Synthetic radiomic features improved throat-lesion classification—but only 30 patients tested the claim
A generative model helped a classifier distinguish benign from malignant laryngeal lesions in one retrospective dataset. The held-out test was small and no external hospital or clinical workflow was evaluated.
The short version
- What happened
- The retrospective dataset contained 11,144 narrow-band images from 210 adults; 180 patients were used for development and 30 for an internal held-out test.
- Who it affects
- Earlier, more consistent recognition of suspicious throat lesions could shorten diagnosis, but the published experiment cannot show that this model changes biopsy decisions or outcomes.
- What remains unproven
- Only 30 patients in the held-out test
Keep in mind
Only 30 patients in the held-out test
Single retrospective data source
No external-centre validation
No clinician, workflow or patient-outcome comparison
Source analysis · 1 linked source. Primary analysed source date: 24 October 2026.
Read the full report · 2 min The evidence, methods, implications and limitations
The experiment augmented measurements rather than inventing clinical pictures
The researchers analysed 11,144 narrow-band endoscopy images from 210 adults with benign or malignant laryngeal lesions. They extracted numerical radiomic measurements, then trained generative adversarial networks to create extra feature vectors for under-represented classes. The generated material was therefore tabular data derived from images, not synthetic endoscopy photographs.
Development used 180 patients and an internal held-out test used 30. Across repeated experiments, the best generative pipeline recorded a median F1 score of 0.851, versus 0.806 for the conventional SMOTE oversampling baseline. F1 combines precision and recall; it does not by itself show how often a clinician would miss a cancer or recommend an unnecessary biopsy in practice.[1]
Thirty held-out patients cannot establish clinical readiness
The patient-level split is important because images from the same person should not leak across training and testing. Even so, a 30-patient test from the same retrospective source is too small to establish transportability. The paper does not report a prospective clinician comparison, workflow effects or patient outcomes.
A stronger assessment would freeze the full pipeline and test consecutive patients at independent centres, reporting sensitivity and specificity with confidence intervals by lesion type and image quality. It should also compare specialists with and without support and track biopsies, delays and missed disease. The authors reported no dedicated external funding and no competing interests.[1]
The distinction that matters
What was generated and what was tested
The study augmented extracted radiomic measurements before classifying lesions.
Source data
11,144 images from 210 adults
Extract radiomic features
Generative step
Create numerical feature vectors
Balance development data
Held-out test
30 patients
Median F1 0.851