การพัฒนา YOLOv8 ด้วยการปรับฟังก์ชันความสูญเสียเพื่อเพิ่มประสิทธิภาพ การตรวจจับก้วนเนื้อในปอดจากภาพเอกซเรย์ทรวงอกและการเปรียบเทียบประสิทธิภาพกับภาพเอกซเรย์คอมพิวเตอร์

Abstract

Lung cancer is a major cause of death worldwide and is often found at a late stage because early symptoms are unclear. Although Chest X-ray and CT scan are used for detection, they have limits in cost, time, and access. Chest X-ray images are especially hard to read due to overlapping structures and low detail. This study develops an AI model using YOLOv8 to detect lung nodules from Chest X-ray images. The dataset includes 130 cases from The Cancer Imaging Archive (TCIA) and LIDC-IDRI. Different data splits and classification loss (Cls loss) values were tested. The best result was achieved with an 80:20 (train:validation) split and Cls loss = 6. The model reached accuracy of 0.27, precision of 0.60, recall of 0.33, F1-score of 0.43, and IoU of 0.40. However, performance is still limited due to small data size and difficulty of Chest X-ray, especially for small nodules. When tested with CT scan images, the model performed much better. This shows that the main limits come from the image type and data size, not the model. This work can be improved and used as a support tool for medical diagnosis in the future.