Related Experiment Video
Updated: Jun 24, 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
2.3K
Comparison of Three Deep Learning Models in Accurate Classification of 770 Dermoscopy Skin Lesion Images
Abdulmateen Adebiyi1, Praveen Rao1, Jesse Hirner2
1Department of Electrical Engineering and Computer Science, University of Missouri.
Summary
Deep learning models can classify skin lesions from dermoscopy images. DenseNet121 showed the best performance, aiding in early skin cancer diagnosis.
Area of Science:
- Dermatology
- Medical Imaging
- Artificial Intelligence
Background:
- Accurate skin cancer diagnosis is vital for early detection and treatment.
- Dermoscopy images are crucial for visual assessment of skin lesions.
- Deep learning offers potential for automated image analysis in dermatology.
Purpose of the Study:
- To evaluate deep learning models for classifying benign and malignant skin lesions using dermoscopy images.
- To compare the performance of ResNet50, DenseNet121, and Inception-V3 models.
- To assess the impact of a hair removal algorithm on classification accuracy.
Main Methods:
- Acquired 770 de-identified dermoscopy images from MU Healthcare.
- Created three datasets: original images and images with hair removal.
- Trained and evaluated ResNet50, DenseNet121, and Inception-V3 models using accuracy and AUC ROC metrics.
Main Results:
- DenseNet121 achieved the highest accuracy (80.52%) and AUC ROC (0.81) on the hair-removed dataset.
- This model demonstrated a sensitivity of 0.80 and specificity of 0.81.
- SHAP values were analyzed for model interpretability.
Conclusions:
- Deep learning, particularly DenseNet121 with hair removal, shows promise for accurate skin lesion classification.
- The findings support the use of AI in improving early skin cancer diagnosis.
- Further research can explore interpretability methods like SHAP for clinical trust.
