A hybrid deep learning and cellular automata framework with fractional derivatives for skin type and skin disease classification.

INTRODUCTION: A hybrid deep learning and mathematical modeling approach to automatic identification of skin type and multi-class classification of skin diseases based on dermatological images are proposed in this paper. A coherent mechanism to combine deep features learning and texture modeling approaches mathematically is a necessity for improved skin type classification and skin diseases categorizations.

METHODS: We propose to combine convolutional neural networks, cellular automata and fractional-order derivatives in one technique. Beside CNN models based on transfer learning, this system makes use of cellular automata with online tessellation to characterize explicitly the dynamic evolution of skin textures. Specifically, a 2-dimensional Moore neighborhood cellular automata based on both totalistic and outer-totalistic rules to illustrate the local relations of neighboring pixels, the smooth aspect of skin textures, and the oily/dry distribution and pore density, which are indicators for characterizing dry, normal, and oily skin type. The models of automata characterize skin by regions and make the texture analysis stable by iterating over tessellated images. In feature extraction part, fractional-order derivatives are used for better sensitivity to edge continuity and detailed texture variations. Many types of fractional-order derivatives can be derived but the Grnwald-Letnikov and the Caputo fractional derivatives were selected for their capability to be used with discrete image grids and long-range relationship modeling. They enable better detection of delicate texture patterns that can be otherwise mistaken by traditional CNN models through an improved multiscale representation. The cellular automata-based and fractional features are combined with deep features taken from fine-tuned ResNet topologies, yielding a robust hybrid representation.

RESULTS & DISCUSSION: The suggested method significantly reduces the misclassification of normal skin and increases skin type classification accuracy by approximately 1.2 percentage points, according to experimental evaluation on publicly available benchmark skin datasets. The integrated model achieved an accuracy of 92.8%, sensitivity of 91.4%, and F1-score of 91.7% for skin disease classification of five common dermatological conditions, while achieving an accuracy of 92.4%, sensitivity of 91.1%, and F1-score of 91.4% for skin type classification. The proposed framework therefore offers an effective and scalable approach for intelligent dermatological assessment and skin image analysis.