Improving the identification of haploid maize seeds using convolutional neural networks

A critical step toward the success of the doubled haploid (DH) technique is the haploid identification within induction crosses. The R1-nj marker is the principal mechanism employed in this task enabling the selection of haploids at the seed stage. Although it seems easy to identify haploid seeds, this task is performed manually by visual classification, which becomes an inefficient process in terms of time and labor. Also, differential phenotypic expression of the R1-nj marker results in high rates of false positives among haploid seeds. For the first time, an image-based convolutional neural network (CNN) was trained to identify true positives among putative haploid seeds. The experiment was conducted using 3,000 maize (Zea mays L.) seeds from induction crosses classified as haploid (1,000), diploid (1,000), and inhibited (1,000) class. Images were taken from each seed, and then seeds were planted in the field to confirm their ploidy. For putative haploids (R1-nj phenotype), the classification accuracy on average was 94.39%, 97.07% for the haploid class, and 91.71% for the diploid class. However, the CNN model was unable to distinguish true haploid seeds among the putative haploid class, which indicates that CNN did not recognize different patterns between them. Finally, we provided a highly accurate and trained CNN model to the scientific community to classify haploid maize seeds via R1-nj, which can support maize breeders to optimize DH pipelines, mainly for small breeding programs with limited resources.