Automated Skin Lesion Classification on Ultrasound Images.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Dermatology

Background:

• Skin cancer incidence is rising, necessitating advanced diagnostic tools.
• Ultrasound imaging offers complementary data to optical dermoscopy for skin lesion analysis.
• Computer-aided diagnosis (CAD) systems are crucial for efficient and accurate skin cancer detection.

Purpose of the Study:

• To develop and evaluate a fully automated (FA) ultrasound-based classification framework for skin cancer.
• To compare the performance of FA segmentation with two semi-automated (SA) segmentation methods.
• To assess the impact of FA classification on diagnostic accuracy and efficiency.

Main Methods:

• Ultrasound recordings of 310 skin lesions (melanoma, basal cell carcinoma, benign nevi) were analyzed.
• A Support Vector Machine (SVM) model was trained on 62 features using ten-fold cross-validation.
• Classification performance was evaluated using receiver operating characteristic (ROC) area under the curve (AUC) and accuracy (ACC) for various class combinations.

Main Results:

• The FA framework achieved high performance, distinguishing nevi from cancerous lesions with AUCs over 90% and ACCs over 85% across all segmentation methods.
• FA classification demonstrated comparable or superior results to SA methods, with performance degradation not exceeding 5% when switching from SA to FA.
• The study highlights the potential of FA ultrasound classification for reducing diagnosis time and operator dependency.

Conclusions:

• Fully automated ultrasound-based classification is a viable and effective tool for skin cancer diagnosis.
• FA segmentation in ultrasound imaging does not compromise, and can potentially improve, classification accuracy compared to SA methods.
• This approach offers a promising avenue for more accessible and objective skin cancer screening and diagnosis.