Classifying Skin Lesions Using Convolutional Spiking Neural Networks With Unsupervised Learning Rule

Skin cancer is one of the most common tumors in the world,and malignant melanoma is a kind of malignant tumor with very high mortality rate.The early detection of melanoma malignant lesions can greatly improve the survival chance of patients.Deep learning methods have made some achievements in skin cancer image classification tasks,but there are still some problems such as limited training samples,too complicated networks structure,and expensive computational cost.Considering the inherent high energy saving,biological credibility and good image recognition performance of the spiking neural networks,in this paper unsupervised spike-timing-dependent plasticity(STDP)convolutional spiking neural networks are used to classify the images of malignant melanoma and benign melanocytic nevi.The efficient spiking coding,event-driven learning rules and winner-take-all(WTA)mechanism in the convolutional neural networks jointly ensure the network’s sparse spiking coding and efficient learning efficiency.Convolutional spiking neural networks are used to extract the characteristic information in the skin cancer images,and then SVM and random forest classifiers are used to make the classification,the average classification accuracy rate of the two classifiers reach 0.838 and 0.845,respectively,and their average area under the receiver operating characteristic curve(AUC)values reach 0.907 and 0.893,respectively.We further propose to use filtered feature selection to replace the last pooling layer of convolutional spiking neural networks to select more diagnostic features and remove redundant features in order to improve the classification performance of the networks model.The comparison experiments of the effectiveness of feature selection are carried out on the standard dataset with high similarity and the skin cancer image dataset.After using the feature selection,the AUC value of the standard dataset classification is increased from 0.739 to 0.789,and the AUC value of the skin cancer image classification is increased from0.917 to 0.945.Further,we compare the performance of our networks with existing deep learning-based skin cancer image classification algorithms.The experimental results show that compared with the convolutional neural networks(CNNs)need to be trained from scratch,our networks not only have better classification accuracy,but also have significantly better calculation efficiency.In addition,compared with the pre-trained CNNs models,the training time of our networks is similar to that of the pre-trained CNNs,but the classification accuracy is better,and the classification performance is more stable and easier to use.In addition,compared with the CNNs models commonly used in skin cancer image classification tasks,our networks have only three convolutional layers,which greatly reduces the complexity of the model and the parameters that need to be trained in the network.Our results shows that our convolutional spiking neural networks based on unsupervised STDP learning rules have good and stable classification performance,and they don’t consume that much computing resources,which is very beneficial to implement automatic skin cancer image classifiers on small portable devices(such as smartphones).