown promising results in diagnosing pneumonia, most have been limited to COVID-19 and non-COVID-19 cases. This study presents the Mix-inResNet model, an innovative convolutional neural network (CNN) that utilizes CT scan images for multi-category image classification. This approach aims to support clinicians in simultaneously screening various types of pulmonary inflammation across large-scale samples.\u003c/p\u003e\n\u003cp\u003eOur deep learning system demonstrates exceptional proficiency in differentiating among various categories of pneumonia. Additionally, our findings indicate a higher accuracy level than prior studies that employed artificial intelligence techniques for CT diagnosis\u003csup\u003e[35]\u003c/sup\u003e. We compared the model with physicians to assess sensitivity, specificity, and 95% confidence intervals. The sensitivity of the Mix-inResNet model for detecting bacterial, fungal, viral, tuberculous pneumonia, and normal CT was 88.14%, 92.11%, 91.94%, 98.00%, and 100%, respectively. The specificities were 98.68%, 100.00%, 97.77%, 97.46%, and 99.04%, respectively. These results indicate that the model can accurately distinguish between different types of pneumonia. Compared to radiologist, sensitivity and specificity were validated, demonstrating the potential for clinical application.\u003c/p\u003e\n\u003cp\u003eThe Mix-inResNet model demonstrates 100% specificity in diagnosing fungal pneumonia, ensuring that negative results are not misclassified. The model\u0026apos;s advantages are further validated by a 95% confidence interval ranging from 0.88 to 0.97. Additionally, the Mix-inResNet model shows performance of 94.61% precision, 94.41% accuracy, 94.22% F1 score, and 93.09% Gwet\u0026apos;s kappa value, compared to radiologist performance of 95.23%, 95.10%, 94.89%, and 93.94% respectively. Compared to radiologist, the AI system utilizing CT images can rapidly identify potentially pathogenic infectious pneumonias. This technology significantly enhances diagnostic accuracy, offering timely guidance on clinical treatment options and ultimately maximizing patient benefits. The artificial intelligence system exhibits real-time capabilities and high accuracy, playing a pivotal role in facilitating early treatment decisions, thereby reducing hospitalization duration and curtailing treatment expenses. Simultaneously, it effectively mitigates the risk of complications arising from delayed diagnosis and treatment due to waiting for pathogen detection\u003csup\u003e[36, 37]\u003c/sup\u003e. Experience and excellent qualifications are the main reasons for the high accuracy of imaging physicians, consistent with previous studies\u003csup\u003e[36]\u003c/sup\u003e. Additionally, we utilized the Grad-CAM technique to localize the site of pneumonia, providing more intuitive and clear results\u003csup\u003e[38]\u003c/sup\u003e. The loss function curve (yellow line) of the Mix-inResNe model exhibits the best convergence performance and has the fastest learning effect in efficiently acquiring knowledge from the training data. This model significantly dominates in terms of learning when compared to other models. Zhang\u003csup\u003e[39]\u003c/sup\u003e et al. categorized pneumonia into four types: bacterial pneumonia, fungal pneumonia, common viral pneumonia, and novel coronavirus pneumonia, based on deep learning, with AUCs of 0.989, 0.996, 0.994, and 0.997, respectively. Their findings align with our study, however, the sample size is limited, particularly in regards to fungal pneumonia (n=4), which may potentially impact the model\u0026apos;s performance when compared to other forms of pneumonia within the test group.We included 38 cases of fungal pneumonia and demonstrated the excellent classification ability of Mix-inResNet in distinguishing between bacterial, fungal, viral, tuberculosis pneumonia, and routine chest CT scans. The area under the ROC curve (AUC) for each category was 0.99, 1.00, 0.99, 1.00, and 1.00, respectively. The Mix-inResNet model outperformed the other models with the highest AUC value, demonstrating excellent classification performance. We expanded the scope of our study by employing CT imaging to differentiate between bacterial, fungal, viral, and tuberculous pneumonia, which are not commonly found in existing literature. We conducted a comparative analysis of the confusion matrices and observed that the Mix-inResNet model exhibited remarkable proficiency in distinguishing among various pneumonia categories. This model achieved significantly fewer misclassifications when compared to the other three models evaluated.