An adaptive neural architecture optimization model for retinal disorder diagnosis on 3D medical images

Abstract Neural architecture design is one of the critical tasks for deep neural models because of the high variety of structure options. This research proposes an adaptive neural architecture optimization (ANAO) model to optimize the convolutional neural network (CNN) structure based on neural blocks, which are collected from the existing state-of-the-art CNNs. An integer programming model is proposed for the optimization process, where the objective is to maximize the designed model accuracy and convergence speed. Constraints are constructed to restrict the CNN design requirements. To enhance the training efficiency of the designed CNN model, a novel objective function is proposed, which considers the accuracy and the training convergence trend. A recurrent neural network is applied to evaluate the performance of the candidate models to boost the efficiency of the optimization process. A heuristic process is proposed to conduct the optimization. The proposed ANAO model is applied for the retinal disorder diagnosis. Eight state-of-the-art CNNs are tested for comparison with the proposed ANAO model from both accuracy and convergence trend perspectives. Experimental results show that the proposed ANAO model can optimize the CNN architecture to adaptively fit a given dataset and achieve quite high-level performance.