Related Experiment Video
Updated: Dec 17, 2025
09:37
Combining Reflectance Confocal Microscopy with Optical Coherence Tomography for Noninvasive Diagnosis of Skin Cancers via Image Acquisition
Published on: August 18, 2022
2.8K
Automatic diagnosis of melanoma using hyperspectral data and GoogLeNet
Ginji Hirano1, Mitsutaka Nemoto2, Yuichi Kimura1
1Department of Biological System Engineering, Graduate School of Biology-Oriented Science and Technology, Kindai University, Wakayama, Japan.
Summary
This study developed a diagnostic system using hyperspectral data and a convolutional neural network for melanoma detection. The system achieved improved accuracy with data augmentation, highlighting its potential for early melanoma diagnosis.
Area of Science:
- Medical imaging
- Machine learning
- Dermatology
Background:
- Melanoma, a superficial tumor, has a poor prognosis in advanced stages, necessitating early detection.
- Current diagnostic methods require improvement for quantitative melanoma assessment.
- Hyperspectral data (HSD) offers potential for detailed tissue analysis.
Purpose of the Study:
- To develop a quantitative diagnostic system for melanoma using HSD and a convolutional neural network (CNN).
- To evaluate the system's diagnostic performance for melanoma detection.
Main Methods:
- Acquisition of HSD using a hyperspectral imager.
- Development of a CNN model (GoogLeNet) with a "Mini Network" layer to process 84 HSD channels into 3 input channels.
- Training and evaluation using 619 lesions (278 melanoma, 341 non-melanoma).
Main Results:
- The system achieved 72.7% accuracy without data augmentation.
- With data augmentation, accuracy improved to 77.2%, with sensitivity at 72.3% and specificity at 81.2%.
- 5-fold cross-validation was used for performance evaluation.
Conclusions:
- The developed HSD and CNN system shows promise for quantitative melanoma diagnosis.
- Data augmentation significantly enhances the system's diagnostic performance.
- Future work will focus on refining the network architecture and expanding the dataset.
