Experimental and Numerical Investigations on Robustness of a Deep Neural Network-based Multi-class Classification Model of CT Images with respect to Image Noise

Abstract

Robustness of Deep Neural Networks (DNNs) is an important aspect to consider for their clinical applications. This work examined robustness issue for a DNN-based multi-class classification model via comprehensive experimental and simulation studies. We constructed a DNN-based multi-class classification model that classifies an axial CT image as one of the four body sites of brain & neck (BN), chest (C), abdomen & pelvis (AP), and leg & foot (LF). The model was trained with whole-body CT images of 37 patients each scanned once and 10 scans of a whole-body phantom with different mAs levels to achieve a F1 score of 99.7% averaged over four classes on a testing dataset. We evaluated robustness of the model against noise perturbations under different mAs levels using simulated CT noisy images based on noise power spectrum (NPS) and experimental CT images acquired in repeated scans of the orange-man phantom. To quantify robustness of the trained model, we defined Successful Attack Rate (SAR) as the ratio of predictions changed with noises and Confusion Matrix for Robustness (CMR) that represents the percentage of a predicted class without noise being predicted to different classes under noise. Besides, we repeatedly trained the model for 100 times using exactly the same training procedure, hyper-parameter settings, as well as the training and validation dataset to investigate robustness of models due to inherent randomness in the training process. Finally, to improve model robustness, we employed an adaptive training scheme and demonstrated its effectiveness. Experimental and numerical studies demonstrated robustness issue of a DNN-based multi-class classification model. The discoveries highlight the needs for evaluating and ensuring DNN model robustness.

Ähnliche Arbeiten