Calamansi (Citrofortunella microcarpa), a vital citrus crop in the Philippines, is increasingly threatened by leaf diseases such as Huanglongbing (HLB), which severely impact yield and quality. While artificial intelligence (AI) and Convolutional Neural Networks (CNNs) have shown potential for automated plant disease detection, most models are trained on ideal, studio-quality images with labor-intensive manual annotation, limiting their applicability in real-world farming conditions. This study investigates how three key factors: image enhancement using Contrast Limited Adaptive Histogram Equalization (CLAHE), AI-assisted annotation, and dataset condition (controlled vs. uncontrolled) affect the performance of CNNs in calamansi leaf disease detection. A 2×2×2 factorial experiment was conducted using 3,000 annotated images collected from farms in Oriental Mindoro, Philippines. These images were utilized in both training and evaluation phases of the four deep learning models: YOLOv11, YOLOv12, Faster R-CNN, and RF-DETR. Results show that CLAHE significantly improved model accuracy under uneven lighting by 4.8%, while AI-assisted labeling reduced annotation time by over 70% with minimal loss in precision. Among the models, YOLOv12 achieved the best overall balance of accuracy and speed (mAP@50:95 = 64.2%; Recall = 73.6% with a variation of 0.9), making it most suitable for scenarios prioritizing sensitivity. RF-DETR achieved the highest precision (72.3% with a variation of 0.5), underscoring its strength in minimizing false positives and providing reliable detections. This study underscores the value of preprocessing, semi-automated labeling, and realistic field datasets in building robust, deployable AI systems for plant disease detection, contributing to practical, low-cost solutions for resource-constrained agricultural environments.