Sinkhorn-based Quantization for Reversed Disease Progress: Further Investigations

Chih-Chieh Chen, Chang-Fu Kuo

Keywords: Interpretable AI, Medical Image, Clinical Diagnosis, VQGAN, OptVQ, flow matching

Abstract:

Interpretability plays a pivotal role in the collaboration between artificial intelligence (AI)
systems and clinicians. It enables clinicians to critically reassess the rationale underlying
AI-generated predictions. Moreover, translating these interpretations into clinically mean-
ingful quantifications is feasible even for more granular algorithms, thereby potentially
reducing the extensive annotation efforts typically required. Recently, a novel approach
was introduced to generate reversed disease progression trajectories by applying condi-
tional flow matching within the latent space of an autoencoder, jointly training a linear
classifier. However, the architectural design, training procedures, and objective functions
associated with the flow matching network warrant further investigation and refinement. In
the present study, we implement this concept utilizing a recently proposed vector-quantized
autoencoder framework incorporating Sinkhorn-based quantization. Our findings indicate
that reversed disease progression can be consistently generated even in the absence of joint
classifier training. Additionally, the method preserves strong spatial correspondences be-
tween the pixel domain and latent representations, enabling the synthesis of desired images
through a CutMix-inspired algorithm. We demonstrate the efficacy of our approach by
applying it to the weakly supervised quantization of midline shift distances.

Primary Subject Area: Interpretability and Explainable AI

Secondary Subject Area: Unsupervised Learning and Representation Learning

Registration Requirement: Yes

Reproducibility: https://github.com/chihchiehchen/Sinkhorn-based-Quantization-for-Reversed-Disease-Progress/tree/main

Visa & Travel: No

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 49