Leveraging Self-supervised Pretraining Using Transformers for Enhanced Lung Nodule Detection in CT Scans
摘要
Lung nodule detection is critical for early diagnosis of lung cancer, but remains challenging due to the nodules’ resemblance to normal tissues. Recent transformer-based approaches have made significant progress; however, their large number of parameters necessitates extensive annotated datasets to achieve robust and reliable results. To address this, we leverage state-of-the-art self-supervised training methods, specifically Masked Image Modeling, on a large domain-specific dataset of lung screening CTs, followed by finetuning on the annotated LUNA16 dataset. Our method achieves an AP of 82.63% and an mAP of 81.23%, outperforming the baseline nnDetection. The experiments demonstrate the effectiveness of pretraining, yielding an increase of 24.0% in performance on the Video-ViT backbone and 4.1% on the Swin Transformer. Additionally, we examine the effect of RGB video pretraining and architectural variations during both pretraining and fine-tuning stages. This work highlights the potential of self-supervised learning in improving efficiency and accuracy in lung cancer screening. Code: github.com/vit-swin-lung-nodule-detection .