Skin lesion detection based on an ensemble of deep convolutional neural network (1705.03360v1)

Abstract: 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.

• The paper shows that ensemble DCNNs using weighted majority voting significantly improve skin lesion classification accuracy.
• It integrates models like GoogLeNet, AlexNet, ResNet-50, and VGGNet, utilizing extensive data augmentation to boost performance.
• The study highlights the potential of deep learning ensembles for enhancing automated skin cancer screening and clinical diagnostics.

Ensemble of Deep Convolutional Neural Networks for Skin Lesion Detection

This paper, authored by Balazs Harangi, presents an ensemble approach using Deep Convolutional Neural Networks (DCNNs) for classifying dermoscopy images into three categories: melanomas, nevi, and seborrheic keratoses. Recognizing the high public health impact of skin cancer, specifically melanoma, the paper responds to the imperative need for more accurate automated diagnostic tools. This work was carried out in the context of the 2017 ISBI Challenge on Skin Lesion Analysis towards Melanoma Detection, illustrating the ongoing efforts in the medical imaging domain to leverage deep learning methodologies for clinical advancements.

Methodology Overview

The proposed solution involves integrating the outputs from multiple DCNN architectures, specifically GoogLeNet, AlexNet, ResNet-50, and VGGNet, through a fusion approach predicated on weighted majority voting. The ensemble benefits from pre-trained models on ImageNet, fine-tuned to accommodate the dermoscopic dataset provided by the challenge. Data augmentation was critically applied to address the limitations of the training data size, enhancing the dataset from 2,000 to 14,300 images via transformations such as rotations and random cropping.

Ensemble models in machine learning are known to enhance prediction accuracy by integrating multiple classifiers, each contributing its strengths to the decision-making process. In this case, each neural network's outputs were weighted based on their receiver operating characteristic curves (AUCs), with weights determined during the fine-tuning phase. The experimental setup involved extensive computational resources, showcasing the feasibility of utilizing GPUs to expedite training.

Experimental Results

The ensemble approach demonstrated superior performance when compared to individual DCNNs, achieving an overall score of 0.932 in the AUC metric on the validation set. This demonstrates the efficacy of the fusion-based methodology, evident in increased average precision and specificity across the diagnostic tasks. Table I of the paper offers a robust statistical presentation of these results, delineating the collective accuracies and other pertinent metrics such as sensitivity at varying specificity levels.

Given these results, the paper posits that ensemble DCNNs could set a new standard for skin lesion classification, delivering higher precision necessary for the intricacies of dermatological diagnostics. The balanced performance across diverse lesion types highlights the adaptability of deep learning frameworks when enhanced with techniques such as weighted voting.

Implications and Future Directions

The practical implications of this work are significant for automated skin cancer screening, promising to extend the reach and reliability of diagnostic measures, especially in resource-limited settings where clinical expertise may be scarce. Theoretically, the success of this ensemble approach underscores the potential synergies obtainable through the fusion of heterogeneous model architectures.

Looking forward, further exploration into optimizing ensemble methods for medical image analysis could yield even greater accuracies and efficiencies. Future research could tackle challenges such as real-time diagnosis on mobile platforms, leveraging the computational capabilities discussed. Furthermore, expanding datasets to include more diverse skin types and conditions could refine the robustness of such models, promoting broader generalization and applicability.

In conclusion, this paper exemplifies the progressive impact of advanced AI methodologies in clinical diagnostics, reinforcing the importance of multi-model strategies in enhancing the predictability and generalizability of machine learning applications in healthcare.