---
title: "PANDA: Prototype-Anchored Alignment for Partially Unpaired Multimodal Learning"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2608.25970v1"
permalink: "https://feed.astralor.com/items/itm_74f2a5d3a5cd8fc6f40cd52c"
published: "2026-08-26T16:28:20.000Z"
score: 72
channel: "ai"
tags: ["PANDA","Medical Imaging","Prototype Alignment","Multimodal Learning"]
---

# PANDA: Prototype-Anchored Alignment for Partially Unpaired Multimodal Learning

## AI commentary

> PANDA elevates multimodal alignment from sample level to class-prototype level, allowing completely unpaired external modalities to contribute during training. This is valuable for medical scenarios with scarce data and constrained deployment. It is a training paradigm worth trying for researchers in multimodal learning and medical imaging AI.

## At a glance

PANDA proposes a two-stage framework to address incomplete pairing of auxiliary modalities in multimodal medical prediction. In the first stage, it learns a shared embedding from paired subjects and estimates class prototypes from auxiliary modalities; in the second stage, it trains the primary encoder on all subjects using cross-entropy plus alignment to frozen prototypes. The method accommodates arbitrary pairing rates and requires no auxiliary inputs at inference. On ADNI, it improves MRI-only baseline AUC by 7.9 percentage points; on TCGA-Lung survival prediction, it outperforms full-fusion training while requiring no RNA-seq at inference.

## Key points

- Proposes PANDA, a prototype-anchored alignment framework for partially unpaired multimodal learning with no auxiliary inputs at inference.
- On ADNI, PANDA improves MRI-only baseline AUC by 7.9 percentage points and reduces 1.5T CN false positives significantly.
- On TCGA-Lung survival prediction, PANDA outperforms full-fusion training while requiring no RNA-seq at inference.
- The method supports arbitrary pairing rates, including zero overlap, providing strong deployment flexibility.

## Full text

Multimodal medical prediction often faces incomplete pairing: auxiliary modalities with complementary signal are available for only a subset of subjects (or none) and cannot be assumed at deployment. We introduce PANDA (Prototype Anchored Data Alignment), a two-stage framework that transfers auxiliary information to a primary-modality model without auxiliary inputs at inference. Stage 1 learns a shared embedding from the paired subset and estimates class prototypes from auxiliary modalities; Stage 2 trains the primary encoder on all subjects using cross-entropy plus alignment to the frozen prototypes. Because supervision is defined at the class-prototype level, PANDA accommodates arbitrary pairing rates, including zero subject overlap. We evaluate PANDA on two applications. On a 1,021-subject multi-scanner ADNI cohort, we perform AD/CN classification with three auxiliary modalities at distinct pairing rates: tabular scores (44.8%), FDG-PET (18.7%), and external handwriting kinematics (0% overlap). Relative to the same-backbone MRI-only baseline, PANDA attains AUC 0.868 +-0.020 (+7.9pp) and reduces 1.5T CN false positives by 24.3pp; on a fully trainable Conv5-FC3 backbone it reaches AUC 0.893 (best overall). A pairing-rate ablation shows that the joint anchor remains within seed noise from 75% to 5% pairing. On TCGA-Lung survival prediction from whole-slide images with RNA-seq as auxiliary data, PANDA improves over WSI-only on 2-year OS (AUC +3.5pp) and Cox PH (C-index +9.0pts) and outperforms full-fusion training, which underperforms WSI-only, while requiring no RNA at inference; wide confidence intervals on this smaller cohort keep the gains below conventional significance. Overall, PANDA provides a deployment-oriented mechanism for leveraging incomplete auxiliary modalities to improve primary-modality prediction.

---

Verify against the original before citing; curated by Astral Feed: https://feed.astralor.com/items/itm_74f2a5d3a5cd8fc6f40cd52c