\u003c/p\u003e\n\u003cp\u003eThe Mix-inResNet model demonstrated strong performance in accurately identifying different pneumonia types, significantly reducing the risk of misdiagnosis. Additionally, it provides timely information on pathogens that can inform clinical decisions regarding the treatment of infections. The model\u0026apos;s diagnostic efficacy was further confirmed by validating key performance metrics such as precision, accuracy, sensitivity, specificity, recall, and F1 score, which are crucial for reducing misdiagnosis in healthcare settings.\u0026nbsp;\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eThe Mix-inResNet model demonstrated superior accuracy and reliability, showing great potential for improving the diagnostic process and patient outcomes. The Mix-inResNet model has demonstrated significant potential in clinical settings, enhancing the accuracy and reliability of lung disease diagnostic procedures. This study indicates that the Mix-inResNet model has the potential to revolutionize AI-driven disease diagnosis, particularly in pulmonary infections. Its high accuracy and robust performance across various class configurations demonstrate its potential as a valuable tool in medical diagnosis.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eAcknowledgement\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWe would appreciate all participants for their contributions to this study and thank the sequencing technical support from the Beijing Genomics Institute (Beijing, China).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eBM: Data curation, Manuscript preparation. CL: Data curation, Investigation, Writing \u0026ndash; review \u0026amp; editing. QC: Data curation, Resources, Sample collection. ZH: Funding acquisition, Writing \u0026ndash; review \u0026amp; editing. HL: Data curation，Investigation \u0026amp; editing. CM: Data curation, Writing - review \u0026amp; editing. TL: Data curation, Formal analysis, Writing \u0026ndash; review \u0026amp; editing. JF: Methodology, Resources, Writing - review \u0026amp; editing. AG: Project administration, Writing \u0026ndash; review \u0026amp; editing. YF: Funding acquisition, Project administration, Writing \u0026ndash; review \u0026amp; editing.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe study was supported by Natural Science Foundation of Hubei Province (Project No. JCZRYB202500657), Young Talent Project of General Hospital of Center Theater (ZZ20231693).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData availability\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe data used to support the findings of this study are available from the corresponding author upon request.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEthics approval and consent to participate\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe studies involving humans were approved by the Animal Ethical and Experimental Committee of the General Hospital of Center Theater of PLA (Wuhan, Hubei province. Permit No. 2024085-01 in accordance with their rules and regulations. The participants provided their written informed consent to participate in this study.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConsent for publication\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting interests\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare that they have no competing interests.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eLei J, Wang L, Li Q, Gao L, Zhang J, Tan Y. Identification of RAGE and OSM as New Prognosis Biomarkers of Severe Pneumonia[J]. Can Respir J, 2022,2022:3854191.\u003c/li\u003e\n\u003cli\u003eGlobal burden of 369 diseases and injuries in 204 countries and territories, 1990-2019: a systematic analysis for the Global Burden of Disease Study 2019[J]. Lancet, 2020,396(10258):1204-1222.\u003c/li\u003e\n\u003cli\u003eMagill S S, Edwards J R, Bamberg W, et al. Multistate point-prevalence survey of health care-associated infections[J]. N Engl J Med, 2014,370(13):1198-1208.