Identification of subthreshold depression based on deep learning and multimodal medical image fusion / 中国医学影像技术

ABSTRACT

Objective: To explore the value of convolutional neural network (CNN) algorithm based on deep learning (DL) for identification of subliminal depression (StD) patients using medical image data. Methods: MRI and fMRI data of 56 StD patients (StD group) and 70 normal controls(NC group) were collected and input into the constructed CNN, respectively. Then the network fusion technology was used to comprehensively analyze the two different modalities to obtain the classification result. Finally, the network fusion technology was used to integrate two different modal data and optimize the classification effect. Results: The identification accuracy of the structural image data alone was 73.02%, of the functional image data alone was 65.08%. With combination of the two modes, the final classification accuracy raised to 78.57%. Conclusion: DL can classify patients with StD and normal subjects. Multiple modal input methods can improve classification accuracy.