Abstract:

To improve patient outcomes, skin cancer and dermatological disorders must be accurately and early detected. These conditions are becoming a major global health concern. While conventional convolutional neural networks (CNNs) have shown impressive results in medical image analysis, they frequently have difficulty striking a balance between high diagnostic performance and computational efficiency. The Fibonacci sequence and golden ratio are incorporated into convolutional filter scaling, kernel sizing, and learning rate decay in this paper's novel Fibonacci-Scaled CNN (Fib-CNN) architecture. By simulating natural growth patterns, this bio-inspired method seeks to enhance feature extraction in dermoscopic imaging and facilitate more effective receptive field expansion. The suggested Fib-CNN was tested using the Private dermoscopic dataset and compared to the most advanced CNN architectures, such as MobileNetV2, ResNet50, and VGG16. The results show that Fib-CNN outperforms baseline models while lowering computational complexity by up to 22%, achieving 90.8% accuracy, 0.907 precision, 0.911 recall, and a ROC-AUC score of 0.958. Fib-CNN reduces background interference and creates more focused lesion attention maps, as shown by Grad-CAM visualizations. The results show that Fibonacci scaling improves the depth-to-width ratio of convolutional layers, which is advantageous for the efficiency and interpretability of the model. This establishes Fib-CNN as a viable instrument for diagnosing skin diseases in real time with limited resources, which could be implemented on edge devices for teledermatology applications.