\u003c/li\u003e\n\u003cli\u003evan Vugt S F, Verheij T J, de Jong P A, et al. Diagnosing pneumonia in patients with acute cough: clinical judgment compared to chest radiography[J]. Eur Respir J, 2013,42(4):1076-1082.\u003c/li\u003e\n\u003cli\u003eCui Y, Wang X, Lin F, et al. MiR-29a-3p Improves Acute Lung Injury by Reducing Alveolar Epithelial Cell PANoptosis[J]. Aging Dis, 2022,13(3):899-909.\u003c/li\u003e\n\u003cli\u003eHao S, Xin Q, Xiaomin Z, et al. Group membership modulates the hold-up problem: an event-related potentials and oscillations study[J]. Soc Cogn Affect Neurosci, 2023,18(1).\u003c/li\u003e\n\u003cli\u003eJiang B, Guo N, Ge Y, Zhang L, Oudkerk M, Xie X. Development and application of artificial intelligence in cardiac imaging[J]. Br J Radiol, 2020,93(1113):20190812.\u003c/li\u003e\n\u003cli\u003eHarmon S A, Sanford T H, Xu S, et al. Artificial intelligence for the detection of COVID-19 pneumonia on chest CT using multinational datasets[J]. Nat Commun, 2020,11(1):4080.\u003c/li\u003e\n\u003cli\u003eOzturk T, Talo M, Yildirim E A, Baloglu UB, Yildirim O, Rajendra Acharya U. Automated detection of COVID-19 cases using deep neural networks with X-ray images[J]. Comput Biol Med, 2020,121:103792.\u003c/li\u003e\n\u003cli\u003eJi G W, Zhu F P, Xu Q, et al. Machine-learning analysis of contrast-enhanced CT radiomics predicts recurrence of hepatocellular carcinoma after resection: A multi-institutional study[J]. EBioMedicine, 2019,50:156-165.\u003c/li\u003e\n\u003cli\u003eNazir M, Shakil S, Khurshid K. Role of deep learning in brain tumor detection and classification (2015 to 2020): A review[J]. Comput Med Imaging Graph, 2021,91:101940.\u003c/li\u003e\n\u003cli\u003eIbrahim D M, Elshennawy N M, Sarhan A M. Deep-chest: Multi-classification deep learning model for diagnosing COVID-19, pneumonia, and lung cancer chest diseases[J]. Comput Biol Med, 2021,132:104348.\u003c/li\u003e\n\u003cli\u003eSun W, Zheng B, Qian W. Automatic feature learning using multichannel ROI based on deep structured algorithms for computerized lung cancer diagnosis[J]. Comput Biol Med, 2017,89:530-539.\u003c/li\u003e\n\u003cli\u003eHao S, Jiali P, Xiaomin Z, et al. Group identity modulates bidding behavior in repeated lottery contest: neural signatures from event-related potentials and electroencephalography oscillations[J]. Front Neurosci, 2023,17:1184601.\u003c/li\u003e\n\u003cli\u003eLi L, Qin L, Xu Z, et al. Using Artificial Intelligence to Detect COVID-19 and Community-acquired Pneumonia Based on Pulmonary CT: Evaluation of the Diagnostic Accuracy[J]. Radiology, 2020,296(2):E65-E71.\u003c/li\u003e\n\u003cli\u003eZhang K, Liu X, Shen J, et al. Clinically Applicable AI System for Accurate Diagnosis, Quantitative Measurements, and Prognosis of COVID-19 Pneumonia Using Computed Tomography[J]. Cell, 2020,181(6):1423-1433.\u003c/li\u003e\n\u003cli\u003eKermany D S, Goldbaum M, Cai W, et al. Identifying Medical Diagnoses and Treatable Diseases by Image-Based Deep Learning[J]. Cell, 2018,172(5):1122-1131.\u003c/li\u003e\n\u003cli\u003eDu W, Luo X, Chen M. A Practical Deep Learning Model in Differentiating Pneumonia-Type Lung Carcinoma from Pneumonia on CT Images: ResNet Added with Attention Mechanism[J]. J Oncol, 2022,2022:8906259.\u003c/li\u003e\n\u003cli\u003eHuang J, Jiang E, Yang D, et al. Metagenomic Next-Generation Sequencing versus Traditional Pathogen Detection in the Diagnosis of Peripheral Pulmonary Infectious Lesions[J]. Infect Drug Resist, 2020,13:567-576.\u003c/li\u003e\n\u003cli\u003eLi H, Gao H, Meng H, et al. Detection of Pulmonary Infectious Pathogens From Lung Biopsy Tissues by Metagenomic Next-Generation Sequencing[J]. Front Cell Infect Microbiol, 2018,8:205.\u003c/li\u003e\n\u003cli\u003eMiao Q, Ma Y, Wang Q, et al. Microbiological Diagnostic Performance of Metagenomic Next-generation Sequencing When Applied to Clinical Practice[J]. Clin Infect Dis, 2018,67(suppl_2):S231-S240.\u003c/li\u003e\n\u003cli\u003eDe Pauw B, Walsh T J, Donnelly J P, et al. Revised definitions of invasive fungal disease from the European Organization for Research and Treatment of Cancer/Invasive Fungal Infections Cooperative Group and the National Institute of Allergy and Infectious Diseases Mycoses Study Group (EORTC/MSG) Consensus Group[J]. Clin Infect Dis, 2008,46(12):1813-1821.\u003c/li\u003e\n\u003cli\u003eKunihiro Y, Tanaka N, Kawano R, et al. Differential diagnosis of pulmonary infections in immunocompromised patients using high-resolution computed tomography[J]. Eur Radiol, 2019,29(11):6089-6099.