EndoStyle is a new artificial intelligence system designed to improve how AI detects colorectal polyps during colonoscopy. Colonoscopy is the gold standard for colorectal cancer screening, and AI tools that automatically flag polyps can help doctors catch more precancerous growths. However, a critical challenge has limited these AI tools: they often perform poorly when used on camera equipment different from what they were trained on.
The core problem is called domain shift, which means that colonoscopy images look visually different depending on which video processor brand and model is used. Differences in color depth, contrast, saturation, brightness, and image resolution all make images from one device look noticeably different from another, even when capturing the same tissue.
EndoStyle solves this by converting colonoscopy images recorded with one device to look as if they were recorded with a different device, while keeping the actual medical content (such as polyps) intact. This gives AI training datasets far more variety without requiring doctors to collect and annotate new images from every possible camera system.
The researchers tested EndoStyle extensively: they measured how realistic the generated images looked, asked experienced endoscopists to judge them, and then trained polyp-detection AI models using EndoStyle-augmented data to see whether real-world detection accuracy improved.
Each year in the United States alone, more than 17.7 million gastrointestinal endoscopic procedures are performed. Computer-aided polyp detection (CADe) systems have become an important part of clinical colonoscopy, helping endoscopists spot polyps they might otherwise miss. Studies have shown that AI assistance can raise adenoma detection rates, which directly translates to better colorectal cancer prevention.
However, most AI systems are trained using images from many different endoscopic camera systems, while in practice a given hospital or clinic typically uses only one brand. This mismatch means a model trained on mixed data may not perform optimally on the specific camera a clinic owns. Conversely, a model trained on data from just one camera may fail on another clinic's equipment entirely.
Researchers have identified that hardware differences produce variation in image shape, resolution, color depth, contrast, white balance, and lens distortion. Even commercially available AI systems are often restricted to specific hardware configurations as a result. This leaves many clinical sites under-served by AI tools optimized for equipment they do not use.
Traditional solutions include domain-invariant feature learning and standard image augmentation (such as adding noise or flipping images), but these cannot replicate the characteristic visual style of a completely different camera brand. Fine-tuning models with locally acquired annotated data is effective but requires significant effort to collect new labeled images. EndoStyle offers a practical middle path.
EndoStyle is built on two separate AI components working in sequence. The first is a StarGANv2 generative model, which is a type of generative adversarial network capable of learning multiple visual styles simultaneously. It takes a source colonoscopy image and a reference image representing the target camera style, then generates a new image that preserves the anatomical content while adopting the visual appearance of the target device.
Because the StarGANv2 output is only 256x256 pixels (too low for clinical use), the second component is an Enhanced Deep Residual Network (EDRN) that doubles the resolution to 512x512 pixels through super-resolution. All final synthetic images used in evaluations were at this higher resolution.
Training used nearly 240,000 frames from 3,452 colonoscopies recorded on five different processor brands: Olympus CV-170, CV-190, CV-1500, Karl Storz Image1 S, and Pentax EPK-i7000. A 90-10 patient-level split prevented data leakage. Two specialized training losses were added to the standard GAN training: a depth-guided loss to preserve luminal structure, and a polyp-segmentation loss to prevent the model from accidentally erasing subtle flat polyps during style conversion.
These additional losses were important because early experiments without them showed that EndoStyle sometimes flattened delicate tissue features or lost hard-to-see polyps during conversion. The depth and segmentation constraints ensured that generated images remained medically faithful even as their visual style changed.
The quality of EndoStyle's generated images was evaluated with two standard image-quality metrics. The mean Frechet Inception Distance (FID) score was 27.2, and the mean Learned Perceptual Image Patch Similarity (LPIPS) score was 0.3 across all five processor types. Lower values on both metrics indicate that generated images are visually close to real images, and these scores indicate high fidelity.
Semantic similarity was measured using three independent AI foundation models: MedSigLIP, BiomedCLIP, and DINOv3. Across all three, the similarity between a source image and its EndoStyle-converted version was statistically comparable to the similarity between the converted image and the target style reference. This confirms that EndoStyle successfully captures the target style without distorting the underlying anatomy.
A multicenter realism study asked 22 endoscopists from 16 clinical centers to judge whether images belonged to a given colonoscopy video. EndoStyle-generated images were accepted as realistic 86.5% of the time, compared to 88.8% for real images from the same recording. The difference was not statistically significant, meaning experienced specialists could not reliably distinguish real from synthetic images.
Importantly, 19 of the 22 participating endoscopists had each performed more than 1,000 colonoscopies, making this a particularly demanding test. The fact that highly experienced specialists rated synthetic images as realistic at nearly the same rate as genuine frames provides strong validation of the system's image quality.
