CitrusGAN: sparse-view X-ray CT reconstruction for citrus based on generative adversarial networks.

Area of Science:

• Agricultural Engineering
• Computer Vision
• Biotechnology

Background:

• Accurate 3D phenotyping is crucial for developing new fruit varieties.
• Manual phenotyping methods are inefficient, time-consuming, and prone to errors.
• High-throughput, precise, and cost-effective solutions are needed for fruit breeding.

Purpose of the Study:

• To develop a generative adversarial network (GAN)-based method for reconstructing 3D citrus CT models from sparse-view X-ray images.
• To enable high-throughput and precise 3D phenotyping of citrus fruits.
• To reduce the cost and time associated with traditional phenotyping.

Main Methods:

• Utilized a generative adversarial network (CitrusGAN) for 3D model reconstruction.
• Employed sparse-view X-ray images arranged in orthogonal pairs as input.
• Developed customized loss functions to enhance the learning of 2D X-ray features to 3D CT volume mapping.

Main Results:

• Achieved high-quality 3D citrus CT volume reconstruction using only 6 X-ray views.
• Obtained a structural similarity index of 92.1% and a peak signal-to-noise ratio of 26.374 dB compared to real CT models.
• Demonstrated precise measurement of phenotypic traits like fruit dimensions, volume, surface area, peel thickness, and segment count.

Conclusions:

• CitrusGAN effectively reconstructs 3D citrus models from sparse X-ray data, enabling precise phenotyping.
• The method offers a high-throughput, low-cost alternative to manual phenotyping.
• Potential applications include integration into fruit production lines or development of portable devices for in-field phenotyping.