|
|
 |
| |
|
|
|
AMDS: A Multistage Framework Using Deep Learning Models for Early Diagnosis of Melanoma and Non-Melanoma Skin Lesions |
|
|
|
PP: 791-807 |
|
|
|
doi:10.18576/isl/130405
|
|
|
|
Author(s) |
|
|
|
Afnan M. Hassan,
Khaled Mahar,
Mohamed M. Fouad,
|
|
|
|
Abstract |
|
|
| Skin cancer, an aggressive cancer with a global frequency, presents an evolving public health challenge requiring novel diagnostic approaches. The traditional method for diagnosing skin cancer involves a thorough examination of tissue samples obtained from skin lesions. Healthcare professionals must perform the complex task of identifying specific early symptoms to make an accurate diagnosis. Early detection of skin cancer is particularly challenging because of the tendency for misdiagnosis due to similarities with other dermatological conditions and variations in specialist expertise. Researchers have used machine learning algorithms to improve the performance of numerous medical applications in recent years to improve the dependability, productivity, efficiency, predictability, and precision of medical diagnostics. Current research presents a Multistage Deep Learning model for Skin Cancer Classifier (AMDS), a framework designed to improve the early detection of melanoma and non-melanoma skin lesions. The AMDS consists of several crucial phases, beginning with precise preprocessing techniques to remove extraneous components surrounding skin lesions. Given the inherent imbalances within most skin cancer datasets, the subsequent stage employs Generative Adversarial Networks (GANs) to generate synthetic images for enhancing dataset diversity and equip the classifier to handle a broad spectrum of skin lesions. In the subsequent stage, an attention-based U-Net model is introduced that is capable of generating masks for regions of interest while removing background noise. The process ended with the classification stage which uses distinct forms of the cutting-edge EfficientNet, ResNet, and DenseNet architectures, carefully trained using the segmented images, to find the best model for skin lesion classification. The proposed deep learning models are systematically evaluated by utilizing the International Skin Imaging Collaboration dataset (ISIC), a dermatology benchmark. The experimental results demonstrate that the proposed framework using a modified EfficientNetV2S with attention mechanism outperforms other tested architectures as well as most recent research. Notably, it achieves a 0.96 accuracy rate, 0.91 F1- score, 0.90 recall rate, and 0.93 precision rate on the test benchmarking datasets. These results highlight the importance of the proposed multistage framework as a potential transformative instrument for early skin cancer detection.
|
|
|
|
|
|
|
|
