Deep Learning-Based Tumor Detection in CT and PET Lung Images Using YOLOv7 and YOLOv8
Aliya Thaseen, Durgesh Nandan, Sheshikala Martha
Abstract
The current paper reports a new deep learning model accurate tumor localization in lung CT and PET-based scans using new YOLOv7 and YOLOv8 models. In contrast to the previous research, the presented method framework will combine individual pre-processing steps and dataset-individual augmentation to maximize tumor position localization in multimodal imaging. The models are also trained and tested on a large dataset that consists of a detailed guaranteed that the models can detect tumors of different size and contrast. Results in experiments show that YOLOv8 has an accuracy higher than 89.3, which is higher than that of YOLOv7 (88.6), at better precision (0.87 vs. 0.85), recall (0.88 vs. 0.86), and shorter inference time (11.8 ms vs. 12.5 ms). The suggested solution saves the fields on the interactivity of GPU memory and the size of models, which enables their implementation into clinical settings. This study confirms that such optimized YOLO models have practical potential of developing tumors at an early phase, fast and precise performance that aims to transform tumor imaging diagnostics in oncological radiology.