Source-linked AI summary
Automated Skin Lesion Classification Using Ensemble of Deep Neural Networks in ISIC 2018: Skin Lesion Analysis Towards Melanoma Detection Challenge
Md Ashraful Alam Milton
TL;DR
The paper addresses automated melanoma and skin-lesion detection from dermoscopic images, motivated by the importance and limitations of conventional diagnosis. It evaluates augmented and pretrained deep neural-network models, including an ensemble, on the ISIC 2018 dataset. The proposed approach achieved competitive classification performance, with a best validation score of 0.76 for PNASNet-5-Large, while dataset imbalance and size remain limitations.
Problem
Melanoma is a threatening skin cancer requiring early diagnosis, while conventional diagnosis is time-consuming, expensive, and potentially biased.
Method
The study preprocesses and augments dermoscopic images and evaluates pretrained PNASNet-5-Large, InceptionResNetV2, SENet154, InceptionV4, and ensemble models.
Results
The ensemble-based deep-neural-network approach achieved competitive dermoscopic-image classification performance, with the best validation score of 0.76 for PNASNet-5-Large.
Takeaways & Limitations
Automated dermoscopic-image classification could be used along with experienced dermatologists as a highly automated and accurate system.
Takeaways & Limitations
The unbalanced dataset makes lesion-feature generalization harder, while larger and more varied data and further hyper-parameter optimization may improve performance and robustness.
Abstract
from arXiv · showhide
In this paper, we studied extensively on different deep learning based methods to detect melanoma and skin lesion cancers. Melanoma, a form of malignant skin cancer is very threatening to health. Proper diagnosis of melanoma at an earlier stage is crucial for the success rate of complete cure. Dermoscopic images with Benign and malignant forms of skin cancer can be analyzed by computer vision system to streamline the process of skin cancer detection. In this study, we experimented with various neural networks which employ recent deep learning based models like PNASNet-5-Large, InceptionResNetV2, SENet154, InceptionV4. Dermoscopic images are properly processed and augmented before feeding them into the network. We tested our methods on International Skin Imaging Collaboration (ISIC) 2018 challenge dataset. Our system has achieved best validation score of 0.76 for PNASNet-5-Large model. Further improvement and optimization of the proposed methods with a bigger training dataset and carefully chosen hyper-parameter could improve the performances. The code available for download at https://github.com/miltonbd/ISIC_2018_classification
1. Introduction
Melanoma is a deadly form of skin cancer whose early diagnosis is important, but conventional diagnosis is time-consuming, expensive, and potentially biased. Automated classification of dermoscopic images is presented as a possible alternative.
- Melanoma comprises about 1% of skin cancers but causes a large majority of skin cancer deaths.
- About 91,270 new melanomas and 9,320 melanoma deaths were estimated in the United States for 2018.
- Traditional diagnosis uses physical examination and biopsy, with results taking about a week and possible cost, time, and dermatologist-bias drawbacks.
- Automated skin lesion classification from dermoscopic images could help address melanoma diagnosis challenges.
2. Background And Related work
Prior work applies digital image processing, deep neural networks, transfer learning, and pretrained convolutional features to skin-lesion classification. Clinical background methods include a three-point checklist and ABCD parameters for identifying melanoma features.
- Deep neural networks have achieved dermatologists-level accuracy in classifying skin lesions from dermoscopic images.
- The three-point checklist considers asymmetry, atypical pigment networks, and blue-white structures, with melanoma considered when at least two are present.
- The ABCD method evaluates asymmetrical shape, irregular borders, and multiple colors in melanoma lesions.
- Prior studies used pretrained ConvNets as feature extractors, transfer learning with fine-tuning, and ImageNet-derived descriptors for skin-image classification.
3. ISIC 2018 Melanoma Detection Challenege and Dataset
The ISIC 2018 challenge includes three skin-lesion analysis tasks, and this paper uses its disease-classification dataset containing seven disease categories.
- ISIC 2018 defines lesion segmentation, lesion attribute detection, and disease classification as its three tasks.
- The training set contains 10,015 images spanning seven skin diseases, while the validation set contains 193 images.
- Task 3 seeks improved automated disease-classification predictions from dermoscopic images.
4. Proposed Methodology
The proposed system preprocesses and augments dermoscopic images, then uses pretrained deep networks with frozen or fine-tuned layers for seven-class lesion classification. Models include several architectures and their ensemble.
- Pre-Processing: Images are normalized using ImageNet mean RGB values, converted to a 0–1 pixel range, and resized for network input.
- Data Augmentation: Augmentation uses rotations, flips, crops, brightness and contrast changes, pixel jitter, aspect-ratio changes, shearing, zoom, and shifts.
- Data Augmentation: The training data are augmented to reduce class imbalance, and HAM 1000 images are also used for training.
- Models: The model set includes PNASNet-5-Large, InceptionResNetV2, SENet154, InceptionV4, and an ensemble of all models.
- Classification: The final output uses softmax probabilities for seven disease classes in Task 3.
- Fine Tuning: Pretrained ImageNet models are fine-tuned by initially freezing weights, then unfreezing later layers and updating them by back-propagation.
5. Training Details
The models were initialized from ImageNet-pretrained weights and trained with staged freezing and learning-rate adjustment. Adam optimization and cross-entropy loss were used for backpropagation.
- ImageNet-pretrained models began training with all weights frozen except the last layer.The initial learning rate was 0.0001 for the first 2 epochs.
- After 2 epochs, the training schedule changed the learning rate while continuing model optimization.
- Adam optimization updated weights using cross-entropy loss and backpropagation.
6. Implementation Details
The project was implemented in PyTorch because the authors found it suitable for flexible model modification, using multi-GPU hardware for training.
- The project was implemented in PyTorch for its simplicity and flexibility in modifying the dynamic computation graph.
- Training used multiple Nvidia 1080ti GPUs.
8. Conclusion
The study found that an ensemble of deep neural networks can achieve competitive dermoscopic skin-lesion classification performance. The authors identify dataset imbalance and limited training diversity as constraints, while reporting a best validation score of 0.76 for PNASNet-5-Large.
- An ensemble of deep neural-network methods achieved competitive performance for dermoscopic skin-lesion classification.
- The proposed approach could support a highly automated and accurate classification system used alongside experienced dermatologists.
- 0.76 was the best validation score achieved by the PNASNet-5-Large model.
- Class imbalance made it harder to generalize visual lesion features.
- Larger, more balanced datasets and additional regularization or hyper-parameter tuning may improve performance and robustness.