Skin Diseases Detection and Diagnosis Support System Using Yolov12s and Late Fusion Technique

Abstract

Millions of people worldwide suffer from dermatological conditions, but in areas with limited resources, such as Pakistan, access to prompt diagnosis is still restricted. In order to increase diagnostic reliability, this study suggests a multimodal framework for the detection of skin diseases that integrates patient-reported symptoms with visual lesion analysis. Using a late fusion approach that combines a Logistic Regression classifier trained on structured symptom features with a YOLOv12s-based object detection model for lesion localization, the system targets 16 common skin conditions. While the symptom model encodes clinical indicators such as redness, itching, and pustules, the visual model captures discriminative lesion patterns. To handle visually ambiguous situations that are challenging for Image-only approaches, outputs from both modalities were combined at the decision level. Experiments on a multimodal dataset show that the proposed fusion framework outperforms unimodal baselines in terms of accuracy and F1-score. The robustness of the proposed method under realistic dataset conditions is demonstrated through a comparison with previous research. The findings indicate that multimodal late fusion improves the performance of skin disease screening, making it suitable for tele dermatology and initial clinical decision support applications in low-resource environments.