We propose a modified convolutional neural network (CNN) tailor-made for computed tomography (CT) image disease recognition to assist doctors in disease diagnosis. First, we analyze the effects of varying the CNN activation function and pooling parameters, as well as the CNN's performance using one data set. Second, we address the activation error that occurs when the sample data size is increased by preprocessing images by an enhancement technique, adjusting the activation function and initialization weighting, training/testing the target, and adaptively extracting features. We found that our method alleviates overfitting with these techniques. The experimental results show that our proposed scheme improves the recognition rate and can better generalize findings.
