Lung cancer (LC) is the second most prevalent type of cancer worldwide and the deadliest, accounting for one in every five cancer-related deaths globally. The chances of survival for patients detected with this type of cancer increase considerably when the diagnosis is made early, with the 5-year survival rate reaching up to 70%. Radiologists perform LC diagnosis through Computed Tomography (CT) images, but such diagnosis is a complex and error-prone task. Through computer-aided tools, this diagnostic process can be automated, reducing time and effort for specialists, as well as improving confidence in the diagnosis. The objective of this work was to evaluate and compare the effectiveness of Convolutional Neural Network (CNN) and Transformer architectures in detecting small lung nodules (≤15mm), where the guiding research question of this work was “What is the impact of the size of lung nodules on the detection accuracy of CNN and Transformer architectures?”. The dataset used was based on the public database LUNA16, filtering the test set to include only sections with nodules smaller than 15mm. The models chosen for our comparisons were YOLOv8, a CNN considered state-of-the-art in object detection, and DEtection TRansformer (DETR), which combines the transformer architecture with a CNN layer, where we obtained results such as mAP50 = 0.70, Sensitivity = 0.91 and Λ = 0.85 for the DETR and mAP50 = 0.90, Sensitivity = 0.83 and Λ = 0.77 for the YOLOv8. We also assessed the impact of nodule size on the performance of both models, where the performance of YOLOv8 was impacted by the decrease in nodules size, while DETR continued to show satisfactory results regardless of how small the nodules were.

Comparative Analysis of Detection Transformers and YOLOv8 for Early Detection of Pulmonary Nodules