Large language model for interpreting the Paris classification of colorectal polyps
Why this work is in the frame
A frame that forgets how it found something cannot be audited. These are the routes that admitted this work.
Bibliographic record
Abstract
Abstract Reporting of colorectal polyp morphology using the Paris classification is often inaccurate. Multimodal large language models (M-LLMs) may support morphological assessment. This study aimed to evaluate the accuracy of an M-LLM (GPT-4o) in classifying colorectal polyp morphology compared with expert and non-expert endoscopists. We used the SUN dataset of colonoscopy videos from 100 unique colorectal polyps, each labeled with the validated Paris classification. An M-LLM (GPT-4o) classified five representative frames per lesion. Three expert and three non-expert endoscopists, blinded to one another, performed the same task. The primary outcome was accuracy in differentiating non-polypoid (IIa/IIc) from polypoid (Is/Ip/Isp) lesions. The secondary outcome was accuracy in differentiating sessile (Is) from pedunculated (Ip/Isp) lesions. Given the exploratory design, no multiplicity correction was applied; point estimates are presented with 95% confidence intervals (CIs), and P values are interpreted descriptively. M-LLM accuracy for differentiating non-polypoid from polypoid lesions was 73% (95% CI 63%-81%), comparable to experts (75%, 65%-83%; P = 0.84) and non-experts (77%, 68%-85%; P = 0.52), with similar sensitivity and specificity. Accuracy for differentiating sessile from pedunculated lesions was 55% (95% CI 42%-67%), lower than experts (76%; P = 0.02) and non-experts (77%; P = 0.01), primarily due to poor specificity (12% vs. experts 82% and non-experts 88%; P < 0.01 for both comparisons). M-LLMs performed comparably to endoscopists in distinguishing non-polypoid from polypoid lesions but failed to reliably identify pedunculated morphology.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
Full frame distilled prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. Learned from the 10,348 direct Codex labels and 10,348 direct Gemma labels. Candidate is the union of thresholded teacher heads; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels or direct frontier model labels.
Codex and Gemma teacher scores by category
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.000 | 0.000 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.000 |
| Bibliometrics | 0.000 | 0.000 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.000 | 0.000 |
| Open science | 0.000 | 0.000 |
| Research integrity | 0.000 | 0.000 |
| Insufficient payload (model declined to judge) | 0.000 | 0.000 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from it