XFruitSeg-A general plant fruit segmentation model based on CT imaging

Area of Science:

• Plant science
• Computer vision
• Bioinformatics

Background:

• Accurate fruit phenotyping is crucial for understanding genetic traits.
• Computed tomography (CT) imaging offers noninvasive 3D visualization of internal fruit structures.
• Existing segmentation methods struggle with the unique characteristics of plant fruit CT images.

Purpose of the Study:

• To develop a general deep learning model for segmenting internal plant fruit tissues from CT images.
• To improve the accuracy and reliability of fruit phenotyping analysis.
• To introduce a novel model, XFruitSeg, and an associated dataset, XrayFruitData.

Main Methods:

• Developed XFruitSeg, a U-shaped encoder-decoder deep learning model integrating multitask learning.
• Incorporated RepLKNet for expanded receptive fields, multiscale skip connections, and deep supervision for feature learning.
• Added a contour feature learning branch and an optimized composite loss function for robustness.
• Created the XrayFruitData dataset with high-resolution CT images of twelve fruit varieties.

Main Results:

• XFruitSeg demonstrated superior segmentation performance compared to four advanced models on orange, mangosteen, and durian datasets.
• Achieved high mean Dice coefficients (e.g., 95.21% for orange) and mIoU scores (e.g., 91.09% for orange).
• Ablation experiments confirmed the effectiveness of individual components within the XFruitSeg model.

Conclusions:

• XFruitSeg is an effective deep learning model for high-precision segmentation of internal fruit tissues in CT images.
• The model facilitates accurate internal fruit phenotyping, aiding in the understanding of complex genetic traits.
• XFruitSeg provides a robust foundation for future research in plant science and agricultural applications.