Evaluation of wheat blast resistance and AI-based digital phenotyping for parent selection in wheat breeding

Abstract

Wheat blast, caused by the fungus Pyricularia oryzae Triticum lineage (PoTl), is one of the major diseases threatening wheat crops in tropical and subtropical regions, which can cause severe yield losses. Since conventional resistance assessment is labor-intensive and genetic resistance is essential for management, the objective of this study was to evaluate the resistance of 20 wheat cultivars to four PoTl isolates, analyze the cultivar × isolate interaction via diallel analysis to identify promising parents, and develop an image analysis pipeline integrated with convolutional neural networks (CNN) to automate severity estimation. Greenhouse experiments were conducted in 2024 and 2025. The evaluated traits were Disease Severity (DS), Area Under the Disease Progress Curve (AUDPC), Number of Grains per Spike (NGS), Number of Shriveled Grains (NSG), and Grain Weight (GW). Statistical analysis revealed significant effects for the interactions among cultivar, isolate, and year. Diallel analysis evidenced differential aggressiveness among isolates and identified genotypes with high general combining ability. Based on the MGIDI index, the cultivars TBIO Convicto, MGS Brilhante, TBIO Calibre, ORS Soberano, and BRS 404 were selected as promising for use as parents in breeding programs in Brazil. Concurrently, 11 days after inoculation, images of the spikes were obtained, which were segmented from the background in PNG format with a transparent background. To perform the segmentation and separation between healthy and symptomatic tissues, the k-means algorithm was applied to the Hue (H) and Saturation (S) channels. Severity was quantified, and the spikes were categorized into four reaction classes. Three CNN models based on the YOLOv11 nano architecture were trained using transfer learning. The use of image processing effectively estimated the diseased area with higher accuracy than human visual assessment, showing a Lin's concordance correlation coefficient of 0.85. The general classification model achieved up to 99% accuracy in internal validation and maintained satisfactory performance in validation with independent images. It is concluded that the image pipeline is promising for digital phenotyping and that the selected genotypes constitute strategic sources of resistance for wheat farming.