SKIN CANCER DETECTION USING DEEP MACHINE LEARNING

Abstract

Background: Skin cancer is one of the most prevalent malignancies worldwide, responsible for nearly one-third of all diagnosed cancers each year. Early detection significantly improves survival rates, yet access to dermatological care remains limited in remote and resource-poor regions. Deep learning (DL) techniques, particularly convolutional neural networks (CNNs), have shown remarkable success in automating image-based diagnosis, offering an opportunity to bridge gaps in timely cancer detection and care.

Objective: The objective of this study was to develop and evaluate a deep learning–based model capable of accurately distinguishing between benign and malignant skin lesions using dermoscopic images, while addressing issues of dataset imbalance and model generalization.

Methods: A total of 3,000 dermoscopic images representing four major skin cancer types were used, with data augmentation applied through rotation, flipping, and scaling to improve robustness. Preprocessing steps included noise reduction, normalization, and lesion segmentation to isolate the region of interest. A lightweight CNN architecture incorporating convolution, max pooling, dropout, and batch normalization layers was employed. The dataset was divided into training and validation subsets (80:20). Model performance was assessed using precision, recall, F1-score, and accuracy metrics. Training was conducted on TensorFlow and Keras frameworks, and statistical evaluation was performed using SPSS v26.

Results: The model achieved an overall accuracy of 94.4%, with precision and recall values of 0.94 and 1.00, respectively, for malignant lesions. The F1-score reached 0.97, indicating excellent balance between sensitivity and specificity. Validation accuracy peaked at 90%, demonstrating effective learning with minimal overfitting.

Conclusion: The proposed deep learning model exhibited strong performance and computational efficiency, proving suitable for real-world deployment in clinical and teledermatology applications. Its scalability for mobile platforms highlights its potential to enhance early diagnosis and improve outcomes for patients in underserved regions.