Strided Self-Supervised Low-Dose CT Denoising for Lung Nodule Classification | JoVE Visualize

Area of Science:

Medical imaging
Artificial intelligence
Radiology

Background:

Low-dose computed tomography (LDCT) is crucial for lung cancer screening, but image noise hinders accurate lung nodule classification.
Existing denoising methods for LDCT require normal-dose CT (NDCT) images for supervision, which are often unavailable in clinical settings.
This limitation necessitates the development of self-supervised or unsupervised approaches for effective LDCT image analysis.

Purpose of the Study:

To develop a novel self-supervised method (SN2N) for joint blind medical image denoising and lung nodule classification using only LDCT images.
To eliminate the need for paired NDCT images in the denoising process, making the method clinically practical.
To enhance the performance of lung nodule classification by integrating denoising within a unified framework.

Main Methods:

Introduced strided Noise2Neighbors (SN2N), a self-supervised framework for blind denoising and lung nodule classification.
SN2N generates supervision for denoising directly from noisy LDCT images using neighboring pixel information.
Employed a joint training strategy with self-supervised loss for denoising and cross-entropy loss for classification.

Main Results:

SN2N demonstrated competitive performance against supervised methods using paired NDCT images on the Mayo LDCT dataset.
Joint training of denoising and classification tasks significantly improved LDCT-based lung nodule classification accuracy on the LIDC-IDRI dataset.
The proposed self-supervised approach effectively addresses the noise challenge in LDCT images without requiring additional high-dose scans.

Conclusions:

The SN2N method offers a practical and effective self-supervised solution for denoising LDCT images and classifying lung nodules.
Jointly training denoising and classification tasks in a unified framework enhances diagnostic accuracy in lung cancer screening.
This approach has significant implications for improving the reliability and accessibility of LDCT-based lung cancer diagnosis.