Related Experiment Video
Updated: Sep 29, 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
2.5K
A Non-Invasive Interpretable Diagnosis of Melanoma Skin Cancer Using Deep Learning and Ensemble Stacking of Machine
Iftiaz A Alfi1, Md Mahfuzur Rahman2,3, Mohammad Shorfuzzaman4
1Department of Electrical and Computer Engineering, North South University, Dhaka 1229, Bangladesh.
Diagnostics (Basel, Switzerland)
|March 25, 2022
Summary
This study introduces an interpretable deep learning and ensemble stacking method for accurate melanoma skin cancer detection. The interpretable approach aids dermatologists by providing visual heatmaps for better understanding of diagnostic results.
Area of Science:
- Dermatology and Medical Imaging
- Artificial Intelligence in Healthcare
- Computational Biology
Background:
- Skin lesions require early detection for improved patient survival rates.
- Current deep learning and machine learning models face challenges with image occlusions and imbalanced datasets, compromising diagnostic accuracy.
- Melanoma, a type of skin cancer, necessitates reliable and interpretable diagnostic tools.
Purpose of the Study:
- To develop an interpretable, non-invasive diagnostic method for melanoma skin cancer.
- To enhance the accuracy of skin lesion classification by combining deep learning and ensemble machine learning models.
- To provide dermatologists with understandable model predictions through interpretability techniques.
Main Methods:
- Utilized a balanced dataset of benign and malignant skin moles.
- Employed hand-crafted features for base machine learning models (logistic regression, SVM, random forest, KNN, gradient boosting machine).
- Implemented ensemble stacking for machine learning models and transfer learning with pre-trained deep learning models (MobileNet, Xception, ResNet50, ResNet50V2, DenseNet121).
- Incorporated Shapley Additive exPlanations (SHAP) for generating heatmaps to identify critical image regions.
Main Results:
- Evaluated individual deep learning models and assessed various ensemble combinations.
- Achieved improved classification performance through ensemble stacking and transfer learning.
- Generated heatmaps via SHAP, highlighting image regions indicative of melanoma, enhancing model interpretability.
- Performance metrics included accuracy, F1-score, Cohen's kappa, confusion matrix, and ROC curves.
Conclusions:
- The proposed interpretable method effectively diagnoses melanoma skin cancer.
- Ensemble stacking and transfer learning significantly improve classification accuracy.
- The interpretability approach provides valuable insights for dermatologists, facilitating clinical decision-making.
