A multi-task segmentation framework for quantitative phenotypic analysis of stone cells in rubber tree bark

Stone cells (sclereids) in the bark of Hevea brasiliensis are critical determinants of mechanical strength and cracking resistance, and their relative area proportion provides an essential quantitative basis for phenotypic evaluation and breeding decisions. However, accurate segmentation and stable quantification under complex tissue structures remain challenging due to blurred boundaries, dense distributions, and heterogeneous backgrounds. To address this, we formulate stone cell phenotyping as an intelligent perception–optimization problem and develop a unified end-to-end multi-task deep learning segmentation framework. The proposed system integrates a Stackelberg game-based tri-modal fusion mechanism to enable adaptive coordination of image, edge, and textual features, and an enhanced topological feature modeling module to capture long-range dependencies and preserve structural continuity. In addition, a global–local synergistic optimization strategy is designed to balance exploration and exploitation during multi-task training, improving convergence stability and system-level robustness. Experiments on a newly constructed dataset of 1500 annotated images from 19 varieties demonstrate superior performance over existing methods, achieving mean intersection over union (mIoU) scores of 80.35% and 91.88% for stone cell and sand bark segmentation, respectively, and reducing area proportion prediction error to 0.96% (coefficient of determination, R 2 = 0.9480). The proposed multi-task deep learning framework enables scalable and automated stone cell quantification for high-throughput rubber tree phenotyping, thereby supporting data-driven breeding decisions and offering extensibility to related biological image analysis tasks.