Ros-NET: A deep convolutional neural network for automatic identification of rosacea lesions
Hamidullah Binol1, Alisha Plotner2, Jennifer Sopkovich2
1Center for Biomedical Informatics, Wake Forest School of Medicine, Winston-Salem, NC, USA.
Summary
A new computer-aided diagnosis system, Ros-NET, accurately identifies rosacea lesions using deep learning. This quantitative approach overcomes subjective diagnostic variability in facial skin conditions.
Area of Science:
- Dermatology
- Medical Imaging
- Artificial Intelligence
Background:
- Rosacea is a common skin condition with subjective diagnosis.
- Current diagnostic methods for rosacea have high observer variability.
- Objective assessment of rosacea is needed to improve patient outcome evaluation.
Purpose of the Study:
- To develop a quantitative and reproducible computer-aided diagnosis system for rosacea.
- To improve the accuracy of rosacea lesion identification.
- To overcome the limitations of subjective human assessment in rosacea diagnosis.
Main Methods:
- Developed Ros-NET, a computer-aided diagnosis system using Inception-ResNet-v2 and ResNet-101.
- Integrated multi-scale and multi-resolution image information for feature extraction.
- Utilized facial landmarks to define regions of interest for refined detection.
Main Results:
- Achieved high weighted average Dice coefficients using leave-one-patient-out cross-validation.
- Inception-ResNet-v2 achieved 89.8% ± 2.6% and ResNet-101 achieved 87.8% ± 2.4% accuracy.
- Demonstrated the effectiveness of deep learning models in identifying rosacea lesions.
Conclusions:
- Pre-trained networks with transfer learning are beneficial for rosacea lesion identification.
- The Ros-NET system offers a quantitative and reproducible diagnostic approach.
- Future work will focus on expanding the database for broader applicability.
Keywords:
computer-assisted diagnosisconvolutional neural networksdeep learningrosaceasemantic segmentationtransfer learning
