Development of Skin Lesion Classification System Based on Watershed Algorithm and Custom CNN

---

Abstract

Skin cancer poses a substantial health threat worldwide, underscoring the importance of timely identification and precise categorization for effective treatment. This paper presents a comprehensive framework for skin cancer image analysis, encompassing preprocessing, segmentation, and classification of nine types of skin cancer using a custom Convolutional Neural Network (CNN) model. The preprocessing stage addresses missing or corrupted regions in images using inpainting algorithms, enhancing image quality and completeness. After that, a dataset obtained from the International Skin Imaging Collaboration (ISIC) on Kaggle is segmented using the watershed technique. Accurately identifying tumor locations throughout the segmentation process makes feature extraction for classification easier later on. Deep learning techniques play a pivotal role in this framework, enabling automatic feature extraction and learning from raw image data. The use of CNNs enables the development of a personalized model specifically designed to accurately identify nine types of skin cancer. The evaluation process of our skin cancer classification system includes the assessment of various metrics, such as predicting cancer types and confidence scores. Through this evaluation, the model assigns prediction values to all nine types of skin cancer, ultimately identifying the specific type based on the highest prediction value. This comprehensive assessment highlights the system’s potential as a valuable aid for dermatologists in accurately diagnosing and effectively treating skin cancer cases.

---

Keywords

Skin cancer, Dataset, Preprocessing, Segmentation, CNN, Evaluation metrics