Quantifying acceptable artefact ranges for dermatologic classification algorithms.

Area of Science:

• Dermatology
• Artificial Intelligence
• Medical Imaging

Background:

• Machine learning classifiers are used to differentiate skin lesions.
• Artefacts in images can negatively impact classifier performance.
• Quantitative data on how blur, color, and lighting variations affect accuracy is lacking.

Purpose of the Study:

• To develop a system for measuring the impact of artefacts on machine learning accuracy.
• To identify the most detrimental artefacts in dermatologic images.
• To demonstrate methods for assessing classifier accuracy with artefact-containing images.

Main Methods:

• Artefacts were systematically introduced into dermatologic images.
• Two convolutional neural networks were trained for binary classification tasks.
• The impact of induced artefacts on diagnostic accuracy was measured using area under the curve and specificity-at-a-given-sensitivity.

Main Results:

• General blur most negatively affected the melanoma versus other classification task.
• Blue hue shifts had a greater impact on the suspicious versus follow task.
• Different artefacts had varying effects depending on the classification task.

Conclusions:

• Classifiers must either mitigate or detect image artefacts.
• Images with artefacts exceeding a quality threshold should be excluded from AI-driven diagnosis.
• Failure to address artefacts will reduce accuracy and hinder regulatory approval.