Automatic lung tumor segmentation from CT images using improved 3D densely connected UNet.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Accurate lung tumor segmentation is crucial for effective lung cancer treatment planning.
• Tumor heterogeneity and similar visual characteristics of surrounding tissues pose challenges for robust segmentation.
• Existing methods may struggle with diverse lung tumor appearances.

Purpose of the Study:

• To develop and evaluate an improved 3D dense connected UNet (I-3D DenseUNet) for segmenting various lung tumors from CT images.
• To enhance feature propagation and reuse for improved segmentation accuracy.
• To provide a tool that assists radiologists in lung cancer treatment planning.

Main Methods:

• Development of an improved 3D dense connected UNet (I-3D DenseUNet) incorporating nested dense skip connections.
• Implementation of dense connections within encoder-decoder blocks to facilitate feature reuse.
• Application of a data augmentation strategy using a 3D thin plate spline (TPS) algorithm to mitigate overfitting.
• Evaluation on three datasets comprising 938 lung tumors (TCIA, LIDC, private).

Main Results:

• The I-3D DenseUNet achieved excellent Dice similarity coefficients (DSC): 0.8316 for TCIA and LIDC datasets, and 0.8167 for the private dataset.
• The nested dense skip connections effectively contributed similar feature maps between encoder and decoder sub-networks.
• The dense connections promoted robust feature propagation and reuse, enhancing segmentation quality.

Conclusions:

• The proposed I-3D DenseUNet demonstrates a strong capability for accurate lung tumor segmentation across diverse datasets.
• The method's performance indicates its potential to significantly aid radiologists in lung cancer diagnosis and treatment planning.
• Further validation and integration into clinical workflows could improve patient outcomes in lung cancer care.