JoHof / lungmask

Automated lung segmentation in CT
Apache License 2.0
669 stars 150 forks source link

Request for Possibility to Adapt Lungmask from 2D U-Net to 3D U-Net #81

Open catkatze opened 1 year ago

catkatze commented 1 year ago

I would like to inquire whether it is feasible to adapt Lungmask, which currently utilizes a 2D U-Net architecture for lung segmentation, to a 3D U-Net architecture. Due to its layer-by-layer recognition process, the 2D U-Net sometimes produces hollow segments within the lung region.

By transitioning to a 3D U-Net, which considers the volumetric information from neighboring slices, we can potentially overcome the hollow segmentation problem and achieve more robust and consistent lung segmentation results. The 3D U-Net can capture the complete structure of the lungs, even in cases with overlapping structures or pathologies.

I kindly request information on whether the development team of Lungmask has plans to investigate the adaptation to a 3D U-Net model. If this enhancement is not currently being considered, I would appreciate the opportunity to contribute to the project by exploring the implementation of a 3D U-Net architecture for lung segmentation.

Thank you for your attention and consideration. I eagerly await your response regarding the possibility of adapting Lungmask to a 3D U-Net model.

Sincerely, Zechary ZHU