Abstract
Deep learning has advanced automatic skin lesion recognition, yet challenges remain regarding limited training data, model complexity, and high computational costs. This paper proposes a solution using Convolutional Spiking Neural Networks (SNNs) with an unsupervised Spike-Timing-Dependent Plasticity (STDP) learning rule. By integrating efficient temporal coding, event-driven learning, and a winner-take-all (WTA) mechanism, the network achieves sparse spike coding and efficient learning. The proposed SNN, enhanced with a statistical feature selection layer, achieves an average accuracy of 87.7% on melanoma classification. Experimental results demonstrate that this lightweight 3-layer architecture outperforms much deeper traditional CNNs in both accuracy and runtime efficiency, making it highly suitable for deployment on portable diagnostic devices.
Introduction: The Efficiency Challenge in Medical Imaging
Automated melanoma recognition is critical for improving survival rates through early detection. However, traditional Convolutional Neural Networks (CNNs) often struggle with medical imaging due to data scarcity, high computational demands, and the "domain gap" when using models pretrained on natural images.
This research explores Spiking Neural Networks (SNNs) as a biologically inspired alternative. Unlike standard CNNs that process continuous values, SNNs communicate via discrete "spikes" over time, mimicking the energy efficiency of the human brain.
Methodology
The core of the proposed system is a 3-layer Convolutional SNN designed for unsupervised feature extraction.
Spike Encoding and Learning
The process begins with Intensity-to-Latency Encoding. Raw images are processed through a Difference of Gaussians (DoG) filter, converting pixel intensities into temporal spike trains where high-contrast features fire earlier.
Learning is achieved through Unsupervised STDP, where synaptic weights are updated based on the precise timing of spikes. A Winner-Take-All (WTA) mechanism ensures that neurons compete to represent distinct features, preventing redundancy.
Statistical Feature Selection
A key innovation is the replacement of standard global max pooling with a Chi-Square Feature Selection layer. While traditional pooling focuses on the most prominent features, this statistical approach identifies the most diagnostic features, allowing the network to ignore visual noise and focus on subtle markers of malignancy.
Datasets and Preprocessing
The study utilized the ISIC 2018 Challenge dataset, focusing on Malignant Melanoma (MEL) and Benign Melanocytic Nevi (NV).
To prepare the data, the authors employed the "DullRazor" algorithm for hair removal, median filtering for noise reduction, and global contrast normalization. A balanced training set of 1,081 images per class was used, with a large test set of 5,557 images to validate robustness against class imbalance.
Key Results and Insights
The SNN approach demonstrates that architectural simplicity can yield superior performance:
- Accuracy: The SNN with feature selection reached 87.7%, compared to 83.8% without the selection layer.
- Efficiency: The model uses only 44,830 parameters, a fraction of the 62 million parameters in ResNet152, yet it achieved higher classification accuracy.
- Stability: The unsupervised SNN proved more stable and easier to deploy than pretrained CNNs, which often require extensive hyperparameter tuning for medical tasks.
Conclusion
This work demonstrates that STDP-based SNNs are highly effective for medical image classification. By focusing on diagnostic features rather than just prominent ones and leveraging the inherent efficiency of spike-based computation, the researchers have created a model that is both accurate and lightweight. Future work will focus on expanding the model to multi-class classification and exploring more sophisticated spike encoding methods to capture finer textures.
Summary updated