Related Experiment Video
Updated: Oct 16, 2025
Combining Reflectance Confocal Microscopy with Optical Coherence Tomography for Noninvasive Diagnosis of Skin Cancers via Image Acquisition
Published on: August 18, 2022
Melanoma classification using light-Fields with morlet scattering transform and CNN: Surface depth as a valuable tool
Pedro M M Pereira1, Lucas A Thomaz2, Luis M N Tavora3
1Instituto de Telecomunicações, Morro do Lena - Alto do Vieiro, Leiria 2411-901, Portugal; University of Coimbra, Centre for Informatics and Systems of the University of Coimbra, Department of Informatics Engineering, Pinhal de Marrocos, Coimbra 3030-290, Portugal.
This study enhances melanoma detection by incorporating skin surface depth (3D) with traditional 2D images. Combining both dimensions significantly improves classification accuracy for medical image analysis.
Area of Science:
- Dermatology
- Medical Imaging
- Computer Vision
Background:
- Learning-based approaches are vital for medical image classification, particularly for melanoma detection.
- Current methods predominantly utilize 2D dermoscopic or RGB images, potentially overlooking crucial surface characteristics.
Purpose of the Study:
- To investigate the utility of incorporating 3D depth information from light-field images into skin lesion classification.
- To compare classification performance using 2D data, 3D data, or a combination of both.
Main Methods:
- A processing pipeline utilizing morlet scattering transform and a Convolutional Neural Network (CNN) model was developed.
- The pipeline was used to analyze 2D images, 3D depth data, and combined 2D/3D data from the SKINL2 dataset.
Main Results:
- Classification accuracy for Melanoma vs. Nevus reached 84.00% (2D only), 74.00% (3D only), and 94.00% (2D and 3D).
- Incorporating 3D depth significantly increased sensitivity and specificity compared to 2D data alone.
- In a more imbalanced setting (Melanoma vs. all other lesions), 3D data integration also showed notable improvements in sensitivity.
Conclusions:
- The integration of 3D depth information from light-field images offers significant advantages over conventional 2D image analysis for skin lesion classification.
- This multimodal approach demonstrates superior performance in discriminating melanoma, highlighting the value of characterizing skin surface rugosity.
