Source-linked AI summary
Skin Lesion Classification Using Hybrid Deep Neural Networks
Amirreza Mahbod, Gerald Schaefer, Chunliang Wang, Rupert Ecker, Isabella Ellinger
TL;DR
Skin lesion classification is difficult because dermoscopic interpretation is subjective and conventional computer-aided methods can require extensive processing. The paper combines multilevel features from three pretrained CNNs with SVM classifiers and fuses their outputs, achieving very good performance on ISIC 2017 validation images. Its scope is limited by the small set of pretrained networks, limited training data, and possible information loss from small resized inputs.
Problem
Dermoscopic lesion classification is challenging because human interpretation is subjective and conventional methods often require segmentation, handcrafted features, and extensive preprocessing.
Method
The method extracts features from multiple layers of pretrained AlexNet, VGG16, and ResNet-18 models, trains SVM classifiers, and fuses their outputs.
Results
The fused multi-CNN approach outclasses single-network approaches and demonstrates significantly better classification performance on the evaluated skin-lesion tasks.
Takeaways & Limitations
Pretrained natural-image CNNs and fusion of features or outputs can provide very good classification results for dermoscopic skin lesions.
Takeaways & Limitations
The study examines a limited number of pretrained networks and training images, while very small resized patches may remove useful lesion information.
Abstract
from arXiv · showhide
Skin cancer is one of the major types of cancers with an increasing incidence over the past decades. Accurately diagnosing skin lesions to discriminate between benign and malignant skin lesions is crucial to ensure appropriate patient treatment. While there are many computerised methods for skin lesion classification, convolutional neural networks (CNNs) have been shown to be superior over classical methods. In this work, we propose a fully automatic computerised method for skin lesion classification which employs optimised deep features from a number of well-established CNNs and from different abstraction levels. We use three pre-trained deep models, namely AlexNet, VGG16 and ResNet-18, as deep feature generators. The extracted features then are used to train support vector machine classifiers. In the final stage, the classifier outputs are fused to obtain a classification. Evaluated on the 150 validation images from the ISIC 2017 classification challenge, the proposed method is shown to achieve very good classification performance, yielding an area under receiver operating characteristic curve of 83.83% for melanoma classification and of 97.55% for seborrheic keratosis classification.
1. INTRODUCTION
Skin lesion diagnosis is important because late detection of melanoma is associated with much lower survival, while dermoscopic interpretation remains difficult and subjective. The paper addresses limitations of conventional computer-aided methods by combining multiple pretrained CNNs, multilevel features, and ensemble classification.
- Early skin cancer diagnosis matters because five-year survival exceeds 95% for early-stage disease but falls below 20% after late detection.
- Dermoscopic images provide better quality and morphological information than ordinary macroscopic images for differentiating pigmented skin lesions.
- Dermoscopic visual inspection is challenging because it depends on dermatologist experience and is affected by subjective interpretation, lesion variety, and image artefacts.
- Conventional computer-aided classification commonly combines lesion segmentation, handcrafted feature extraction, classification, and extensive image preprocessing.
- Prior deep-feature studies were limited by using particular architectures or layers, often relying on a single pretrained network.
- The proposed approach uses AlexNet, VGG16, and ResNet-18 as feature extractors, trains SVM classifiers, and fuses their outputs for three lesion classes.
2. MATERIALS AND METHODS
The method trains classifiers on ISIC dermoscopic images using minimally preprocessed inputs and features extracted from multiple pretrained CNNs. Separate multiclass SVM outputs are averaged and converted into probabilities for the two ISIC binary tasks.
- The training data comprise 2,037 dermoscopic images from ISIC 2016 and ISIC 2017, including 411 melanoma, 254 seborrheic keratosis, and 1,372 benign nevi images.
- Preprocessing is kept minimal: ImageNet mean normalization, network-specific resizing, and rotation-based training augmentation are applied.
- Deep Learning Models: AlexNet, VGG16, and ResNet-18 pretrained on ImageNet generate features from different abstraction levels, with shallower VGGNet and ResNet variants chosen to limit overfitting.
- Classification and Fusion: Multiclass nonlinear SVMs are trained separately for each network, and their class scores are averaged to produce the fused classification.
- Classification and Fusion: The fused multiclass scores are converted to probabilities and combined into melanoma-versus-all and seborrheic-keratosis-versus-all outputs.
3. RESULTS
Results are evaluated on the ISIC 2017 validation set using AUC, with the ROC curve reported for the best-performing fusion of all networks. The section also identifies incorrectly classified examples for both binary tasks.
- The ISIC 2017 validation set contains 150 images: 30 melanoma, 42 seborrheic keratosis, and 78 benign nevus cases.
- Performance is measured using area under the receiver operating characteristic curve on the provided validation images.
- The reported results include all individual networks and fused models, with the best-performing approach being the fusion of all networks.
- Figure 4 presents incorrectly classified examples for malignant melanoma-versus-all and seborrheic-keratosis-versus-all classification tasks.
4. DISCUSSION
The proposed hybrid method combines deep feature fusion, multiple SVM classifiers, and output fusion, outperforming single-network approaches. Results also indicate benefits from multi-level features, while dataset size, network coverage, and image resizing remain limitations.
- The hybrid approach fuses deep features, trains multiple SVM classifiers, and combines their probabilities for skin lesion classification.
- Feature fusion across different abstraction levels improves performance over using features from a single fully connected layer for every single network.
- AlexNet produces the best performance among the single-network approaches.
- The multiple-CNN method with fused SVM outputs outperforms the single-network approaches.
- The experiments are reported on the ISIC 2017 validation dataset.
- The study is limited by the number of pretrained networks, available training data, and potentially information loss from resizing images to small patches.
5. CONCLUSIONS
The paper presents a fully automatic skin lesion classification method using pretrained deep learning models and feature fusion. It reports very good classification results on challenging ISIC 2017 images, with further CNN fusion identified as a potential future improvement.
- The paper proposes a fully automatic method for skin lesion classification.
- Pretrained natural-image models and fusion of features from multiple layers or CNNs improve dermoscopic image classification performance.
- The method achieves very good classification results on challenging ISIC 2017 competition images.