Abstract
Accurate differentiation between actinic keratosis (AK) and cutaneous squamous cell carcinoma (cSCC) is crucial for effective treatment planning. While histopathology remains the gold standard, routine biopsy is often impractical for several reasons and dermoscopic evaluation is limited by overlapping features that lead to diagnostic uncertainty, even among experienced dermatologists. Artificial intelligence (AI), particularly convolutional neural networks (CNNs), has emerged as a powerful tool for automating image-based diagnosis in dermatology, achieving promising results in lesion classification. However, most of the existing models rely solely on raw images, overlooking the dermoscopic features that guide clinical reasoning. We developed a CNN-based model designed to classify AK versus cSCC in situ using dermoscopic images, integrating a dual-branch architecture that combines an EfficientNetB0 backbone for RGB inputs with a lightweight convolutional branch for two additional channels generated through targeted preprocessing to enhance vascular and keratinization patterns. Our dataset comprised 2,000 images, expanded through geometric and deep learning-based augmentation, exposing the model to nearly 200,000 training instances across epochs. Using repeated hold-out validation across 10 iterations, our best-performing model achieved an accuracy of 98.61%, sensitivity of 98.33%, specificity of 98.90%, precision of 98.90%, F1‑score of 98.61% and loss of 0.3120. These results surpass previously reported models for this task, demonstrating that incorporating clinically informed preprocessing significantly improves CNN performance. This approach represents a step toward clinically aligned AI systems capable of supporting dermatologists in differentiating between AK and cSCC with greater confidence and precision.