Developing a Machine Learning Model to Predict Pleural Adhesion Preoperatively Using Pleural Ultrasound

Lung cancer is currently the leading cause of death from malignant tumors worldwide. Surgery is the primary treatment method for lung cancer, and breakthroughs in thoracic surgical techniques play a crucial role in the diagnosis and treatment of lung cancer. Medical ultrasound technology, due to its non-invasive, flexible, convenient, and economical characteristics, is widely used in clinical practice. With the advancement of ultrasound technology and the need to address clinical challenges, ultrasound imaging technology has seen new developments in the field of thoracoscopic surgery.

In the era of minimally invasive surgery, intraoperative pleural adhesions are one of the main factors affecting the implementation of video-assisted thoracoscopic surgery (VATS). Especially under the concept of enhanced recovery after surgery (ERAS), the day surgery model for VATS has gradually taken shape. However, pleural adhesions significantly increase intraoperative trauma and prolong hospital stays. Additionally, pleural adhesions increase the risk of lung injury during VATS and, in severe cases, may hinder access to the pleural space, necessitating conversion to open thoracotomy. Pleural adhesions increase intraoperative time and morbidity in thoracic surgery due to poor visibility, bleeding, and lung and vascular injuries. The presence, location, and degree of pleural adhesions are useful for determining the initial port placement or choosing between open or VATS approaches. Therefore, accurately predicting the presence and specific location of pleural adhesions preoperatively is crucial for the development of day surgery under thoracic ERAS, ensuring the safety and efficiency of future VATS day surgeries.

Previous studies have shown that chest CT is difficult to predict pleural adhesions, with a sensitivity of only 72% and a sensitivity of only 46% for determining adhesions at specific locations. In contrast, ultrasonography of the pleura (USP) can dynamically display pleural sliding and adhesions with surrounding lung tissue, and has real-time monitoring capabilities based on movement, providing unique advantages for detecting pleural adhesions. Preoperative prediction of pleural adhesions using USP has significant application value. Studies have already demonstrated the advantages of using transthoracic pleural ultrasound to identify pleural adhesions. Nicola et al. conducted 1,192 ultrasounds to predict pleural adhesions, confirming 1,124 positive cases and 68 negative cases, with a sensitivity of 80.6%, specificity of 96.1%, positive predictive value of 73.2%, and negative predictive value of 97.4%. However, there are still some issues with using USP to predict pleural adhesions. Physicians who can identify pleural adhesions need to be trained in lung ultrasound, and ultrasound examination and interpretation are skill-dependent techniques. The burden of training thoracic ultrasound physicians remains a clinical challenge.

Three-dimensional convolutional neural network (3D-CNN) technology is an emerging technology in the field of artificial intelligence and machine learning. Unlike traditional convolutional neural networks (CNN), 3D-CNN can process three-dimensional data that includes a time dimension, making it suitable for analyzing the real-time dynamic image features of ultrasound images. This technology holds promise for developing a machine learning model to interpret USP images, potentially replacing physician interpretation and improving the accuracy of USP in identifying pleural adhesions.

In summary, this study intends to use USP for preoperative identification of pleural adhesions in patients scheduled for VATS surgery. It aims to investigate the accuracy of USP in predicting intraoperative pleural adhesions and to develop a diagnostic model using 3D-CNN technology to process USP-related images and video data for machine learning. The study will explore the sensitivity, specificity, positive predictive value, and negative predictive value of the 3D-CNN-USP model in identifying pleural adhesions. Additionally, it will examine the feasibility and effectiveness of using 3D-CNN-USP for preoperative identification of pleural adhesions to support the implementation of day surgery in thoracic surgery.