An Improved Skin Lesion Classification Using a Hybrid Approach with Active Contour Snake Model and Lightweight...

Area of Science:

• Dermatology
• Artificial Intelligence
• Medical Imaging

Background:

• Skin cancer diagnosis is critical for patient survival but faces challenges like misinterpretation and time delays.
• Accurate classification of skin lesions is difficult due to their complex nature and visual variability.
• Current deep learning models struggle with border delineation, spatial feature connections, and contextual information for classification.

Purpose of the Study:

• To develop a novel deep learning approach for improved skin lesion classification.
• To enhance the accuracy and generalization capabilities of automated dermatological diagnosis.
• To provide a robust tool for medical practitioners in identifying skin malignancies.

Main Methods:

• A hybrid model combining active contour (AC) segmentation, ResNet50 feature extraction, and a capsule network with attention mechanisms.
• Utilized stochastic gradient descent (SGD) for model parameter optimization.
• Implemented and evaluated on the HAM10000 and ISIC 2020 public datasets.

Main Results:

• Achieved a diagnostic accuracy of 98% and an Area Under the Receiver Operating Characteristic Curve (AUC-ROC) of 97.3%.
• Demonstrated superior model generalization compared to existing state-of-the-art (SOTA) methods.
• Effectively delineated lesion borders and utilized contextual information for enhanced classification.

Conclusions:

• The proposed model shows significant potential for reshaping automated dermatological diagnosis.
• The approach offers a reliable and accurate tool for early skin cancer detection.
• Highlights the effectiveness of integrating AC segmentation, ResNet50, and capsule networks with attention for medical image analysis.