A Multiresolution Convolutional Neural Network with Partial Label Training for Annotating Reflectance Confocal...

Area of Science:

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

Background:

• Skin cancers, particularly melanoma, are a significant global health concern.
• Reflectance confocal microscopy (RCM) offers non-invasive, high-resolution imaging for skin cancer diagnosis.
• Interpreting RCM images is challenging due to complexity, low contrast, variability, and requires extensive expert training.

Purpose of the Study:

• To develop an automated method for annotating key diagnostic patterns in RCM images of human skin.
• To improve the accuracy and efficiency of skin cancer diagnosis using RCM.
• To facilitate clinical training and adoption of RCM technology.

Main Methods:

• A novel multiresolution 'nested encoder-decoder' convolutional network architecture was designed.
• A selective loss function was implemented to address partially labeled images.
• The network was trained and validated on large (12k × 12k pixels), partially labeled RCM images of melanoma-suspicious skin lesions.

Main Results:

• The developed network achieved high sensitivity and specificity in automatically annotating diagnostic morphological patterns in RCM images.
• The system provided consistent annotations for unlabeled image sections.
• The approach effectively handled challenges like large image size, pattern scale variance, and class imbalance.

Conclusions:

• The nested encoder-decoder network provides an effective tool for automated RCM image annotation, aiding skin cancer diagnosis.
• This technology can significantly reduce the time and expertise needed for RCM interpretation, accelerating clinical adoption.
• The multiresolution deep network architecture may have broader applications in biomedical image analysis.