📄 medicine

SY-CNN+: Synthetic-Data-Enhanced Hybrid Learning for Clinical Photograph-Based Eyelid Tumor Classification

This study introduces SY-CNN+, a hybrid learning framework that leverages GAN-based synthetic data augmentation to achieve high accuracy in classifying malignant versus benign eyelid tumors from clinical photographs, outperforming human clinicians in an exploratory comparison while demonstrating robustness across retrospective and prospective cohorts.

Original authors: Guanchao Tong, Yan Yan, Jiashu Li, Yixuan Sun, Jia Rong, Zixuan Chen, Jiamin Wang, Gaurav Gupta, Jiayou Chao, Kuan Huang, Sung Soo King, Qi Gao, Yongqing Shao, Guohai Wu

Published 2026-07-07

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Guanchao Tong, Yan Yan, Jiashu Li, Yixuan Sun, Jia Rong, Zixuan Chen, Jiamin Wang, Gaurav Gupta, Jiayou Chao, Kuan Huang, Sung Soo King, Qi Gao, Yongqing Shao, Guohai Wu

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine your eyelid is a small, busy neighborhood. Sometimes, harmless little houses (benign tumors) pop up there, but occasionally, a dangerous, invasive construction crew (malignant tumors) moves in. The problem is that from a distance, the dangerous crew often looks exactly like the harmless ones. Even experienced neighborhood watch captains (doctors) can get confused, especially if they haven't seen many cases before.

This paper introduces a new tool called SY-CNN+ designed to help tell the difference between these "good" and "bad" eyelid growths just by looking at photos.

Here is how the researchers built this tool, explained in simple terms:

1. The Problem: Not Enough Photos

To teach a computer to spot a dangerous tumor, you usually need thousands of photos. But these specific eyelid tumors are rare, and collecting photos is hard because of patient privacy. It's like trying to teach a child to recognize a specific type of rare bird when you only have a handful of pictures. The computer gets confused and makes mistakes.

2. The Solution: "Fake" Photos to Fill the Gaps

To solve the lack of photos, the researchers used a trick. They used two types of "artificial artists" (AI models called GANs and Diffusion models) to create brand new, realistic-looking photos of eyelid tumors that don't actually exist in real life.

The Analogy: Imagine you are training a security guard to spot a specific type of thief. You only have 10 photos of the thief. So, you hire an artist to draw 100 more pictures of what that thief might look like wearing different hats or standing in different lights. You then show all 110 pictures to the guard.
The Test: The researchers tested both types of artists. They found that the GAN artist (who creates images by playing a game of "real vs. fake") did a better job at helping the computer learn than the Diffusion artist (who creates images by slowly turning static noise into a picture), even though the Diffusion pictures looked slightly more "perfect" on paper.

3. The Brain: A Two-Step Team

The system doesn't just use one brain; it uses a team with two distinct roles:

The Detective (ResNet50): This is a deep-learning computer that looks at the photo and pulls out the important clues (like the shape of the edges or the texture). It's like a detective who spots the tiny details a human eye might miss.
The Judge (XGBoost): Once the Detective gathers the clues, they hand them to the Judge. The Judge is a classic machine-learning algorithm that makes the final decision: "Guilty" (Malignant) or "Not Guilty" (Benign).
Why a team? The researchers found that this "Detective + Judge" combo worked better than just letting the Detective make the final call, especially when the amount of real data was small.

4. The Results: How Well Did It Work?

The researchers tested their system in two ways:

The Practice Run (Retrospective Test): They tested it on old photos they hadn't shown the computer before. The best version (using the GAN artist and the Detective/Judge team) got it right 92% of the time.
The Real-World Test (Prospective Test): They then tested it on brand new photos taken recently. It got it right 89.8% of the time.

They also compared the computer to two human doctors (one senior, one junior).

The Senior Doctor: Got it right about 81% of the time.
The Junior Doctor: Got it right about 73% of the time.
The Computer: Got it right 92% of the time.

5. Important Caveats (What the Paper Says, and What It Doesn't)

The authors are very careful not to overhype the results. Here is the reality check they provide:

It's not a replacement for a doctor: The computer still missed some dangerous tumors (about 18% of the time in the practice run). The authors explicitly state this tool should not be used alone to diagnose a patient.
It's a "Second Opinion": The best use for this right now is as a helper. It can flag suspicious photos so a specialist can take a closer look.
Limited Scope: The study was done at one hospital with a specific set of cameras. The computer was trained on low-resolution images (small, pixelated squares), which means it might miss tiny details like fine blood vessels.
Future Work: Before this can be used in real hospitals, it needs to be tested on much larger groups of people from many different hospitals to make sure it works everywhere.

