Source-linked AI summary
Data Augmentation for Skin Lesion Analysis
Fábio Perez, Cristina Vasconcelos, Sandra Avila, Eduardo Valle
TL;DR
Limited annotated skin-lesion data motivates this study of 13 augmentation scenarios for melanoma classification across three CNNs, dataset sizes, and test-time prediction. The experiments find benefits from augmentation during training and testing, with scenario J reaching an AUC of 0.882 without external data, while lesion-mix augmentation remains limited by handcrafted image processing.
Problem
Annotated skin-lesion images are difficult to collect, while deep-learning models require substantial data for automated analysis.
Method
The study compares 13 augmentation scenarios across Inception-v4, ResNet, and DenseNet, including training, test-time, and varying-dataset-size evaluations.
Results
Data augmentation benefits melanoma classification during training and testing, with scenario J achieving an AUC of 0.882 without additional data.
Takeaways & Limitations
In some cases, data augmentation can provide more performance gains than obtaining new images.
Takeaways & Limitations
Lesion-mix augmentation performs worse than other scenarios, and its handcrafted image processing may be unsuitable for producing reliable images.
Abstract
from arXiv · showhide
Deep learning models show remarkable results in automated skin lesion analysis. However, these models demand considerable amounts of data, while the availability of annotated skin lesion images is often limited. Data augmentation can expand the training dataset by transforming input images. In this work, we investigate the impact of 13 data augmentation scenarios for melanoma classification trained on three CNNs (Inception-v4, ResNet, and DenseNet). Scenarios include traditional color and geometric transforms, and more unusual augmentations such as elastic transforms, random erasing and a novel augmentation that mixes different lesions. We also explore the use of data augmentation at test-time and the impact of data augmentation on various dataset sizes. Our results confirm the importance of data augmentation in both training and testing and show that it can lead to more performance gains than obtaining new images. The best scenario results in an AUC of 0.882 for melanoma classification without using external data, outperforming the top-ranked submission (0.874) for the ISIC Challenge 2017, which was trained with additional data.
1 Introduction
The paper addresses limited annotated skin-lesion data by investigating data augmentation for melanoma classification across CNN architectures, dataset sizes, and test-time prediction. Experiments use the ISIC Challenge 2017 dataset.
- Annotated skin-lesion images are difficult to collect, while deep-learning models require substantial data.
- Data augmentation modifies training images while preserving semantic information and labels, helping reduce overfitting on small datasets.
- The study evaluates diverse augmentation techniques across Inception-v4, ResNet, and DenseNet CNNs.
- The experiments examine augmentation across dataset sizes and during test-time to assess effects on generalization error.
- Experiments use the ISIC Challenge 2017 dataset for melanoma classification.
2 Related Work
Prior work establishes data augmentation as common in CNN-based skin-lesion classification, while this paper broadens the comparison to 13 scenarios, test-time methods, and varying dataset sizes.
- CNNs commonly use color modifications and geometric transforms such as rotation, scaling, and random cropping.
- Test-time augmentation averages predictions from multiple transformed views, but extensive cropping can be expensive in production.
- Data augmentation is widespread among top-ranked ISIC Challenge 2017 submissions.
- Earlier skin-lesion studies report gains from geometric, color, symmetry-preserving, and test-time augmentations.
- This work compares testing techniques, 13 augmentation scenarios including a novel method, and effects across dataset sizes.
3 Methodology
The methodology evaluates 13 augmentation scenarios across three deep CNN architectures, using pretrained networks and multiple training, validation, and test-time prediction procedures.
- CNN Architectures: Experiments use Inception-v4, ResNet-152, and DenseNet161, selected as deep architectures with different design choices.
- CNN Architectures: The authors compare broad trends rather than numeric values alone because identical optimization hyperparameters were used across networks.
- Data Augmentation Techniques: The study evaluates 13 augmentation scenarios comprising individual image-processing techniques and combinations.
- Data Augmentation Techniques: Scenarios J to M are compositions whose augmentations are applied in the presented order.
- Training and Evaluation: Validation predictions average 16 augmented copies, while test evaluation uses the original image, 64 augmented copies, or 144 crops.
- Training and Evaluation: Each setup runs six separate trainings, and documented source code is provided for reproducibility.
4 Results and Discussion
Augmentation generally improves melanoma-classification performance during training and testing, with scenario J producing the strongest AUCs and benefits often increasing when training data are limited.
- Augmentation on Training and Testing: Geometric transforms such as affine changes, random crops, and elastic transformations provide more consistent improvements across the three networks.
- Augmentation on Training and Testing: Random erasing provides little benefit for Inception and DenseNet and worsens results relative to no augmentation with ResNet.
- Augmentation on Training and Testing: Combining elastic transformations with common augmentations performs worse than scenario J, possibly because of the resulting deformations.
- Augmentation on Training and Testing: Lesion-mix scenarios perform worse than other augmentations, suggesting the generated images do not preserve relevant features from both source lesions.
- Augmentation on Training and Testing: Scenario J achieves AUCs of 0.854 for Inception-v4, 0.882 for ResNet, and 0.879 for DenseNet.
- Augmentation on Training and Testing: Averaging augmented samples or 144 crops outperforms prediction on the original image alone for every scenario.
- Augmentation on Training and Testing: Test-time augmentation can provide faster inference than 144 crops when ResNet and DenseNet achieve similar results with 64 versus 144 samples.
- Impact of Data Augmentation on Different Dataset Sizes: Training and testing with scenario J noticeably improves performance for datasets containing 500 or more images.
5 Conclusion
Data augmentation improves melanoma classification during both training and testing, with the strongest scenario exceeding the ISIC Challenge 2017 top-ranked AUC without additional data. However, handcrafted lesion-mix augmentation performed worse than other scenarios and may produce unreliable images.
- Data augmentation positively affects melanoma classification during both training and testing.
- The best scenario, combining geometric and color transformations, surpasses the ISIC Challenge 2017 top-ranked AUC without additional data.
- Lesion mix augmentation scenarios I and M perform worse than other scenarios.The authors implemented lesion mixing with handcrafted image-processing techniques.
- Handcrafted lesion mixing may be unsuitable for producing reliable images, while generative architectures could improve it.