Artificial intelligence (AI) is changing the way healthcare services are delivered when it comes to radiodiagnosis. The motive to ensure quicker and precise diagnostic efficacy is what drives deep learning-based research. In situations that demand immediate and apt responses in the absence of health-care representatives, AI seems to stand out. The automated detection of chest pathologies from routine chest radiographs (CXRs) will save time and give physicians a choice to prioritize high-risk patients.
Pleural effusion is a commonly encountered pathological condition that goes unnoticed until the patient experiences proper clinical signs and symptoms. Pleural effusion is the excess fluid that accumulates in the pleural cavity, which is the fluid-filled space between the lungs and the chest cavity walls. It is seen in patients with congestive heart failure, nephrotic syndrome, liver cirrhosis, and severe malnutrition. Pleural effusion is a significant sign in infectious diseases like tuberculosis and pneumonia. A minimum of 500 ml of fluid is necessary for clinical detection of pleural effusion. The patient experiences chest pain while breathing, apnea (shortness of breath), cough, and fever, if associated with infections. The fluid accumulates in the lower parts of the lung as a result of gravity. Radiological investigations are commonly used for diagnosing pleural effusion. The blunting of the costophrenic angle, i.e., the angle that forms between the ribs (costo) and diaphragm (phrenic), is the striking feature observed in pleural effusion. Effusions are diagnosed with a minimum of 200 ml of fluid deposition. Large effusions have a meniscus that tapers towards the lateral aspect of the lung. It is frequently observed that mild fluid deposits in the pleural space are an incidental finding. But if undiagnosed for a prolonged period, the underlying cause can exacerbate the condition leading to multiple complications like the formation of abscess due to infection (empyema), scarring the lining of the lungs (pleural thickening), or permanent lung damage in cases with underlying malignancy. A common complication observed after drainage of the effusion is pneumothorax, i.e., air accumulation in the chest cavity.
DeepTek’s AUGMENTO platform
DeepTek aims to utilize artificial intelligence-based deep learning models to assist experts in delivering care faster and easier. Our work focuses on identifying chest pathologies and developing models for the same. The detection of pleural effusion is a part of DeepTek's AUGMENTO platform, which includes several other chest pathologies like tuberculosis, pneumonia, lung mass, pneumothorax, nodules, etc. AI models for detecting chest, knee, and spine pathologies are included in AUGMENTO. AUGMENTO's AI tools provide an impressive turnaround time for reports by improving the physician's efficacy. This platform enables radiologists and imaging experts to read more scans in a shorter period, reduces stress, and lowers burnout rates.
How AI proves valuable
Timely detection of pleural effusion has a greater value in clinical settings, including intensive care units and emergency rooms. Leveraging AI for diagnosis can bolster critical care treatment decisions. Using it in combination with a point-of-care device like an X-ray scanner can enhance and augment patient care using the instant notification feature and worklist reprioritization.
Figure 1: original image (left), radiologist annotated image (centre), and model output (right).
Working of the model
All the procured dataset images are manually annotated by drawing bounding boxes for marking the areas that show pathologic findings. The radiographs are sourced from multiple publicly available datasets as well as from private hospitals. These annotations are done by experienced radiologists. The model learns from these annotations to give its predictions for new radiographs that are provided. The U-Net with attention model with an Xception encoder is used to detect the presence of pleural effusion. Automated detection of pleural effusion from chest X-rays, though not easy, is an achievable task. To make the AI model more robust, a dataset that is wide and varied is essential, which led us to include positive (pathological CXRs) and negative (non-pathological as well as CXRs with differential pathologic findings) to be included. The AI model predictions (marked in white) co-relate with the radiologist-annotated images (marked in blue) - ref. Fig1-2
Figure 2: original image (left), radiologist annotated image (centre), and model output (right).
Although in an experimental phase, this model is being trained on a versatile dataset, and we plan to train the model on a more extensive and diverse dataset. The final goal is that the model should distinguish pleural effusion from normal lung findings and similar pathologies like pleural thickening. Since precise detection of various chest pathologies is a necessary factor for treatment planning and prioritization, we aim to make progress in providing additional tools to assist healthcare setups in delivering efficient patient care.