---
title: "Zhejiang University's MedVeriSeg Teaches Medical AI to Refuse Hallucinated Tumor Masks"
subtitle: "A training-free framework catches hallucinated masks in medical segmentation models by reading the geometry of similarity heatmaps between [SEG] tokens and image features."
url: "https://alphasignal.ai/news/zhejiang-university-s-medveriseg-teaches-medical-ai-to-refuse-hallucinated"
author: "AlphaSignal Newsroom"
author_url: "https://alphasignal.ai/editorial-team"
company: "AlphaSignal"
company_url: "https://alphasignal.ai/company/alphasignal"
published: "2026-04-14T00:00:00.000Z"
updated: "2026-08-24T00:35:26.563Z"
type: "Paper"
topics:
  - "Image"
  - "Security"
subtopics:
  - "Vision Language"
  - "Red Teaming"
upvotes: 935
source_url: "https://arxiv.org/abs/2604.10242"
comments: 0
truncated: true
---

# Zhejiang University's MedVeriSeg Teaches Medical AI to Refuse Hallucinated Tumor Masks

> A training-free framework catches hallucinated masks in medical segmentation models by reading the geometry of similarity heatmaps between [SEG] tokens and image features.

Medical segmentation models built on multimodal LLMs have a quiet reliability problem: ask them to segment a tumor that isn't there, and they will happily draw one anyway. A group from Zhejiang University and Southern Medical University has proposed [MedVeriSeg](https://arxiv.org/abs/2604.10242), a training-free verification layer that lets these models say no when the queried target doesn't exist in the image, without retraining a single weight.

## The hallucination hiding inside LISA

Most recent MLLM-based medical segmenters follow the LISA blueprint. LISA employs LLaVA to output a text response based on the input image, text query, and a \[SEG\] token, and the embedding for that customized token is decoded into a segmentation mask via a SAM decoder. In medical variants like [MediSee](https://arxiv.org/html/2604.10242) and Med-PLIB, the vision backbone is swapped for MedSAM and the language side for LLaVA-Med, but the mechanism is the same.

These models were trained almost exclusively on valid query-mask pairs, so existing LISA-like methods typically generate masks regardless of whether the target specified in the query is present, leading to hallucinated segmentation. Ask a chest CT model to outline a clavicle in an abdominal slice and you still get a confident mask. In a teaching setting that misleads students; in a clinical setting it risks misdiagnosis.

![Comparison of segmentation with and without MedVeriSeg on a brain MRI](https://arxiv.org/html/2604.10242v4/teaser.png) 

Retraining with negative samples is expensive and tends to overfit. Feeding the query into a general MLLM like Qwen-3-VL as a gatekeeper also underperforms, because general models struggle with fine-grained medical visual cues.

## The geometry of a real answer

_Preview only. Full text requires an AlphaSignal Pro account._

## Sources

- [MedVeriSeg: Teaching LISA-Like Medical Segmentation Models to Verify Query Validity Without Extra Training](https://arxiv.org/html/2604.10242)
- [GSVA: Generalized Segmentation via Multimodal Large Language Models](https://arxiv.org/html/2312.10103v3)
- [Image Segmentation in Foundation Model Era: A Survey](https://arxiv.org/pdf/2408.12957)
- [MedVeriSeg GitHub repository](https://github.com/Edisonhimself/MedVeriSeg)