Y Peng*, C Shen, Y Gonzalez, S Zhang, X Jia, The University of Texas Southwestern Medical Ctr, Dallas, TX

• 
  Y Peng

SU-H430-IePD-F6-6 (Sunday, 7/10/2022) 4:30 PM - 5:00 PM [Eastern Time (GMT-4)]

Exhibit Hall | Forum 6

Purpose: Deep neural network (DNN) is powerful in numerous classification tasks. However, robustness of DNN models has been pointed out, such that prediction results may not be stable against noise perturbations to the input data. Previous studies investigated this issue via simulation studies. In this work, we examine the robustness of a DNN-based classification model and for the first time, experimentally demonstrate this issue.

Methods: We trained a DNN-based multi-class classification model that classifies a CT image as one of the five body sites (brain, chest, abdomen, leg, and foot). This model is relevant, as it often serves as the starting point of organ segmentation task. We trained the model with whole body CT images of patients and an orange-man phantom to yield a satisfactory classification performance. After that, robustness of the model was evaluated against noise perturbations under different mAs levels. We first generated simulated noise based on noise power spectrum (NPS) acquired in real scans. We then performed repeatedly scans of the orange-man phantom to examine robustness experimentally.

Results: In simulation study, change percentage (CP), namely percentage of predictions changed by noise to the input, generally increased, as mAs was reduced, indicating reduced model robustness with increased noise amplitude. At 25 mAs, CP for chest and leg sites were 21.60% and 18.86%, respectively, and those of the other sites were above 50%. CP of chest and leg sites dropped to ~0 at ~37.5 mAs. Repeated CT scans of orange-man chest at 25 mAs found CP = 22.67%, in agreement with simulation results.

Conclusion: To our knowledge, this is the first experimental study clearly demonstrating robustness issue of a DNN-based multi-class classification model. The discoveries call for future studies to address this issue.

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email