[CVPR 2026] SUREON: Moving Surgical AI from Perception to Reasoning with GRPO and Expert Narrations

Executive Summary

TL;DR: Deep learning in surgery has historically been "eyes without a brain"—capable of identifying a scalpel but unaware of the risks involved in its current trajectory. SUREON bridges this gap by introducing a massive dataset of 206k QA pairs extracted from surgical lectures. By combining supervised fine-tuning with Group Relative Policy Optimization (GRPO), the authors developed SureonVLM-R1, a model that not only performs surgical tasks but "thinks" through clinical rationales using explicit reasoning tokens.

Background: This work is a pivotal shift from traditional Perception-only AI (SOTA on Cholec80/HeiChole) to Cognitive Surgical AI. It treats surgical lectures not just as video data, but as a source of "hidden" reasoning supervision.

The Problem: The "Silent" Surgeon

Current surgical AI systems are trained on "what" is visible: tools, organs, and phases. However, surgery is a game of intent and anticipation. Prior works fail because:

• Rigid Ontologies: Models are trapped in fixed label spaces (e.g., "Instrument: Grasper").
• Annotation Bottleneck: It is nearly impossible to get surgeons to sit and label "Why am I doing this?" for millions of frames.
• Lack of Interpretability: General VLMs (GPT-4/Gemini) suffer from domain shift and lack the "surgical common sense" required for safety-critical decisions.

Methodology: Extracting Wisdom from Lectures

The core innovation lies in the SUREON Data Curation Pipeline. Instead of manual labeling, the authors used a multi-agent GPT-5 framework to harvest Semantic Grounding Moments (SGMs) from expert-narrated videos.

1. The Multi-Agent Pipeline

The pipeline uses specialized Generators to identify moments in transcripts where the narrator explains intent or safety, and Validators to ensure these temporal windows align with the visual scene. Fig 1: The end-to-end pipeline transforming narrated videos into 12 structured question types, from safety assessment to forecasting.

2. SureonVLM-R1: Thinking before Acting

The model architecture (based on Qwen3-VL) undergoes a two-stage evolution:

• Stage 1 (SFT): Progressive adaptation of the MLP and Vision Encoder.
• Stage 2 (GRPO): Unlike standard RL, GRPO optimizes the model's reasoning trajectory without a massive reward model, encouraging the model to generate <think> tokens. This allows the model to "verbalize" its logic—for example, identifying "tissue blanching" as a cue for energy application—before providing a final answer.

Experiments: Surpassing the Giants

SureonVLM-R1 was tested against state-of-the-art generalist models and specialized surgical baselines.

Performance Highlights

• Safety First: In "Safety Action Identification," the model scored 93%, dwarfing GPT-5.1's 62%.
• Beyond Perception: Even on traditional tasks like Phase Recognition, the reasoning-trained model (SureonVLM) outperformed general VLMs significantly (F1 0.63 vs 0.36 for GPT-5.1 on Cholec80).

Table 1: Performance across the 12 taxonomy categories. Note the massive lead in Decision Reasoning and Action Description.

Visual Reasoning Case Study

The model demonstrates true Inductive Bias for surgery. In qualitative tests, it correctly deduced that a vessel branch was sacrificed to remove an enlarged lymph node—a high-level clinical inference that goes far beyond simple tool tracking.

Critical Analysis & Conclusion

Takeaway

The success of SUREON proves that the "intelligence" required for surgical AI is already encoded in the way surgeons teach. By algorithmically structuring this pedagogical data, we can move past the limitations of manual labeling.

Limitations & Future Work

• Pedagogical Bias: Lectures focus on "interesting" moments; routine, boring steps of surgery remain underrepresented.
• Hallucination Risk: While <think> tokens provide transparency, the reasoning traces themselves are not yet expert-validated and could potentially hallucinate clinical logic.
• Future Path: The next frontier will likely be Intra-operative Decision Support, where these reasoning models are integrated into real-time robotic systems to provide "co-pilot" suggestions to surgeons.

Conclusion: SUREON sets a new standard for surgical VLM benchmarks, proving that specialized reasoning models are not just a luxury, but a necessity for the next generation of safe, interpretable surgical assistants.