Accurate delineation of individual tree crowns (ITC) is essential for forest inventory, biomass estimation, and ecological monitoring. Airborne LiDAR data provide an effective basis for this task, but accurate delineation remains challenging due to complex canopy structure and limited labelled training data. In this study, a modular fusion framework is developed that combines bounding box detection, multi-model fusion, and foundation model segmentation to produce tree crown masks from canopy height models (CHMs). Two detection architectures, DINO and Faster R-CNN, are implemented with both ImageNet-pretrained ResNet-50 and domain-specific Masked Autoencoder (MAE) backbones. Due to the small size of the target dataset, supplementary training data from the Finnish Taiga–Tundra Ecotone are incorporated to stabilize transformer training. Predictions from the individual detectors are fused using Weighted Box Fusion, with score normalization applied to account for differing confidence distributions between architectures. The fused detections serve as prompts for the Segment Anything Model (SAM), enabling a decoupled detection and segmentation framework for tree crown delineation. On a mixed-wood forest site in the Great Lakes–St. Lawrence forest region of Ontario, Canada, the framework achieves mask-level F1 scores of 0.79 and 0.61 at IoU thresholds of 0.25 and 0.50 respectively, matching 193 of 233 ground truth trees on the primary test plot. Visual inspection indicates that many delineated crowns align more closely with canopy structure than the reference polygons. The proposed framework demonstrates the potential of modular detector fusion and detection-guided segmentation for ITC delineation from LiDAR-derived canopy height models.