Generalization and Visual Comprehension of CNN Models on Chromosome Images

Computer-aided image classification has achieved start-of-the-art performance since Convolutional Neural Network structures were employed. Classical neural networks such as AlexNet and VGG-Net inspired several rules of designing network models. Besides benchmark datasets such as MNIST, CIFAR and ImageNet, classification performance of medical images such as chromosome karyotyping images also improved via Convolutional Neural Network. However, there are few studies on generalization among different datasets. In this paper, we designed a neural network with nine layers, and achieved classification accuracy of 0.984, 0.816 and 0.921 on the dataset of MNIST, CIFAR and chromosome karyotype images. We also visualized the output of several layers of the model and explained that smooth output between neural network layers may induce lower accuracy on classification.