Related Experiment Video
Updated: Aug 21, 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
2.4K
ChimeraNet: U-Net for Hair Detection in Dermoscopic Skin Lesion Images
Norsang Lama1, Reda Kasmi2, Jason R Hagerty3
1Missouri University of Science & Technology, Rolla, MO, 65409, USA.
Journal of Digital Imaging
|November 17, 2022
Summary
A new deep learning model, ChimeraNet, effectively detects and removes hair and ruler marks from skin lesion images. This method improves diagnostic accuracy by enhancing the visibility of critical network features in dermoscopic images.
Area of Science:
- Dermatology
- Medical Imaging
- Artificial Intelligence
Background:
- Hair and ruler marks in dermoscopic images hinder accurate segmentation and analysis of skin lesions.
- Removing these artifacts is challenging due to variations in hair thickness, color, and overlap with lesion textures.
Purpose of the Study:
- To develop a novel deep learning (DL) technique for detecting hair and ruler marks in skin lesion images.
- To improve the accuracy of subsequent diagnostic analyses by effectively removing image artifacts.
Main Methods:
- Proposed ChimeraNet, an encoder-decoder DL architecture utilizing EfficientNet and squeeze-excitation residual (SERes) structures.
- Evaluated the model on the HAM10000 (ISIC2018 Task 3) dataset at various image sizes.
- Utilized the Dice loss function for optimization.
Main Results:
- The largest image size (448x448) achieved the highest accuracy (98.23%) and Jaccard index (0.65) on the HAM10000 dataset.
- ChimeraNet outperformed U-Net and ResUNet-a in artifact detection.
- Achieved state-of-the-art accuracy on additional test images compared to eight classical methods.
Conclusions:
- The proposed ChimeraNet architecture demonstrates significant potential for improving the detection of fine structures in dermoscopic images.
- This DL technique can enhance the accuracy of skin lesion analysis by removing obstructive artifacts.
- Further research into DL for dermoscopy structure detection is recommended.