\u003c/li\u003e\n\u003cli\u003eCilloniz C, Martin-Loeches I, Garcia-Vidal C, San Jose A, Torres A. Microbial Etiology of Pneumonia: Epidemiology, Diagnosis and Resistance Patterns[J]. Int J Mol Sci, 2016,17(12).\u003c/li\u003e\n\u003cli\u003eRea G, Sperandeo M, Lieto R, et al. Chest Imaging in the Diagnosis and Management of Pulmonary Tuberculosis: The Complementary Role of Thoraci Ultrasound[J]. Front Med (Lausanne), 2021,8:753821.\u003c/li\u003e\n\u003cli\u003eSzegedy C, Liu W, Jia Y, et al. Going deeper with convolutions[C], Boston, MA, 2015.\u003c/li\u003e\n\u003cli\u003eHe K, Zhang X, Ren S, et al. Deep Residual Learning for Image Recognition[C], Las Vegas, 2016.\u003c/li\u003e\n\u003cli\u003eSzegedy C, Ioffe S, Vanhoucke V, et al. Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning[C], San Francisco, 2017.\u003c/li\u003e\n\u003cli\u003eSzegedy C, Vanhoucke V, Ioffe S, et al. Rethinking the Inception Architecture for Computer Vision[C], Las Vegas, 2016.\u003c/li\u003e\n\u003cli\u003eDellinger R P, Levy M M, Rhodes A, et al. Surviving Sepsis Campaign: international guidelines for management of severe sepsis and septic shock, 2012[J]. Intensive Care Med, 2013,39(2):165-228.\u003c/li\u003e\n\u003cli\u003eChaudhuri A, Martinez-Martin P, Kennedy P G, et al. EFNS guideline on the management of community-acquired bacterial meningitis: report of an EFNS Task Force on acute bacterial meningitis in older children and adults[J]. Eur J Neurol, 2008,15(7):649-659.\u003c/li\u003e\n\u003cli\u003eLi H, Wang C. MIS-Net: A deep learning-based multi-class segmentation model for CT images[J]. PLoS One, 2024,19(3):e0299970.\u003c/li\u003e\n\u003cli\u003ePham T D. A comprehensive study on classification of COVID-19 on computed tomography with pretrained convolutional neural networks[J]. Sci Rep, 2020,10(1):16942.\u003c/li\u003e\n\u003cli\u003ePolsinelli M, Cinque L, Placidi G. A light CNN for detecting COVID-19 from CT scans of the chest[J]. Pattern Recognit Lett, 2020,140:95-100.\u003c/li\u003e\n\u003cli\u003eChumbita M, Cilloniz C, Puerta-Alcalde P, et al. Can Artificial Intelligence Improve the Management of Pneumonia[J]. J Clin Med, 2020,9(1).\u003c/li\u003e\n\u003cli\u003eWang F, Li X, Wen R, et al. Pneumonia-Plus: a deep learning model for the classification of bacterial, fungal, and viral pneumonia based on CT tomography[J]. Eur Radiol, 2023,33(12):8869-8878.\u003c/li\u003e\n\u003cli\u003eChung M, Bernheim A, Mei X, et al. CT Imaging Features of 2019 Novel Coronavirus (2019-nCoV)[J]. Radiology, 2020,295(1):202-207.\u003c/li\u003e\n\u003cli\u003eSelvaraju R R, Cogswell M, Das A, et al. Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization[C], Venice, 2017.\u003c/li\u003e\n\u003cli\u003eZhang Y H, Hu X F, Ma J C, et al. Clinical Applicable AI System Based on Deep Learning Algorithm for Differentiation of Pulmonary Infectious Disease[J]. Front Med (Lausanne), 2021,8:753055.\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"bmc-medical-imaging","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"bmim","sideBox":"Learn more about [BMC Medical Imaging](http://bmcmedimaging.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/bmim/default.aspx","title":"BMC Medical Imaging","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"pulmonary infection, CT scan, deep learning, artificial intelligence, accuracy","lastPublishedDoi":"10.21203/rs.3.rs-6323914/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6323914/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cstrong\u003eBackground:\u003c/strong\u003eCurrently, clinical pathogen diagnosis primarily consists of two methods: high-throughput sequencing and traditional culture. However, considering the evident limitations in terms of time and cost associated with the diagnostic process, there is an urgent need for a highly sensitive diagnostic and detection system to complement traditional methods. Therefore, this study proposes a Mix-inResNet model based on Convolutional Neural Network (CNN) aiming to enhance the accuracy of pulmonary infection diagnosis and effectively detect various types of lung diseases.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003ePurpose\u003c/strong\u003e: This study aimed to develop a deep learning model, Mix-inResNet, based on computed tomography (CT) scannings for the preclassification of pneumonia-related pathologies.