[Paper Review] Convolutional Neural Network Committees for Melanoma Classification with Classical And Expert Knowledge Based Image Transforms Data Augmentation

Cristina Nader Vasconcelos, Bárbara Nader Vasconcelos|arXiv (Cornell University)|Feb 22, 2017

Cutaneous Melanoma Detection and Management6 references20 citations

TL;DR

This paper proposes a committee of GoogLeNet-based convolutional neural networks (CNNs) for melanoma classification using a novel hybrid data augmentation strategy combining classical image transforms (geometric/color) and expert-knowledge-guided warping based on dermatological lesion symmetry. The method improves robustness to natural and biological variations, achieving a 0.932 validation score—only 2.6% below the top ISBI 2017 challenge result—demonstrating effectiveness on small, unbalanced biomedical datasets.

ABSTRACT

Skin cancer is a major public health problem, as is the most common type of cancer and represents more than half of cancer diagnoses worldwide. Early detection influences the outcome of the disease and motivates our work. We investigate the composition of CNN committees and data augmentation for the the ISBI 2017 Melanoma Classification Challenge (named Skin Lesion Analysis towards Melanoma Detection) facing the peculiarities of dealing with such a small, unbalanced, biological database. For that, we explore committees of Convolutional Neural Networks trained over the ISBI challenge training dataset artificially augmented by both classical image processing transforms and image warping guided by specialist knowledge about the lesion axis and improve the final classifier invariance to common melanoma variations.

Motivation & Objective

• Address the challenge of training deep learning models on small, unbalanced biomedical datasets like the ISBI 2017 melanoma dataset.
• Investigate whether data augmentation can improve CNN invariance to natural and biological variations in skin lesion images.
• Evaluate the impact of dataset imbalance on CNN performance and propose mitigation strategies.
• Demonstrate that combining classical and expert-knowledge-based data augmentation enhances model generalization and robustness.
• Assess the effectiveness of CNN committees in improving classification performance on a challenging melanoma classification task.

Proposed method

• Employ a pre-trained GoogLeNet architecture as the base CNN model, fine-tuned on the ISBI 2017 training set using transfer learning.
• Apply classical data augmentation techniques including random rotations, flips, and color jittering to increase diversity in training samples.
• Implement expert-knowledge-based image warping using thin-plate spline transformations guided by lesion symmetry and axis orientation, preserving key morphological patterns.
• Approximate lesion boundaries with ellipses and distort the main axis length (±20%) while maintaining direction and central position for realistic variation.
• Form CNN committees by training multiple models on balanced subsets of the augmented data to reduce variance and improve generalization.
• Use 5-fold cross-validation with fixed hyperparameters (10 epochs, 10⁻⁴ learning rate, batch size 256) to ensure reproducibility and fair comparison.

Experimental results

Research questions

• RQ1 Can classical image processing transforms (geometric and color) improve CNN invariance to natural variations in melanoma lesion images?
• RQ2 Can expert-knowledge-guided image warping—based on dermatological principles of lesion symmetry and axis orientation—enhance CNN robustness to biological variations?
• RQ3 How does class imbalance in the training dataset affect CNN performance, and can data augmentation mitigate this issue?
• RQ4 Does forming a committee of CNNs trained on augmented and balanced data improve classification performance compared to single models?
• RQ5 To what extent can a pre-trained GoogLeNet architecture achieve competitive results on a small, unbalanced biomedical dataset using hybrid data augmentation?

Key findings

• The proposed hybrid data augmentation strategy—combining classical and expert-knowledge-based warping—significantly improved model robustness to both natural and biological variations.
• The CNN committee trained on augmented and balanced data achieved a validation score of 0.932, which is only 2.6% below the best ISBI 2017 challenge result (0.958).
• Fine-tuning with ImageNet-pretrained weights reduced overfitting and training time, confirming the benefit of transfer learning on small datasets.
• The use of lesion-specific warping based on symmetry and axis orientation preserved critical diagnostic patterns, enhancing generalization without introducing spurious features.
• The method demonstrated strong performance despite the small training set (2000 images) and high class imbalance (68.6% seborrheic keratosis, 12.7% nevus, 18.7% melanoma).
• The results suggest that the approach is transferable to other CNN architectures, such as residual networks, and could be further improved with stronger models.