Weakly-Supervised Midline Shift Quantification through Simulating the Reversed Disease Progression

Chih-Chieh Chen, Chang-Fu Kuo

Keywords: Medical Image Segmentation, Conditional Flow Matching, Weakly Supervised Midline Shift Measurement

TL;DR: Generating visual and quantitative information from a black-box classifier

Abstract:

Medical segmentation masks are often scarce. To get visual and quantitative information,
we propose constructing trajectories from anomaly data to normal data using conditional
flow matching on an autoencoder, augmented with an auxiliary classification head in the
latent space. We demonstrate the effectiveness of our method through weakly supervised
midline shift estimation.

Submission Number: 2