Dermoscopic dark corner artifacts removal: Friend or foe?

Area of Science:

• Dermatology
• Artificial Intelligence
• Medical Imaging

Background:

• Artifacts in dermoscopic images pose challenges for skin cancer classification.
• Previous methods to remove artifacts yielded inconclusive results due to limited labeled data.

Purpose of the Study:

• Investigate the impact of dark corner artifacts on deep learning binary classification for skin cancer.
• Evaluate strategies to mitigate artifact influence on model performance.

Main Methods:

• Labeled 10,250 skin lesion images, creating a balanced dataset.
• Utilized a training set (6126 images) without artifacts and a testing set (4124 images) with artifacts.
• Conducted experiments using inpainted and synthetically generated artifact examples.

Main Results:

• Superimposing synthetic dark corner artifacts onto the training set improved model performance, especially the true negative rate.
• Deep learning models learned to disregard artifacts when they were present in training data.
• Developed a novel heatmap quantification method using root mean square brightness intensity.

Conclusions:

• Proposed artifact methods can mitigate the impact of dark corner artifacts on model performance.
• The new heatmap analysis aids in understanding the relationship between network focus and performance metrics.