The main clinical test was whether EndoStyle-generated data could improve polyp-detection AI performance. Two test datasets were used, each consisting of full-length colonoscopy videos recorded exclusively with a single Olympus processor (CV-190 or CV-1500). Models were trained on publicly available multi-processor datasets and then augmented with EndoStyle-generated images matched to the target processor's style.
On Test Dataset 1 (CV-190), adding EndoStyle data reduced false-positive detections by 43%, while specificity improved from 96.9% to 98.2% and precision improved from 55.2% to 63.7%. Sensitivity decreased slightly (from 48.8% to 40.8%), primarily because the additional training images gave the model a better understanding of what non-polyp tissue looks like, making it less prone to alerting on harmless mucosa.
On Test Dataset 2 (CV-1500, a brand-new processor whose images were entirely absent from the training data), false positives were cut by 41%, and specificity improved from 95.7% to 97.4%. Again, sensitivity had a modest decrease while precision improved meaningfully. Critically, only one polyp was missed by the EndoStyle-augmented model across this entire dataset.
False positives are a major barrier to clinical adoption of AI polyp detection. Frequent false alarms disrupt the endoscopist's workflow, erode trust in the AI system, and can lead clinicians to ignore true alerts. A 40%-plus reduction in false alarms is therefore a clinically meaningful improvement that could make real-world use of these AI tools more practical.
A particularly rigorous experiment tested whether EndoStyle could help even when the training images came from a camera brand (Fujifilm) that EndoStyle had never seen during its own training. This represents the real-world scenario where a clinic's equipment is entirely different from anything in the AI developer's dataset.
Models trained on only Fujifilm images were augmented with EndoStyle-converted images targeting either the CV-190 or CV-1500 style. On Test Dataset 1, false positives fell by 24.2%, and on Test Dataset 2, by 32.7%, even though the style transfer had never been trained to handle Fujifilm images. Neither the base model nor the augmented model missed any polyp in Test Dataset 2.
These results suggest that EndoStyle has some generalization beyond its training distribution. Although the gains were somewhat smaller than in the multi-processor scenario, they remained meaningful and suggest that the system could benefit clinics with newer or less common equipment that is underrepresented in public datasets.
Compared against commercially available polyp detection systems on Test Dataset 1, the EndoStyle-augmented models showed higher specificity than most commercial alternatives despite being trained on considerably less data, underscoring how much the style-transfer augmentation contributed to the models' ability to reduce nuisance alerts.
The authors acknowledge several important limitations. EndoStyle was trained and validated on only five endoscopic processor models, and its ability to generalize to newer devices needs further study. The realism evaluation used still images rather than video sequences, and applying EndoStyle frame-by-frame to video may introduce temporal inconsistencies if consecutive frames are converted slightly differently.
The polyp detection evaluation used two relatively small test datasets, both from the same manufacturer (Olympus), and larger multi-manufacturer validation studies are needed. The ratio of synthetic to real images used in training (20% positive, 100% negative) was fixed based on exploratory experiments, but future work could explore adaptive sampling strategies that dynamically balance sensitivity and specificity.
An interesting technical consideration is that EndoStyle currently treats vascular patterns as part of the image style rather than preserving them as anatomical content. In some clinical contexts, vascular patterns are diagnostically important, so decoupling them from style transfer could be a valuable refinement.
Looking ahead, the authors suggest exploring diffusion-based generative models as an alternative to GANs for style transfer, comparing style transfer with fine-tuning when limited target-domain data is available, and developing automated content-consistency verification. Beyond polyp detection, the technology may also find use in educational simulators for training endoscopists on a wider variety of equipment styles.
Colorectal cancer is among the most preventable cancers when caught early through screening colonoscopy. Polyps detected and removed before they turn malignant dramatically reduce a person's lifetime risk of developing colorectal cancer. AI-assisted polyp detection has shown real promise for improving the rate at which adenomas are caught, but its real-world adoption has been slowed by the false-alarm problem and by hardware fragmentation across clinical sites.
EndoStyle addresses both issues simultaneously. By adapting AI training data to match a clinic's specific hardware, it enables more accurate and reliable performance on locally used equipment. By specifically targeting the reduction of false positives through style-matched non-polyp training images, it makes the AI less distracting and more trustworthy for the endoscopist.
For patients, this could translate to more consistent and equitable access to AI-assisted screening regardless of which hospital or endoscopy center they visit. Hospitals that use less common equipment brands would no longer be disadvantaged by AI systems optimized only for the most widely used processors.
The practical framework of EndoStyle, converting existing labeled datasets rather than collecting new ones, significantly reduces the cost and effort required to deploy AI across diverse clinical environments. This makes it more feasible for smaller clinics and healthcare systems in lower-resource settings to benefit from AI-enhanced colonoscopy in their colorectal cancer screening programs.