Source-linked AI summary
Skin lesion detection based on an ensemble of deep convolutional neural network
Balazs Harangi
TL;DR
Skin-lesion classification requires automated methods for clinically important melanoma and related lesion categories. The paper proposes an ensemble of four DCNNs that combines accuracy-weighted classifier outputs with softmax confidence values, achieving an overall validation score of 0.932 and outperforming its individual components.
Problem
Automated classification is sought for dermoscopy images containing melanoma, nevus, and seborrheic keratosis, supporting assessment of clinically important skin lesions.
Method
The method ensembles GoogLeNet, ResNet-50, AlexNet, and VGG-VD-16, combining individual accuracies and softmax confidence values through weighted majority voting.
Results
0.932 overall score was achieved on the 150-image validation set, and the ensemble outperformed the individual DCNN components in skin-lesion classification.
Takeaways & Limitations
Accuracy- and confidence-weighted fusion provided the paper’s reported improvement over the individual neural networks.
Takeaways & Limitations
The available training set contained only 2,000 manually annotated images, and the class counts were not sufficiently large for learning DCNNs.
Abstract
from arXiv · showhide
Skin cancer is a major public health problem, with over 5 million newly diagnosed cases in the United States each year. Melanoma is the deadliest form of skin cancer, responsible for over 9,000 deaths each year. In this paper, we propose an ensemble of deep convolutional neural networks to classify dermoscopy images into three classes. To achieve the highest classification accuracy, we fuse the outputs of the softmax layers of four different neural architectures. For aggregation, we consider the individual accuracies of the networks weighted by the confidence values provided by their final softmax layers. This fusion-based approach outperformed all the individual neural networks regarding classification accuracy.
1. INTRODUCTION
The paper addresses automated classification of dermoscopy images for melanoma and other skin lesions, motivated by the clinical importance of skin cancer and expanding access to imaging devices.
- Dermoscopy removes surface reflection, exposing more visual information from deeper skin layers for lesion assessment.
- Automated lesion assessment could support patient care as mobile dermatoscopes become more available for smartphones.
- The 2017 ISBI challenge required classification into melanoma, nevus, and seborrheic keratosis.
- The challenge dataset contained 2,000 training images, 150 validation images, and 600 test images.
2. PROPOSED METHODLOGY
The proposed method ensembles four DCNN architectures trained or fine-tuned for skin-lesion classification, then combines their predictions using accuracy- and confidence-weighted voting.
- Training-set extension: 14,300 augmented images were generated through cropping, horizontal flipping, and rotation of the training images.The augmented classes contained 8,200 nevus, 4,600 melanoma, and 1,500 seborrheic keratosis images.
- Network preparation: The system ensembles GoogLeNet, ResNet-50, AlexNet, and VGG-VD-16 DCNN architectures.
- Network preparation: GoogLeNet and ResNet-50 were fine-tuned from ImageNet weights, whereas AlexNet and VGG-VD-16 used randomly initialized weights.
- Training procedure: Training and fine-tuning used stochastic gradient descent for 500 epochs per neural network, with validation top-1 error monitored after each epoch.The reported validation curves did not increase after the CNN models were overfit.
- Ensemble decision rule: The ensemble first uses majority voting, then resolves ties and improves voting reliability with weights based on validation-set AUCs.
- Ensemble decision rule: Each classifier’s softmax confidence for a class is incorporated into the weighted voting scheme, and the class with the largest weighted sum is selected.
3. EXPERIMENTAL RESULTS
The ensemble was evaluated on the 150-image validation set using AUC-based measures for melanoma and seborrheic keratosis classification.
- 0.932 overall score was achieved by the ensemble on the 150-image validation set.The score was calculated as the average AUC for melanoma and seborrheic keratosis classification.
- Table I summarizes validation results using ACC, AP, SE, and SP for melanoma and seborrheic keratosis classification.M/SK denote melanoma/seborrheic keratosis; ACC/AP/SE/SP denote accuracy/average precision/sensitivity/specificity.
4. CONCLUSION
The paper concludes that an ensemble of DCNNs outperformed its individual components for skin-lesion classification using accuracy-weighted majority voting.
- The ensemble outperformed the accuracy of the individual DCNN components in skin-lesion classification.
- Final class labels were derived by incorporating individual DCNN accuracies into a weighted majority-voting decision rule.