AutoMO-Mixer: An automated multi-objective multi-layer perspecton Mixer model for medical image based diagnosis

Xi Chen, Jiahuan Lv, Xuanqing Mou, Zhiguo Zhou

Abstract: Medical image based diagnosis is one of the most challenging things which is vital to human life. Accurately identifying the patient's status through medical images plays an important role in treatment of diseases. Deep learning has achieved great success in medical image analysis. Particularly, Convolutional neural network CNN) can obtain promising performance by learning the features in a supervised way. However, since there are too many parameters to train, CNN always requires a large scale dataset to feed, while it is very difficult to collect the required amount of patient images for a particular clinical problem. Recently, MLP-Mixer (Mixer) which is developed based multiple layer perceptron (MLP) was proposed, in which the number of training parameters is greatly decreased by removing convolutions in the architecture, while it can achieve the similar performance with CNN. Furthermore, obtaining the balanced outcome between sensitivity and specificity is of great importance in patient's status identification. As such, a new automated multi-objective Mixer (AutoMO-Mixer) model was developed in this study. In AutoMO-Mixer, sensitivity and specificity were considered as the objective functions simultaneously to train the model and a Pareto-optimal Mixer model set can be obtained in the training stage. Additionally, since there are several hyperparameters to train, the Bayesian optimization was introduced. To obtain a more reliable results in testing stage, the final output was obtained by fusing the output probabilities of Pareto optimal models through the evidence reasoning (ER) approach. The experimental study demonstrated that AutoMO-Mixer can obtain better performance compared with Mixer and CNN.