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMethods\u003c/strong\u003e: A total of 273 cases of patients with pulmonary infections were included in this study, and both metagenomic next-generation sequencing (mNGS) and traditional culture methods were employed to comfirm the pathology. We possess a substantial collection of 2,858 imaging datasets with notable levels of infection were included to compare the diagnostic efficacy of the Mix-inResNet model against other models (ResNet 101, Inception V3, and VGG 16), as well as clinicians and imaging specialists.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eResults\u003c/strong\u003e:The Mix-inResNet model achieved precision, accuracy, F1 score, and Gwet's kappa values of 94.61%, 94.41%, 94.22%, and 93.09%, respectively. For bacterial, fungal, viral, and tuberculous pneumonia, as well as normal chest CT tests, the sensitivities were 88.14%, 92.11%, 91.94%, 98.00%, and 100%, respectively. The specificities were 98.68%, 100.00%, 97.77%, 97.46%, and 99.04%, respectively, with 95% confidence intervals ranging from 0.88 to 0.97. Compared to other models or manual identification, the Mix-inResNet model demonstrated the highest area under the curve (AUC) values and achieved more accurate classification. Confusion matrix results indicate that the Mix-inResNet model provides an excellent understanding and accurate identification of pneumonia. Additionally, Grad-CAM technology was utilized to localize the site of pneumonia, offering more intuitive and clear results.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConclusions\u003c/strong\u003e: The Mix-inResNet model performs at a level comparable to imaging specialists in distinguishing pneumonia classifications. It effectively reduces the misdiagnosis risk and can promptly deliver relevant pathogenic information. This provides clinicians with rapid and accurate diagnostic support, aiding the implementation of early therapeutic interventions.\u003c/p\u003e","manuscriptTitle":"Application of deep learning to automated diagnosis of computed tomography images of pulmonary infections","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-05-13 01:29:57","doi":"10.21203/rs.3.rs-6323914/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2026-05-13T10:43:31+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-05-08T13:02:50+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-05-05T05:17:18+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"290942326493589955167199085485305214236","date":"2026-04-29T16:44:43+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"270197022703045659371495712661054485815","date":"2026-04-27T17:06:11+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-05-26T08:28:30+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"184686971527573176692776582902843541024","date":"2025-05-26T06:39:35+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-05-05T17:58:03+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2025-04-09T09:59:28+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-04-08T13:25:30+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-04-08T13:21:56+00:00","index":"","fulltext":""},{"type":"submitted","content":"BMC Medical Imaging","date":"2025-03-28T01:33:58+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"bmc-medical-imaging","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"bmim","sideBox":"Learn more about [BMC Medical Imaging](http://bmcmedimaging.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/bmim/default.aspx","title":"BMC Medical Imaging","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"f49265fc-6798-44f2-8791-405069220c99","owner":[],"postedDate":"May 13th, 2025","published":true,"recentEditorialEvents":[{"type":"decision","content":"Revision requested","date":"2026-05-13T10:43:31+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-05-08T13:02:50+00:00","index":90,"fulltext":""}],"rejectedJournal":[],"revision":"","amendment":"","status":"in-revision","subjectAreas":[],"tags":[],"updatedAt":"2026-05-13T10:58:03+00:00","versionOfRecord":[],"versionCreatedAt":"2025-05-13 01:29:57","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-6323914","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6323914","identity":"rs-6323914","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}