In light of the swift progress in digital technologies and artificial intelligence, deep learning approaches have played an expanding role in medical image detection. In lung CT images, lesions often exhibit complex contextual and fine-grained features. The location and scale of these lesions can significantly influence feature representation, making them difficult to capture accurately and thus reducing detection performance. To solve this challenge, we propose a novel model, YOLO-SPASS, which combines the YOLOv8 framework with the AssemFormer and SaELayer attention mechanisms to refine feature representations at multiple scales. Additionally, the model integrates the SPD-Conv module to improve convolutional feature encoding and adopts Inner-SIoU loss to preserve contextual information and enhance localization accuracy. The experimental results demonstrate that the proposed model achieves mAP@0.5 improvements of 5.1%, 4.3%, and 1.9% on CT images from the DeepLesion, Lung-PET-CT-Dx, and LUNA16 datasets, respectively. Furthermore, it achieves the highest average precision (AP) in multi-scale lesion detection tasks. These findings highlight the model’s potential to significantly enhance performance in lung CT-based multi-scale lesion detection, contributing to more accurate and efficient diagnosis and treatment of lung diseases.