SpikeFlow-4D maps rice panicle grain-filling kinetics by computed tomography

Rice yield and grain quality depend on the rate, timing, and uniformity of grain filling, yet the spatial dynamics of this process remain difficult to quantify within intact panicles. Destructive sampling reconstructs developmental trajectories from different plants and cannot directly follow how superior and inferior grains fill within the same organ. Here, we developed SpikeFlow-4D, an in situ 4D CT phenotyping framework for non-destructive, repeated monitoring of rice grain filling. The framework integrates coarse-to-fine segmentation, including slice-wise stem detection and removal, morphological cleaning, DBSCAN-based panicle separation, and grain instance extraction, while retaining original 32-bit reconstructed intensities for gray-value quantification and using 8-bit masks only for intermediate segmentation. Applied to two rice cultivars, 9311 and Zhonghua 11 (ZH11), SpikeFlow-4D quantified temporal changes in mean CT gray value, used as a relative attenuation proxy, across the upper, middle, and lower vertical strata of rice panicles. Logistic modeling of the stratum-level trajectories revealed delayed attenuation increase, a shorter active accumulation period than the upper stratum, and a smaller net attenuation increment in the lower stratum of 9311, whereas the estimated peak increase rate did not differ consistently among strata. At 16 days after flowering, the stratum-level CT gray-value pattern was consistent with independently measured grain fresh weight and was positively associated with dry weight across genotype × stratum group means, without establishing a direct calibration between CT gray value and dry-matter content. These results support SpikeFlow-4D as a non-destructive framework for quantifying spatially resolved, population-level grain-filling dynamics in rice panicles.