Robust and accurate pulmonary nodule detection with self-supervised feature learning on domain adaptation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer-Aided Diagnosis

Background:

• Medical imaging data annotation is costly and time-consuming.
• Supervised deep learning for computer-aided diagnosis (CAD) faces overfitting with limited data and vendor variability.
• High false-positive rates hinder clinical application of automatic lung nodule detection.

Purpose of the Study:

• To develop a robust and accurate lung nodule detection framework for CT scans.
• To address challenges of limited data, overfitting, and high false-positive rates in CAD.
• To improve the generalizability and clinical applicability of automated lung nodule detection.

Main Methods:

• A novel self-learning schema trains a pre-trained model on unlabeled data for rich feature representation.
• A 3D feature pyramid network (3DFPN) is proposed for high-sensitivity nodule detection using multi-scale features.
• A High Sensitivity and Specificity (HS) network reduces false positives by analyzing appearance changes across CT slices using Location History Images (LHI).

Main Results:

• The proposed detector achieved state-of-the-art sensitivity at a low false positive rate on the LUNA16 dataset.
• The framework demonstrated consistent detection performance across novel datasets.
• Generalizability was validated on diverse CT scanner datasets (SPIE-AAPM, LungTIME, HMS).

Conclusions:

• The developed self-learning framework significantly enhances lung nodule detection accuracy and robustness.
• The proposed 3DFPN and HS network effectively improve sensitivity and reduce false positives.
• This approach offers a promising solution for reliable computer-aided diagnosis in clinical practice.