Convolutional Neural Network Analytics of Melasma in Harmonically Generated Microscopy Images
Journal
2025 International Symposium on Intelligent Signal Processing and Communication Systems, ISPACS 2025
Start Page
1
End Page
6
ISBN (of the container)
979-833158066-7
Date Issued
2026-02-19
Author(s)
Abstract
Melasma is a common skin pigmentation disorder characterized by the overproduction of melanin by melanocytes and its subsequent transfer to neighboring keratinocytes. This paper presents a computer-aided diagnosis method for melasma classification. To quantify these pathological features, a Gabor filter bank was employed to capture texture patterns associated with the clinical categories and observed histopathological traits. Drawing on domain knowledge from computer vision, the initial kernels of the convolutional layers were designed based on Gabor feature extraction principles to provide a more informative starting point for the convolutional neural network (CNN). The CNN was then trained on higher harmonic generation microscopy (HHGM) images classified into four melasma subtypes, leveraging the enhanced initial representation to improve learning performance. Experimental results demonstrate that this approach enables accurate and efficient classification of melasma subtypes. Furthermore, integrating domain-inspired handcrafted kernels reduces training time and annotation effort, while also outperforming randomly initialized CNNs in classification accuracy. Experimental results achieved an accuracy of 89.97 % with sensitivity of 93.8%, 90.2%, 86.0%, and 72.0% for the four severity levels, respectively.
Event(s)
2025 International Symposium on Intelligent Signal Processing and Communication Systems, ISPACS 2025
Subjects
biomedical image
Convolutional Neural Network (CNN)
feature extraction
Gabor filter
Higher Harmonic Generation Microscopy (HHGM)
image segmentation
Melasma
optical in vivo virtual biopsy
Otsu's method
texture segmentation
Publisher
IEEE
Type
conference paper