In situ evaluation of Oudemansiella raphanipes fruiting body growth status using pseudo-label-assisted instance segmentation

In situ assessment of Oudemansiella raphanipes fruiting bodies could support cultivar evaluation and cultivation management, but dense cap clusters and limited annotated data hinder automated analysis. To address these issues, we developed a cascaded framework that combines automatic annotation, image synthesis, instance segmentation, and phenotypic classification. Regarding dataset construction, Depth Anything V2 and the Segment Anything Model were combined to generate cap pseudo-labels. Furthermore, an illumination-adaptive synthesis strategy matched cap instances to local background brightness and augmented small objects. For algorithm optimization, Mask2Former was modified by incorporating deformable convolution and coordinate attention into its backbone. Moreover, the resulting instance masks were used for cap shape classification and distance-based classification of clustered and solitary growth. On 30 real test images, the modified model achieved a mean average precision of 86.4%, exceeding the baseline by 7.5 percentage points. Cap shape classification and clustered/solitary classification achieved accuracies of 88.2% and 94.3%, respectively. Exploratory analysis identified nonlinear associations between the proportion of irregular caps and cap area and fruiting body height. These associations suggest potential phenotypic indicators for cultivation assessment. In summary, the proposed framework reduces annotation dependence, improves segmentation accuracy, and enriches phenotypic analysis, offering practical support for digital cultivation, intelligent harvesting, and smart breeding in industrial mushroom production.