|
||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
|
Role of Artificial Intelligence in Detecting Pneumothorax and Cardiomegaly in Chest X-rays: An Observational Study |
||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
|
Manasa Mayukha Hanumanthu, Harsha Kopuru, Bala Murali Krishna Vadana, Sandeep Velicheti, Sai Preethi Athota, Anveeksha Marineni, Chandra Sekhar Kondragunta 1. Resident, Department of Radiodiagnosis, Dr. PSIMS and RF, Vijaywada, Andhra Pradesh, India. 2. Assistant Professor, Department of Radiodiagnosis, Dr. PSIMS and RF, Vijaywada, Andhra Pradesh, India. 3. Associate Professor, Department of Radiodiagnosis, Dr. PSIMS and RF, Vijaywada, Andhra Pradesh, India. 4. Professor, Department of Radiodiagnosis, Dr. PSIMS and RF, Vijaywada, Andhra Pradesh, India. 5. Resident, Department of Radiodiagnosis, Dr. PSIMS and RF, Vijaywada, Andhra Pradesh, India. 6. Resident, Department of Radiodiagnosis, Dr. PSIMS and RF, Vijaywada, Andhra Pradesh, India. 7. Professor, Department of Radiodiagnosis, Dr. PSIMS and RF, Vijaywada, Andhra Pradesh, India. |
||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
|
Correspondence Address : Manasa Mayukha Hanumanthu, C Block, G-1 SLV, Anjani Heights, Chinautopaly, Vijaywada-521101, Andhra Pradesh, India. E-mail: hmanasamayukha@gmail.com |
||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
|
||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| ABSTRACT |  | |||||||||||||||||||||||||||||||||||||||||||||||||||||||||
: Introduction: Pneumothorax and cardiomegaly are pathological conditions affecting the respiratory and cardiovascular systems, respectively. Early and accurate detection of these abnormalities in Chest X-rays (CXR) is crucial for timely intervention and improved patient outcomes. Artificial Intelligence (AI) has emerged as a promising tool in medical imaging, showing potential in automating the detection of various abnormalities. Aim: To investigate the effectiveness of AI-based algorithms in the assessment of pneumothorax and cardiomegaly through the analysis of CXR images. Materials and Methods: This was a prospective observational study conducted at the Radiology Department of Dr. Pinnamaneni Siddhartha Institute of Medical Sciences and Research Foundation (PSIMS and RF), Vijaywada, Andhra Pradesh, India, from July 2024 to November 2024. A total of 200 patients who were referred to the Radiology Department for CXR evaluation as part of their clinical assessment were included in the study. The study utilised the DeepTek’s Augmento AI model for interpreting CXRs of patients presenting to the emergency department with chest pain or shortness of breath. The sensitivity, specificity, Positive Predictive Value (PPV) and Negative Predictive Value (NPV) of the DeepTek’s Augmento AI model were calculated for its ability to detect pneumothorax and cardiomegaly in chest radiographs. Results: The AI model exhibited 91% sensitivity, 100% specificity, 100% PPV and 97% NPV in detecting the pneumothorax, and 89.5% sensitivity, 100% specificity, 100% PPV and 95% NPV in detecting the cardiomegaly in chest radiographs. Conclusion: The study demonstrates that the DeepTek’s Augmento AI model exhibit high sensitivity and specificity in detecting both pneumothorax and cardiomegaly on CXRs. | ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
|
|
||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| Keywords : Accuracy, Algorithm, Computer-aided detection, Early detection | ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
|
|
||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
|
DOI and Others :
DOI: 10.7860/IJARS/2025/75931.3048
Date of Submission: Sep 29, 2024 Date of Peer Review: Dec 24, 2024 Date of Acceptance: Feb 15, 2025 Date of Publishing: May 01, 2025 AUTHOR DECLARATION: • Financial or Other Competing Interests: None • Was Ethics Committee Approval obtained for this study? Yes • Was informed consent obtained from the subjects involved in the study? No • For any images presented appropriate consent has been obtained from the subjects. Yes PLAGIARISM CHECKING METHODS: • Plagiarism X-checker: Sep 30, 2024 • Manual Googling: Feb 11, 2025 • iThenticate Software: Feb 13, 2025 (13%) ETYMOLOGY: Author Origin EMENDATIONS: 6 |
||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
|
|
||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| INTRODUCTION |
| |||||||||||||||||||||||||||||||||||||||||||||||||||||||||
Although the use of Computed Tomography (CT) and Magnetic Resonance Imaging (MRI) has grown, chest radiography remains the most frequently performed radiologic procedure worldwide. CXRs play a critical role in the emergency department, serving as a rapid and essential diagnostic tool. They are often the first imaging modality used to assess patients with acute symptoms such as chest pain, shortness of breath, trauma, or suspected infections. The speed and accessibility of CXRs make them invaluable in emergency settings where timely diagnosis is crucial (1),(2). The CXR help in quickly identifying life-threatening conditions like pneumothorax (collapsed lung), heart failure, pneumonia and rib fractures, enabling prompt treatment decisions. They are also used to monitor the progression of diseases and the effectiveness of interventions, such as confirming the placement of medical devices like endotracheal tubes, central lines, or chest drains (1). However, interpreting CXRs can be complex due to the overlap of bone structures, inconsistent image quality and low-contrast resolution. Consequently, radiologists often exhibit high interobserver variability in their analysis of CXRs (3). Additionally, the increasing volume of radiographs being performed has led to a shortage of radiologists in many regions, making it challenging to meet reporting demands (4). These issues have spurred the development of AI-based Computer-Aided Diagnosis (CAD) systems. The potential benefits of automated CXR analysis