AutoMO-Mixer: Towards a Balanced, Safe and Robust Prediction Model in Medicine
Z Zhou1*, J Lv2, X Chen3, D Feng4, X Wang5, X Mou6, B Bai7, S Zhang8, (1) University Of Kansas Medical Center, Kansas City, KS (2) Xi'an Jiaotong University, Xi'an, China (3) Xi'an Jiaotong University, Xi'an, China (4) Xi'an Jiaotong University, Xi'an, China (5) Xi'an Jiaotong University, Xi'an, China (6)Xi'an Jiaotong University, Xi'an, China (7) Xi'an Jiaotong University, Xi'an, China (8) Xi'an Jiaotong University, Xi'an, China
Presentations
PO-GePV-M-304 (Sunday, 7/10/2022) [Eastern Time (GMT-4)]
ePoster Forums
Purpose: Medical image has become an important role in diagnosis and treatment outcome prediction. With the revival of artificial intelligence, especially the development of deep learning, it has achieved great success in medicine. To build a more reliable prediction model, an automated multi-objective Mixer (AutoMO-Mixer) model was developed to achieve balance, safe and robustness simultaneously.
Methods: In this study, totally 228 patients collected from Second Affiliated Hospital of Xi’an Jiaotong University (Xi’an, China) with Choroidal neovascularization (CNV) and cystoid macular edema (CME) between October 2017 and October 2019 were used. To reduce the number of parameters in training stage, multi-layer perceptron Mixer (MLP-Mixer) model was employed as base model. In training stage, both sensitivity and specificity are considered as the objective functions simultaneously to obtain more balanced model (between sensitivity and specificity). After training through iterative multi-objective immune algorithm, a Pareto-optimal model set is generated. Instead of manually selecting an optimal Pareto-optimal model, all the Pareto-optimal models are used by fusing them with the relative importance in testing stage. Furthermore, a new evidential reasoning-based entropy (ERE) was developed to estimate the uncertainty through test-time data augmentation strategy to obtain safer model. Meanwhile, ERE can improve the model robustness which is evaluated through fast gradient sign method.
Results: The mean and standard deviation value of AutoMO-Mixer on sensitivity, specificity, AUC and accuracy are 0.778±0.000, 0.779±0.000, 0.844±0.000, and 0.779±0.000, respectively, while ResNet is 0.728±0.075, 0.706±0.071, 0.791±0.046, 0.714±0.052 and MLP-Mixer is 0.611±0.052, 0.703±0.077, 0.709±0.041, 0.671±0.038. Meanwhile, the new developed entropy in AutoMO-Mixer can estimate the uncertainty very well (model performance will increase with the decreasing of uncertainty), and robustness outperformed the other two methods.
Conclusion: A new unified AutoMO-Mixer prediction model was developed in this study. The experimental results demonstrated that AutoMO-Mixer can obtain better performance and achieve balance, safe and robustness simultaneously.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email