Improved brain metastases segmentation using generative adversarial network and conditional random field optimization...

Area of Science:

• Medical Imaging
• Radiotherapy
• Artificial Intelligence

Background:

• Accurate delineation of gross tumor volume (GTV) in brain metastases is critical for radiotherapy planning.
• Current methods lack automated tools for GTV segmentation using CT simulation localization images.

Purpose of the Study:

• To develop an effective tool for automatic GTV segmentation in brain metastases based on CT simulation localization images.

Main Methods:

• A dual-network generative adversarial network (GAN) was employed to refine CT images.
• Mask R-CNN was integrated for meticulous GTV segmentation.
• Conditional random fields (CRFs) were utilized for mask refinement, with an end-to-end training process.

Main Results:

• The integrated GAN+Mask R-CNN+CRF model achieved an average Dice Similarity Coefficient (DSC) of 0.819 and Intersection over Union (IoU) of 0.712 in internal validation.
• External validation showed an average DSC of 0.726 and IoU of 0.640, demonstrating good generalization.
• The method provides a robust automatic segmentation approach for brain metastases without MRI.

Conclusions:

• The developed tool offers a pioneering solution for sophisticated GTV segmentation in brain metastases.
• This integrated approach provides a robust automatic segmentation method, particularly valuable in the absence of MRI.