Pengembangan Sistem Dermatologi Veteriner Kucing dan Anjing Berbasis Two-Stage Lesion-Aware Pipeline dengan Mekanisme Confidence Calibration

The high prevalence of feline and canine dermatological diseases in Indonesia, compounded by significant visual similarities across pathological conditions, frequently induces misdiagnoses. To address the limitations of conventional detectors regarding small lesions, this study develops a two-stage lesion-aware pipeline utilizing YOLOv12 for localization and EfficientNetV2-S for classification as an AI-integrated veterinary dermatology screening tool. The system is enhanced with temperature scaling for model confidence calibration and deployed via a Flutter/FastAPI-based mobile application. Evaluation results demonstrate a mean average precision (mAP@0.5) of 0.89 for felines and 0.64 for canines, with weighted classification F1-scores reaching 0.9976 and 0.9871, respectively. Notably, calibration (T=1.3174) successfully reduced the Expected Calibration Error (ECE) by 32.4% to 0.0209, satisfying the functional target (ECE < 0.15) with a mean system response time of 4.97 seconds. Usability evaluation yielded an "Excellent" score (85), whereas expert assessments validated the user interface quality (4.44/5.00). This calibrated pipeline proves viable as a first-line screening instrument, with recommendations for expanding disease coverage and conducting validation on independent test sets prior to clinical adoption.