Generalizing common tasks in automated skin lesion diagnosis

Area of Science:

• Dermatology and Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Automated skin lesion diagnosis is crucial for early detection and treatment.
• Existing methods often focus on single tasks, lacking a unified approach.
• Dermoscopic imaging provides detailed views of skin structures, aiding diagnosis.

Purpose of the Study:

• To introduce a versatile supervised learning model for multiple automated skin lesion diagnosis tasks.
• To evaluate the model's performance in skin lesion segmentation, hair detection, and pigment network identification.
• To demonstrate the model's capability in generating diagnostic visualizations and confidence metrics.

Main Methods:

• A general model employing supervised learning and Maximum A Posteriori (MAP) estimation.
• Application to skin lesion segmentation, hair occlusion detection, and dermoscopic pigment network identification.
• Utilizing probabilistic outputs for Receiver Operating Characteristic (ROC) curve generation and confidence intervals.

Main Results:

• Quantitative results for segmentation and hair detection are competitive with specialized methods.
• Successful and compelling visualizations of dermoscopic pigment networks were generated.
• Probabilistic outputs enabled ROC curve analysis and provided confidence intervals for segmentations.

Conclusions:

• The developed general model offers a unified and effective approach to multiple automated skin lesion diagnosis tasks.
• The model's performance is comparable to specialized techniques, with added benefits of probabilistic outputs.
• This approach shows promise for enhancing the accuracy and interpretability of computer-aided diagnosis in dermatology.