| Abstrakt: |
Lung cancer is a very dangerous disease and one of the leading causes of cancer-related deaths worldwide. It often goes undetected until it reaches an advanced stage. Early detection of lung nodules, especially those ranging between 3mm-30mm, can aid radiologists in diagnosing the disease, as it poses a significant challenge for them. In this study, we propose a method for the detection and classification of these nodules using the LIDR-IDRI dataset. Our method consists of two parts. The first part focuses on introducing object detection algorithms using one of the recent version of YOLO (YOLO v7) to detect lung nodules. These algorithms enable the drawing of bounding boxes around the lung nodules without losing any vital information, thus assisting radiologists in identifying and tracking the nodules in adjacent computed tomography slices. We also evaluated the impact of different input images on nodule detection, including whole images (images without preprocessing and segmentation), lung segmented images (the lung area is extracted from the whole images), and preprocessed images (applying some filtering methods on the whole images). Our findings revealed that using whole images resulted in the best performance, achieving a detection mAP (mean average precision) of 81.28%. In the second part, we present a multi-class classification using transfer learning with the VGG16 model. This classification process demonstrated good performance in classifying the nodules detected in the first step by the YOLO object detector into three classes: benign, suspect, and malignant. The classification is based on the degree of malignancy given by each radiologist, which varies from (1 to 5) depending on the nodule malignancy. This approach has the potential to enhance the accuracy of nodule classification and improve the overall diagnostic process for lung cancer. [ABSTRACT FROM AUTHOR] |
Full text from SpringerLink