Lung tumor segmentation in 4D CT images using motion convolutional neural networks

Area of Science:

• Medical Imaging
• Radiotherapy
• Artificial Intelligence

Background:

• Manual segmentation of lung tumors in 4D CT is time-consuming and prone to errors due to respiratory motion.
• Accurate tumor delineation is crucial for effective lung cancer radiotherapy planning.

Purpose of the Study:

• To develop a novel deep learning (DL) framework for fast and accurate segmentation of lung tumors in 4D CT datasets.
• To address the challenges of manual delineation, including time constraints and inter-observer variability.

Main Methods:

• A deep learning framework utilizing a motion region convolutional neural network (R-CNN) with integrated global and local motion estimation.
• Incorporation of a self-attention strategy in the mask head to enhance segmentation performance.
• Validation through cross-validation and evaluation on hold-out datasets using metrics like Dice Similarity Coefficient (DSC) and Volume Difference (VD).

Main Results:

• The automated DL method demonstrated high agreement with ground truth tumor segmentation.
• Achieved significantly superior performance (p < 0.05) compared to four other methods, including VoxelMorph and U-Net.
• Reported DSC values of 0.86-0.90 on hold-out datasets, with the smallest tumor volume difference (0.50).

Conclusions:

• The proposed DL framework shows significant promise for fully automated lung tumor segmentation in 4D CT.
• This automated approach can enhance the accuracy and efficiency of lung radiotherapy treatment planning.
• Facilitates integration into clinical workflows for improved patient care.