SkinToneNet: a robust optimised cascaded multi-scale residual attention network for accurate psoriasis and vitiligo detection across diverse skin types

Accurate detection of chronic skin diseases like Psoriasis and Vitiligo remains challenging due to significant variations in skin pigmentation and lesion presentation across different populations. This paper introduces SkinToneNet, a comprehensive framework designed for robust dermatological diagnosis across diverse skin types. The core methodological contributions include a novel hybrid optimisation algorithm (APVCO) that combines the strengths of Volleyball Premier League and Chimp Optimisation for effective hyperparameter tuning in medical image analysis. Additionally, we propose the CMR-GRU architecture, which cascades Multi-Scale Residual Attention Networks with Gated Recurrent Units to capture both spatial hierarchies and sequential dependencies in skin lesion patterns. The framework integrates optimised segmentation using Adaptive TransUNet with optimised classification via CMR-GRU, both fine-tuned using APVCO. Experimental validation demonstrates that SkinToneNet achieves segmentation Dice scores of 0.894 and IoU of 0.812, with classification accuracy of 95.17% for Psoriasis and 95.19% for Vitiligo across Fitzpatrick skin types I-VI. The system maintains specificity above 93.05% and sensitivity above 93.15% for all skin types, demonstrating consistent performance. The work establishes a methodological foundation for skin-type-agnostic dermatological image analysis while addressing critical challenges in automated diagnosis of Psoriasis and Vitiligo.