A rotation meanout network with invariance for dermoscopy image classification and retrieval

Area of Science:

• Medical imaging analysis
• Artificial intelligence in dermatology
• Computational pathology

Background:

• Computer-aided diagnosis (CAD) systems aid clinical skin disease diagnosis.
• Convolutional neural networks (CNNs) excel at extracting features from dermoscopy images.
• CNNs lack rotation invariance, limiting their robustness with rotated skin lesion images.

Purpose of the Study:

• To propose a rotation meanout (RM) network for extracting rotation-invariant features from dermoscopy images.
• To enhance the robustness of CNNs against image rotations in skin lesion analysis.
• To improve the performance of classification and retrieval tasks for dermoscopy images.

Main Methods:

• Developed a rotation meanout (RM) network integrating weight-sharing convolutions and a meanout strategy.
• The RM network generates rotation-equivariant feature maps, yielding rotation-invariant features post-global average pooling (GAP).
• RM is a parameter-free, structure-agnostic operation embeddable within existing CNN architectures.

Main Results:

• The RM network demonstrated theoretical rotation-equivariance and practical rotation-invariant feature extraction.
• Experiments on a dermoscopy dataset showed superior performance compared to other anti-rotation methods.
• Significant improvements were observed in skin disease classification and image retrieval tasks.

Conclusions:

• Rotation invariance is crucial for robust dermoscopy image analysis.
• The proposed RM network offers a generalizable and effective solution for incorporating rotation invariance into CNNs.
• This approach holds significant potential for advancing computer-aided diagnosis in dermatology.