Research map / Vision-language and multimodal / Multimodal LLMs

CG-Reasoner: Centroid-Guided Positional Reasoning Segmentation for Medical Imaging with a Robust Visual-Text Consistency Metric

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Multimodal LLM benchmarks

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes CG-Reasoner, a LLaVA-Med-based framework with a light-weight ConvNeXt-Tiny encoder-decoder and a Text2Centroid module that jointly produces medical segmentation masks and positional reasoning text, plus a new PRScore metric for reasoning-mask alignment.

Problem

Medical image segmentation models optimize pixel accuracy but neglect positional reasoning, and existing reasoning evaluations (ChatGPT yes/no judgments, BLEU/ROUGE) fail to check whether generated text correctly localizes the lesion.

Approach

Couple a LoRA-tuned multimodal LLM with a FiLM-conditioned lightweight decoder, and use a frozen Text2Centroid regressor that maps reasoning text to lesion centroid coordinates to supply a spatial grounding loss; evaluate with PRScore, which averages SBERT semantic similarity, text-derived spatial similarity, and text-to-mask centroid similarity.

Key results

• CG-Reasoner reaches mDice 0.977 / mIoU 0.958 on Lung X-ray, the best in its comparison, and 0.970 / 0.948 on Lung CT-Scan.
• PRScore on Breast Ultrasound is 0.755 for CG-Reasoner versus 0.729 for PRS-Med; PRS-Med is slightly higher on Polyp Endoscopy (0.734 vs 0.732) and Skin RGB (0.735 vs 0.730).
• Adding Text2Centroid raises PRScore across all six datasets (e.g., Breast Ultrasound 0.721 to 0.755).

Surprising findings

• CG-Reasoner has lower mDice than PRS-Med on Breast Ultrasound (0.765 vs 0.817) and lower than BiomedParse on Polyp Endoscopy in Table 2 comparisons, and it is also below PRS-Med on Polyp Endoscopy and Skin RGB PRScore.

Limitations

• PRScore relies on a Text2Centroid model whose own accuracy is not validated against human judgments.
• Gains over PRS-Med on segmentation are mixed; CG-Reasoner trails on Breast Ultrasound and Polyp Endoscopy.
• Ablations of the decoder and PRScore weighting choices are limited in the main text.

Experimental setup

Datasets: MMRS (PRS-Med combined dataset) Breast Ultrasound Brain MRI Lung CT-Scan Lung X-ray Polyp Endoscopy Skin RGB

Benchmarks: Breast Ultrasound Brain MRI Lung CT-Scan Lung X-ray Polyp Endoscopy Skin RGB Image

Metrics: mDice mIoU PRScore

Baselines: G-Dino + SAM-Med2D BiomedParse LISA-7B LISA-13B PRS-Med PixelLM Tiny-SAM Med-SAM Swin-UNet

Builds on: LISA LLaVA-Med ConvNeXt PRS-Med FiLM

Closest papers by meaning

• MedReasoner: Reinforcement Learning Drives Reasoning Grounding from Clinical Thought to Pixel-Level Precision (AAAI 2026 · similarity 0.89)
• CoPRS: Learning Positional Prior from Chain-of-Thought for Reasoning Segmentation (ICLR 2026 · similarity 0.86)
• CXReasonBench: A Benchmark for Evaluating Structured Diagnostic Reasoning in Chest X-rays (NeurIPS 2025 · similarity 0.86)
• IBISAgent: Reinforcing Pixel-Level Visual Reasoning in MLLMs for Universal Biomedical Object Referring and Segmentation (CVPR 2026 · similarity 0.85)
• "Why Is There a Tumor?": Tell Me the Reason, Show Me the Evidence (ICML 2025 · similarity 0.85)
• M3CoTBench: Benchmark Chain-of-Thought of MLLMs in Medical Image Understanding (ICLR 2026 · similarity 0.85)