Efficient multiscale fully convolutional UNet model for segmentation of 3D lung nodule from CT image

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer-Aided Diagnosis

Background:

• Accurate lung nodule segmentation in CT images is crucial for lung cancer diagnosis and treatment.
• Manual segmentation is subjective and relies heavily on specialist expertise.
• Developing automated segmentation methods is essential for consistent and reliable results.

Purpose of the Study:

• To introduce a multiscale fully convolutional three-dimensional UNet (MF-3D UNet) model for automated lung nodule segmentation in CT images.
• To enhance nodule segmentation performance by integrating multiscale feature fusion and trainable downsampling techniques.

Main Methods:

• The MF-3D UNet model fuses multiscale features using Maxout aggregation to prioritize important features.
• Trainable downsampling layers are implemented, replacing fixed pooling operations for improved adaptability.
• The model was evaluated using CT scans from the Lung Image Database Consortium and Image Database Resource Initiative (LIDC-IDRI) dataset.

Main Results:

• The MF-3D UNet model demonstrated reliable segmentation results compared to other UNet variants.
• The model achieved encouraging performance in segmenting various nodule types, including juxta-pleural, solitary pulmonary, and non-solid nodules.
• The proposed method outperformed other Convolutional Neural Network (CNN)-based segmentation models.

Conclusions:

• The proposed MF-3D UNet model accurately segments lung nodules through multiscale feature aggregation and trainable downsampling.
• The utilization of 3D operations facilitates precise segmentation of complex nodules by leveraging inter-slice contextual information.