Quality Control
Using some of our work on anomaly detection as a starting point, we have been working on the development of methods that can
survey the quality of the segmentations produced by cardiac segmentation models. By formulating our framework as an anomaly detection problem, we avoid
annotations on the segmentation quality, which is a limitation of current methods.
Fig. 1. A Convolutional Autoencoder (CA) is trained with ground truth (GT) masks from a cardiac imaging dataset. At inference, the CA reconstructs an input mask $\widehat{X}$, previously segmented by a model. The reconstructed mask $\widehat{X}’$ acts as a pseudo ground truth (pGT) to estimate a function $\rho$, a surrogate measure of the segmentation quality and the model’s performance.
By using the proposed framework, we have been able to reproduce the final rankings of the ACDC Challenge using six different cardiac segmentation models.
All the code and experiments have been made available though our Github repository.
Students
Francesco Galati (funded by the Monaco Government)
References
- F. Galati and MA. Zuluaga. Efficient Model Monitoring for Quality Controlin Cardiac Image Segmentation. In: Functional Modeling and Imaging of the Heart (FIMH) 2020. in press
- R. Candela, P. Michiardi, M. Filippone, MA. Zuluaga. Model Monitoring and Dynamic Model Selection in Travel Time-series Forecasting. In: European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases ECML-PKDD (2020)
- Quality control in Time series data. Github repository