Nondestructive 3D phenotyping method of passion fruit based on X-ray micro-computed tomography and deep learning

Abstract

Passion fruit is a tropical liana of the Passiflora family that is commonly planted throughout the world due to its abundance of nutrients and industrial value. Researchers are committed to exploring the relationship between phenotype and genotype to promote the improvement of passion fruit varieties. However, the traditional manual phenotyping methods have shortcomings in accuracy, objectivity, and measurement efficiency when obtaining large quantities of personal data on passion fruit, especially internal organization data. This study selected samples of passion fruit from three widely grown cultivars, which differed significantly in fruit shape, size, and other morphological traits. A Micro-CT system was developed to perform fully automated nondestructive imaging of the samples to obtain 3D models of passion fruit. A designed label generation method and segmentation method based on U-Net model were used to distinguish different tissues in the samples. Finally, fourteen traits, including fruit volume, surface area, length and width, sarcocarp volume, pericarp thickness, and traits of fruit type, were automatically calculated. The experimental results show that the segmentation accuracy of the deep learning model reaches more than 0.95. Compared with the manual measurements, the mean absolute percentage error of the fruit width and length measurements by the Micro-CT system was 1.94% and 2.89%, respectively, and the squares of the correlation coefficients were 0.96 and 0.93. It shows that the measurement accuracy of external traits of passion fruit is comparable to manual operations, and the measurement of internal traits is more reliable because of the nondestructive characteristics of our method. According to the statistical data of the whole samples, the Pearson analysis method was used, and the results indicated specific correlations among fourteen phenotypic traits of passion fruit. At the same time, the results of the principal component analysis illustrated that the comprehensive quality of passion fruit could be scored using this method, which will help to screen for high-quality passion fruit samples with large sizes and high sarcocarp content. The results of this study will firstly provide a nondestructive method for more accurate and efficient automatic acquisition of comprehensive phenotypic traits of passion fruit and have the potential to be extended to more fruit crops. The preliminary study of the correlation between the characteristics of passion fruit can also provide a particular reference value for molecular breeding and comprehensive quality evaluation.

Challenges and opportunities in applying AI to evolutionary morphology

“…AI methodologies have seamlessly extended their proficiency from extracting 2D univariate traits to 3D, by employing analogous methods to obtain linear measurements of both length and width within 3D images (Hu et al, 2020;Lu et al, 2023). Similar to the techniques applied to their 2D counterparts, these methods can concurrently extract multiple traits from individual images (Wu et al, 2021) and tally features across diverse regions in 3D images (Yu et al, 2021).…”

Advanced deep learning models for phenotypic trait extraction and cultivar classification in lychee using photon-counting micro-CT imaging

“…Not only can it help identifying plants from complex environment, but also can rebuilt 3D models and get phenotypic traits when analyzing CT images of the plant. A nondestructive method for more accurate and efficient automatic acquisition of comprehensive phenotypic traits has been developed in the passion fruit ( Lu et al., 2023 ) and the coconut ( Yu et al., 2022 ).…”

SCGNet: efficient sparsely connected group convolution network for wheat grains classification

“…Machine learning methods, while effective, require substantial agricultural knowledge, manual feature selection, and classifier design (Lu et al, 2022). This process demands significant human effort and may not match the recognition speed achieved by deep learning approaches.…”