Unsupervised SoftOtsuNet Augmentation for Clinical Dermatology Image Classifiers.

Area of Science:

• Artificial Intelligence
• Computer Vision
• Medical Imaging

Background:

• Data augmentation is vital for machine learning (ML) model accuracy and reducing overfitting in deep neural networks (DNNs).
• Clinical dermatology images frequently contain irrelevant background details, which can introduce biases into DNNs.
• Supervised segmentation of foreground/background in dermatology images is costly due to extensive labeling requirements.

Purpose of the Study:

• To develop a novel unsupervised deep neural network (DNN) for data augmentation in clinical dermatology.
• To address the challenge of irrelevant background information in dermatology images that can bias DNN models.
• To propose a cost-effective solution that avoids manual labeling for image segmentation.

Main Methods:

• An unsupervised DNN was developed incorporating a differentiable adaptation of Otsu's Method.
• CutOut augmentation was combined with the Otsu's Method adaptation for dynamic masking.
• The proposed SoftOtsuNet was evaluated on multiple dermatology image datasets.

Main Results:

• SoftOtsuNet demonstrated superior performance compared to other augmentation methods across three datasets.
• Improvements included 0.75% on Fitzpatrick17k, 1.76% on Diverse Dermatology Images, and 0.92% on a proprietary dataset.
• The method only impacts training time, leaving inference costs unaffected.

Conclusions:

• Unsupervised, human-engineered loss functions can still enhance large, data-driven models.
• SoftOtsuNet offers an effective and efficient approach to data augmentation in medical imaging.
• This method mitigates background-induced bias in DNNs without increasing computational inference costs.