Lung Cancer Segmentation With Transfer Learning: Usefulness of a Pretrained Model Constructed From an Artificial...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Accurate lung cancer segmentation is crucial for diagnosis and treatment planning.
• Developing robust segmentation models often requires large, diverse datasets, which can be challenging to obtain.

Purpose of the Study:

• To develop and evaluate a lung cancer segmentation method utilizing a pretrained model and transfer learning.
• To assess the efficacy of an artificial dataset generated via generative adversarial network (GAN) for model pretraining.

Main Methods:

• Utilized three public datasets (LUNA16, Decathlon lung, NSCLC radiogenomics) for model development and evaluation.
• Generated an artificial dataset for lung cancer segmentation using GAN and 3D graph cut on the LUNA16 dataset.
• Constructed pretrained models from the artificial dataset and a main segmentation model using the Decathlon lung dataset.
• Evaluated the main segmentation model on the NSCLC radiogenomics dataset using the Dice Similarity Coefficient (DSC).

Main Results:

• The pretrained models led to an overall improvement in mean DSC on the NSCLC radiogenomics dataset.
• A maximum increase of 0.09 in mean DSC was observed when using the pretrained model compared to not using it.
• The generation of an artificial dataset for segmentation using GAN and 3D graph cut was demonstrated as feasible.

Conclusions:

• The proposed method, incorporating an artificial dataset and a pretrained model, effectively enhances lung cancer segmentation performance.
• The study confirms the feasibility and utility of GAN-generated data for improving deep learning models in medical image analysis.