Classification of Skin Cancer Lesions Using Explainable Deep Learning

Area of Science:

• Medical imaging
• Artificial intelligence in healthcare
• Dermatology

Background:

• Skin cancer is a major global health concern.
• Traditional detection methods can be time-consuming.
• Computer-aided diagnostic systems offer efficient solutions.

Purpose of the Study:

• To enhance deep learning models for effective skin cancer detection.
• To improve early diagnosis of both benign and malignant skin lesions.
• To compare the performance of modified models against original pre-trained networks.

Main Methods:

• Modification of pre-trained MobileNetV2 and DenseNet201 models.
• Addition of three convolutional layers to each model.
• Evaluation of detection performance for benign and malignant skin classes.

Main Results:

• Modified models outperformed original pre-trained versions.
• Modified DenseNet201 achieved 95.50% accuracy.
• Modified DenseNet201 demonstrated high sensitivity (93.96%) and specificity (97.03%).

Conclusions:

• The proposed Modified DenseNet201 model offers state-of-the-art performance for skin cancer detection.
• Deep learning model enhancements can significantly improve diagnostic accuracy.
• This approach can assist dermatologists in timely and accurate skin cancer diagnosis.