Otoendoscopic images stop relying on text alone: CLNA turns unwritten visual variation into an adjustable latent, cutting FID from 52.91 to 10.11
Synopsis
Otoendoscopic image synthesis has been limited because text prompts state a diagnosis but not visual details such as hue and texture, leaving appearance variation within a diagnostic class hard to control. CLNA adds a style encoder and dual cross-attention to a Stable Diffusion 2.1 backbone, compressing such non-prompt attributes into a continuous latent whose sampling range is set by a style parameter. Across normal tympanic membrane, otitis media with effusion and chronic otitis media, FID against the held-out internal test set fell from 52.91 for baseline Stable Diffusion to 10.11, and feature-space recall rose from 57.67% to 90.22%.
Interpretation
Otoendoscopic image synthesis can now adjust visual variation beyond the text prompt while the prompt still fixes the diagnostic class. Earlier controllable text-to-image methods rely on spatial maps, reference images or composable conditioning signals and need extra inputs at inference, which is impractical when only diagnosis-level prompts are available. On the internal test set, CLNA reached an FID of 10.11 against 52.91 for baseline Stable Diffusion, and k-nearest-neighbor recall of 90.22% against 57.67%.
The mechanism adds a dedicated style channel inside the diffusion backbone. The style encoder is a convolutional downsampling network that reads the mean and variance of a Gaussian posterior from two rows of the final feature map of the clean image latent, then samples a style latent by reparameterization. That latent enters the denoising U-Net as a single token through dual cross-attention in three down blocks, the middle block and three up blocks; the loss combines a denoising term with a KL regularization term weighted at 1.
Generated images can replace part of the real data used to train a downstream classifier. A classifier trained on real data alone reached 94.46% F1 on the internal test set and 83.76% in a single run on the external test set. Replacing 10%–40% of real training images with CLNA-generated samples gave 93.57%–94.08% F1 on the internal test set and 84.81%–85.30% on the external test set; with fully synthetic replacement, CLNA reached 78.53% internal and 72.37% external against 70.25% and 54.26% for the baseline.
Perspective
The demonstrated scope is three otoendoscopic diagnostic classes—normal tympanic membrane, otitis media with effusion and chronic otitis media—with internal data drawn from 2018–2020 video frames at a single center and external data from the public AI-HUB otoscopic image set; FID and recall are both computed in a domain-adapted InceptionV3 feature space.
The style parameter adjusts only the magnitude of non-prompt visual variation and cannot steer its direction toward user-specified visual or clinical attributes; the FID and recall gains hold in the domain-adapted feature space and do not establish an extractor-independent quality advantage; the reader study did not include baseline-generated images; each diagnostic class used one fixed prompt, so robustness to varied prompt wording was not evaluated; both datasets were collected in a single country, and anonymized data precluded demographic analysis.