← Vaishak Deep dive 01 / 07 Get in touch

Active · MICCAI 2026 target

FOVEA-Net

Hallucination-resistant vision-language model for chest CT.

Role
Sole researcher
Stack
PyTorch · MedNeXt · LangGraph
Code
Repository →

<15%

target hallucination rate, from a 35% baseline

>0.70

target mean IoU, finding to region

26

papers surveyed to position the design

How it works

  1. 01 MedNeXt-L encoder 3D volume → dense features
  2. 02 3D RPN foveation proposes candidate regions
  3. 03 Region transformer reads each region locally
  4. 04 Confidence-gated decoder mask-conditioned text
  5. 05 Report aggregation findings + region ids

CT volume in → grounded report out. Each stage can only pass forward what the previous one localised.

Hallucination rate

Published baseline ~35%
FOVEA-Net target <15%

The single number the architecture is built to move.

The problem

Report-generation models for chest CT hallucinate roughly a third of their findings. The failure is not random noise — the decoder is free to produce fluent clinical language whether or not the volume contains the evidence. Nothing in the architecture forces the sentence and the scan to agree, and a radiologist reading the output has no way to audit which slice a claim came from.

The approach

Make grounding a structural constraint rather than a loss term. Five stages, each narrowing what the decoder is allowed to say. A finding cannot be emitted unless a proposed 3D region carries it, and the region id travels with the sentence into the final report.

What broke

The first foveation pass proposed regions the decoder promptly ignored — grounding existed on paper but not in the attention. Cross-attention masks conditioned on the proposed masks closed the leak, at the cost of a decoder that now refuses to describe anything the RPN missed. That trade is the point: a missed finding is a recall problem I can measure, a fabricated one is a trust problem I cannot.

Where it stands

Architecture and training pipeline are complete; evaluation is running against the survey baselines. Every sentence the model emits currently carries a region id and an IoU, which is the property the whole design exists to guarantee.