Automated Lung Cancer Segmentation Using a PET and CT Dual-Modality Deep Learning Neural Network

Area of Science:

• Medical Imaging
• Radiotherapy Planning
• Artificial Intelligence in Oncology

Background:

• Accurate lung tumor segmentation is crucial for effective radiation therapy planning.
• Current segmentation methods can be time-consuming and subject to inter-observer variability.

Purpose of the Study:

• To develop an automated lung tumor segmentation method using deep learning on dual-modality PET and CT images.
• To improve segmentation accuracy, especially for small tumors, through a novel stratification strategy.

Main Methods:

• A 3D convolutional neural network (CNN) was designed with parallel convolution paths for feature extraction and a deconvolution path for segmentation.
• The network was trained on 290 PET/CT image pairs, with manual physician contours serving as ground truth.
• A GTV-based stratification strategy was investigated to optimize performance for different tumor sizes.

Main Results:

• The unstratified dual-modality model achieved a mean Dice similarity coefficient of 0.79.
• The GTV-stratified model yielded improved results, with the best combined Dice similarity coefficient of 0.83.
• Automated contours showed high clinical acceptability, with 88.75% accepted or accepted with modifications by radiation oncologists.

Conclusions:

• The proposed 3D deep learning network with GTV-based stratification effectively generates clinically useful lung cancer contours.
• The automated segmentation method demonstrates high accuracy and physician acceptance, aiding radiation therapy planning.
• This approach leverages a dual-modality imaging architecture and a large clinical dataset for robust performance.