Few-Shot Learning for CT Lung Nodule Detection Based on Open-Set Object Detection.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Lung nodule detection is crucial for early lung cancer diagnosis.
• Current deep learning models often require large annotated datasets, limiting their application.

Purpose of the Study:

• To develop a few-shot learning model for lung nodule detection in CT images.
• To leverage visual open-set object detection techniques for improved efficiency.

Main Methods:

• Utilized the Lung Nodule Analysis 2016 (LUNA16) dataset for training and testing.
• Compared classical You Only Look Once (YOLO) models with advanced open-set models like Grounding DINO and Open-Vocabulary DINO (OV-DINO).
• Developed and evaluated a novel few-shot learning model against established benchmarks.

Main Results:

• YOLO models achieved peak precision, recall, and mAP of 82.8%, 73.1%, and 77.4%, respectively.
• OV-DINO showed higher recall than YOLO but no significant advantage in precision or mAP.
• The proposed few-shot model outperformed YOLO and OV-DINO, achieving 89.4% precision, 96.2% recall, and 87.7% mAP with reduced training data and epochs.

Conclusions:

• The developed few-shot learning model exhibits enhanced scene transfer capabilities.
• It requires significantly fewer training samples and epochs for effective lung nodule detection.
• The model demonstrates a substantial improvement in the accuracy of lung nodule detection in CT images.