are numerous, including improved sensitivity for detecting subtle findings, prioritisation of urgent cases and automation of routine tasks. CAD systems can also assist emergency physicians and radiologists-in-training when senior radiologists are unavailable. Among the various AI applications in diagnostic imaging, CAD systems for chest radiography using deep learning algorithms have already demonstrated high effectiveness in detecting lung nodules, screening for tuberculosis, identifying pneumothorax and detecting various other chest abnormalities (5). The pneumothorax, in which air is present outside of the lung but within the pleural cavity, develops when air collects between the parietal and visceral pleurae inside the chest, putting pressure on the lung and potentially causing it to collapse (6). Early detection of pneumothorax via chest radiography is crucial for determining the need for emergent clinical intervention. Radiologists could benefit from an AI model that can quickly triage and detect pneumothorax to aid in earlier diagnosis and enhance patient care (7). Pneumothorax has been evaluated by several AI models so far. (Annalise Enterprise CXR, ChestEye) (8). These approved models are examples of computer-assisted triage tools that can be used to better prioritise and sort urgent results (9). Cardiomegaly has become increasingly prevalent. Enlargement of the heart both in the form of dilation or hypertrophy, leads to a spectrum of clinical heart failure syndromes. Typically, this ailment goes unnoticed until symptoms manifest. Hence, the accurate prediction of abnormal CXRs aids in the early diagnosis of clinical conditions. X-ray Net is an AI-based tool designed to detect a range of thoracic conditions, including cardiomegaly, from CXR images (10). The purpose of the present research was to evaluate the accuracy of AI at detecting pneumothorax and cardiomegaly and to compare the results obtained from an AI system to those obtained from a radiologist’s interpretation of cardiomegaly and pneumothorax. | ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| Material and Methods |
| |||||||||||||||||||||||||||||||||||||||||||||||||||||||||
A prospective observational study was conducted at the Department of Radiodiagnosis, Dr. PSIMS and RF, Vijaywada, Andhra Pradesh, India, from July 2024 to November 2024. Prior clearance was obtained from the Institutional Ethics Committee (IEC). The certificate numbers are PG/1190/24 (cardiomegaly) and PG/1191/24 (pneumothorax). The study utilised the DeepTek’s Augmento AI model for interpreting CXRs of patients presenting to the emergency department with chest pain or shortness of breath. Inclusion and Exclusion criteria: A total of 200 patients who were referred to the Radiology Department for CXR evaluation as part of their clinical assessment were included in the study. CXRs with rotation or poor exposure were excluded. Study Procedure The CXRs (PA view) were obtained from the Radiodiagnosis Department. The radiologist examined each X-ray independently to establish whether or not pneumothorax/cardiomegaly is present. All of the cases were fed into the AI software (as DICOM files) and individually examined. The AI software evaluated each case separately and determined whether the CXR was normal or abnormal. The results of the AI program are compared to radiologist reports (Table/Fig 1). In this study, the AI model used was UNet with Xception backbone and Attention, a UNet-based AI model architecture, as shown in (Table/Fig 2). There are four encoder (Xception) blocks and five decoder (convolution) blocks. Each encoder block was connected to a decoder block based on the input shape. The encoder was connected to the decoder through the bottleneck. The input dimension of the model was 512×512×3, where 512×512 is the spatial dimension of the input image and 3 is the number of colour channels (RGB). The output dimension was 512×512, which corresponds to the segmentation map with one channel for each pixel. During training, the model was fed a set of labeled images, where each pixel was assigned a label indicating which region it belongs to. The model then learns to recognise features in the input image and map them to the corresponding output labels using a combination of loss functions and optimisation algorithms (6),(7),(9). Radiograph collection and assessment: Chest radiographs were collected from the Radiology Department, with the exclusion of those lacking proper exposure, showing signs of rotation, or taken in the Anteroposterior (AP) view. The remaining radiographs were initially assessed by two radiologists: one with seven years of experience and the other with two years of experience. The radiographs were categorised as either normal or abnormal. The abnormal radiographs were further classified into two groups: those with pneumothorax and those with cardiomegaly. AI model analysis: Following the initial assessment, all selected chest radiographs were uploaded into the DeepTek’s Augmento AI software. The AI model analysed the images using the UNet architecture with an Xception backbone and Attention algorithm. The main worklist has various categories of display using which one can sort the radiographs. One of which was the sorting using pathology detected, which seemed to be useful in the emergency department. As shown in the (Table/Fig 3), it displays various pathologies detected and upon selecting pneumothorax, for example, it displayes a list of the cases that the software has diagnosed as pneumothorax. There is also a triage system, as shown in (Table/Fig 4), which has a colour coding that displays red, orange, yellow and green for critical, high, medium and low (green) priority cases, respectively. Both the abovementioned methods-sorting cases based on diagnosis and the triage system-proved useful in emergencies with heavy workloads. It helps to address these critical cases at first, followed by other cases (Table/Fig 5). There is also another display method using probability scoring system, which showed scores based on the probability of that particular pathology that has been detected (Table/Fig 6). The radiologist’s analysis determined the presence of pneumothorax or not (Table/Fig 7)a. For all the cases manual calculation of cardiothoracic ratio was been done using four lines. The first line was a perpendicular line drawn in the midline, from which two other lines were drawn up to the lateral border of the heart on either side. The fourth line measuring the maximum inter-thoracic diameter. The ratio of the maximum cardiac diameter (average of lines two and three) and the maximum inter-thoracic diameter was calculated. The ratio above 0.50 was considered as positive case (Table/Fig 8)a. As shown in (Table/Fig 7)b, AI uses a rectangular box to highlight the side of the chest where pneumothorax (collapsed lung) was detected. This box likely marks the affected area or region of interest where the pathology exists. As shown in (Table/Fig 8)b, the AI draws two rectangular boxes: one for the cardiac diameter (indicating the size of the heart) and another for the maximum inter-thoracic diameter (representing the overall width of the chest cavity). This visual annotation helps healthcare professionals quickly locate and assess the severity of these conditions based on the AI’s analysis. Statistical Analysis Data analysis was performed using Statistical Package for the Social Sciences (SPSS) software. The sensitivity, specificity, PPV and NPV of the DeepTek’s Augmento AI model were calculated for its ability to detect pneumothorax and cardiomegaly in chest radiographs. | ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| Results |
| |||||||||||||||||||||||||||||||||||||||||||||||||||||||||
Out of 200 CXRs, 37 cases are diagnosed with pneumothorax in radiologist’s analysis. The AI software showed almost similar results by accurately identifying pneumothorax, apart from the normal CXRs in 32 cases, but it showed negative result in five cases (Table/Fig 9). The AI model exhibited an 91% sensitivity, 100% specificity, 100% PPV and 97% NPV in detecting the pneumothorax. Out of 200 CXRs, cardiomegaly was detected by the radiologist by calculating cardiothoracic ratio in 67 cases, out of which 60 cases were given as positive and seven cases were given as negative for cardiomegaly by the AI software. All the 133 cases that were given as negative for cardiomegaly by the radiologist were given similar results by the AI software (Table/Fig 10). The AI model exhibited 89.5% sensitivity, 100% specificity, 100% PPV and 95% NPV in detecting the cardiomegaly in chest radiographs. | ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| Discussion |
| |||||||||||||||||||||||||||||||||||||||||||||||||||||||||
The use of AI in medical imaging began to take shape in the 1960s and 1970s, with early efforts focused on developing CAD systems. These initial systems were rudimentary and primarily aimed at pattern recognition in medical images, such as detecting lung nodules or breast cancer in radiographs. However, the technology was mostly theoretical and limited in practical application due to the constraints of computing power and algorithm sophistication at the time (11),(12). Significant progress was made in the 1980s and 1990s as advances in machine learning, especially neural networks, enabled more complex image analysis. The integration of AI into clinical workflows, however, gained real momentum in the 2000s, particularly with the advent of deep learning and the availability of large datasets, which allowed for the development of more accurate and reliable AI tools (13). Neural networks in radiology imaging refer to the application of AI algorithms, particularly deep learning, to analyse and interpret medical images such as X-rays, CT scans, MRI scans and ultrasounds (14). These neural networks are trained on large datasets of images to learn patterns and features that help diagnose various medical conditions. In radiology imaging, neural networks can detect abnormalities, such as tumors or fractures; classify images as normal or abnormal; segment images to identify specific features or structures; enhance image quality and reduce noise; and assist in image-guided procedures (15),(16),(17),(18). Some common neural network architectures used in radiology imaging include Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), Generative Adversarial Networks (GANs) and transfer learning models, such as U-Net and ResNet. Neural networks in radiology imaging aim to improve diagnostic accuracy, reduce interpretation time and enhance patient care (19). However, they are not meant to replace human radiologists but rather to assist and augment their expertise. In a study done by Bennani S et al., AI-assisted chest radiography interpretation has demonstrated significant benefits, particularly in improving sensitivity for six pathologies, including pneumothorax, while their study showed there is no increase in specificity in detection of pneumothorax (20). The present study showed that the AI model exhibited 91% sensitivity, 100% specificity, 100% PPV and 97% NPV in detecting the pneumothorax. In a study done by Lee KH et al., the AI model exhibited an accuracy of 0.95 in detecting the cardiomegaly in CXRs (21), while our study showed 89.5% sensitivity, 100% specificity, 100% PPV and 95% NPV in detecting cardiomegaly in chest radiographs. Limitation(s) The major limitation in the present study is the study sample. As the authors included only the X-rays with proper exposure and without rotation or artifacts, the number is further limited. | ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
|
||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
|
||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
|
|
||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
Original article / research
|
Full Version