PhD Student Z. Yuan and Prof. Yang Team 1st Place on CheXpert ML Competition

Thursday, September 24, 2020 - 9:21am

About the CheXpert Competition:

CheXpert is a medical imaging competition hosted by Stanford University ML Group, which aims to detect and classify multiple lung and chest diseases from X-ray images, such as Cardiomegaly, Pneumonia, etc. The submission is evaluated on average AUC (area under the ROC curve) between the ground truth and predicted probabilities of 5 selected diseases, namely, Cardiomegaly, Edema, Consolidation, atelectasis, Pleural Effusion, on a 500-patient studies annotated by 5 board-certified radiologists individually. This compeition has attracted over a hundred submissions from all over the world since its launch in January 2019.

About Our Solution:

Our team joined the competition in late May 2020 and on 08/31/2020 our result was ranked 1st of 156 in the leaderboard (our method is named DeepAUC-V1 ensemble). Our solution is based on directly optimizing a new surrogate loss function for AUC by our proposed stochastic optimization algorithms. The proposed model performance surpasses the previous competition leader's solution by optimizing regular objective loss on average AUC score and is also better than 2.8 out of 3 radiologists for 5 selected dieases on average. The overview of the proposed solution is shown below:

Overview of the proposed solution

About Our Team:

Main Contributor: Zhuoning Yuan

Advisor: Prof. Tianbao Yang

Others Members: Yan Yan, Zhishuai Guo, Mingrui Liu

Funding Source:

CAREER: Advancing Constrained and Non-Convex Learning, NSF Early Career Development Program (2019-2024).

[Submitted by Main Contributor: Zhuoning Yuan]


Relatedly, this November 19, 2020 Iowa State University Theoretical and Applied Data Science [TADS] Seminar with Tianbao Yang presenting his research group's work on "Deep AUC Maximization and Applications in Medical Image Classification"

Abstract: In this talk, I will present our recent research on a new learning paradigm of deep learning by AUC maximization. I will present a new surrogate loss for AUC and non-convex min-max optimization algorithms for solving deep AUC maximization problem. I will also talk about our results on Stanford CheXpert competition, on which our method is ranked at the 1st place as of today.