Low-cost self-constructing multi-objective multi-mode parallel vestibular schwannoma recognition method

Abstract

Convolutional neural networks (CNNs) have achieved great success in the fields of radiology and pathology. The automatic recognition of vestibular schwannoma (VS) based on magnetic resonance imaging (MRI) can significantly enhance the speed and accuracy of disease diagnosis, and reduce the threat of the disease to patients’ lives. At present, patients are diagnosed using contrast enhanced T1-weighted mode images from MRI but there is growing interest in high resolution T2-weighted mode images. However, due to the complex relationship between these two modes, applying a CNN using a simple multi-mode fusion strategy makes it difficult to learn complex information between the modes, and the feature information cannot be well matched and fused. In addition, most CNN hyper-parameters require fine tuning by experts in numerous “trial and error” experiments to achieve better results, and it is difficult to balance multiple objectives such as the model accuracy and training time. The cost of optimization is very expensive. Therefore, we propose a high-performance “non-deep” VS recognition model with dual-mode multi-channel feature perception coupled with a surrogate-assisted multi-objective particle swarm optimization algorithm based on a Kullback–Leibler (KL)-Dropout network to balance multiple objectives while reducing model optimization costs and human influence. Our experimental results showed that the proposed algorithm reached the optimal level in the benchmark test problem. By combining the proposed algorithm with the proposed model, the accuracy was better in the comparison and the amount calculated by the model was controllable, which verified the effectiveness and generalizability of the proposed method.