Lescnnet: A Sequence-Guided Convolutional Framework for High-Accuracy Skin Cancer Recognition

Abstract

Early and reliable identification of malignant skin lesions remains a central challenge in dermatological practice,
particularly when visual differences between benign and cancerous tissues are subtle and evolve over time. While
conventional convolutional neural networks have shown promise in dermoscopic image analysis, they often treat each image as an isolated sample, overlooking the progressive and structured nature of lesion appearance. To address this limitation, this study presents LesCNNet, a sequence-guided convolutional framework designed to capture both spatial detail and sequential lesion behavior for improved skin cancer recognition.LesCNNet integrates deep convolutional feature extraction with a sequence modeling mechanism that learns how lesion characteristics change across ordered image representations. Instead of relying solely on static texture and color cues, the proposed architecture encodes the continuity and variation of lesion patterns, enabling more stable and discriminative feature learning. The framework is further enhanced through adaptive feature fusion and optimized training strategies to reduce overfitting and improve generalization across diverse skin types and imaging conditions.The proposed model is evaluated on publicly available dermoscopic datasets widely used in skin cancer research. Extensive experiments demonstrate that LesCNNet consistently outperforms conventional CNN-based classifiers and recent deep learning approaches in terms of accuracy, sensitivity, specificity, and F1-score. The results indicate that incorporating sequence-guided learning significantly improves the model’s ability to distinguish malignant lesions from visually similar benign cases, particularly in early-stage and lowcontrast scenarios.The findings suggest that LesCNNet provides a reliable and clinically meaningful tool for automated skin cancer screening. By combining spatial feature learning with sequence-aware modeling, the proposed framework offers a more realistic representation of lesion characteristics, supporting dermatologists in early diagnosis and reducing
the risk of misclassification. This work contributes a robust and scalable deep learning solution that moves automated dermatological analysis closer to real-world clinical deployment.