Optimization of Reconstruction Accuracy of Anomaly Position Based on Stacked Auto-Encoder Neural Networks

The diffuse optical tomography (DOT) technique which uses the traditional linear iterative algorithm has the problems of slow calculation speed and low reconstruction imaging accuracy in the inverse problem reconstruction, which limits its clinical application and development. This paper proposes an inverse-problem solving technology based on a stacked auto-encoder (SAE) network to improve the reconstruction accuracy of anomaly position, size and absorption coefficient in tissues. The reconstruction accuracy of anomaly position, size and absorption coefficient obtained by the traditional algebraic reconstruction technique (ART) method and the SAE based method are experimentally compared. The experimental results show that the SAE based method achieves the prediction accuracy of anomaly position of 96.25%, thus improving the accuracy and shortening the reconstruction time compared with the traditional ART method. Accordingly, the proposed method provides a better solution to the problem of the inaccurate reconstruction of the position and size of the rapid DOT based positioning of anomalies.