In summary: The paper presents a new AI system that uses "fake" photos to train itself to spot dangerous eyelid tumors. It performed better than the human doctors in their specific test, but the authors warn that it is still a work-in-progress tool meant to assist doctors, not replace them.

Technical Summary: SY-CNN+ for Eyelid Tumor Classification

Problem Statement
Differentiating between malignant and benign eyelid tumors from clinical photographs is a critical yet challenging task in ophthalmology. Malignant lesions, such as sebaceous gland carcinoma (SGC) and melanoma, can mimic benign conditions like chalazion or cysts, leading to delayed diagnosis, increased metastasis risk, and unnecessary surgical interventions. This diagnostic difficulty is exacerbated in primary care or low-resource settings where specialist experience is limited. Furthermore, the development of robust AI models for this task is hindered by the scarcity of high-quality, diverse medical image datasets. Rare tumors are difficult to collect, image curation is labor-intensive, and patient privacy restrictions limit data sharing. Traditional data augmentation techniques (e.g., rotation, cropping) often fail to generate clinically relevant pathological variations, such as ulcerated margins or heterogeneous pigmentation, which are essential for distinguishing malignancy.

Methodology: SY-CNN+ Framework
The authors propose SY-CNN+ (Synthetic-Data-Enhanced Convolutional Neural Network+), a hybrid learning framework designed to address data scarcity through synthetic augmentation and a two-stage classification architecture.

Data Sources and Partitioning:
- The study utilized 3,041 clinical images from 881 patients at Ningbo Eye Hospital.
- The dataset was split into a retrospective cohort (2,992 images from 863 patients) and a prospective test cohort (49 images from 18 patients).
- The retrospective cohort was partitioned into training, validation, and test1 sets using a patient-level independent split (approx. 8:1:1 ratio) to prevent data leakage. Stratified allocation ensured consistent malignant-to-benign ratios across sets.
Synthetic Data Generation:
- To augment the training set, the authors employed Generative Adversarial Networks (GANs) and Diffusion models.
- Synthetic images were generated only from the training set using class-conditional training to preserve label integrity. No validation or test images were used for generation.
- The augmentation strategy was controlled by a synthetic-to-real ratio ( $n$ ), tested at values of 0, 0.5, 1, 2, 3, 4, and 5.
Hybrid Architecture (CNN+):
- Feature Extraction: The framework utilizes CNN backbones (ResNet18, ResNet50, VGG11, VGG16) to extract image features. Input images were resized to 64x64x3.
- Downstream Classification: Instead of relying solely on the CNN's final softmax layer, the extracted feature vectors were fed into separate machine learning classifiers: Naive Bayes (NB), Support Vector Machine (SVM), Random Forest (RF), and Extreme Gradient Boosting (XGBoost).
- Selection Process: The validation set was used to determine the optimal model configuration (backbone + classifier) and the optimal synthetic-to-real ratio ( $n$ ).
Evaluation Protocol:
- The selected model was evaluated on the retrospective test1 set and the prospective test2 cohort.
- Performance was measured using accuracy, F1-score, confusion matrices, ROC curves, and Matthews Correlation Coefficient (MCC).
- An exploratory comparison was conducted against two volunteer clinicians (one senior, one junior) on the test1 set.

Key Results

Optimal Configuration: The best-performing model was GAN-based augmentation combined with ResNet50 feature extraction and XGBoost classification, with a synthetic-to-real ratio of $n=4$ .
Performance Metrics:
- Retrospective Test1: Achieved 92.0% accuracy (95% CI: 88.5–95.5%) and an F1-score of 87.2%.
- Prospective Test2: The same configuration achieved 89.8% accuracy and an F1-score of 82.8%.
Comparison with Diffusion Models: While diffusion-generated images exhibited lower Fréchet Inception Distance (FID) scores (indicating higher visual fidelity), GAN-based augmentation yielded superior downstream classification performance. This suggests that image-quality metrics alone do not fully predict utility for specific diagnostic tasks.
Comparison with Clinicians: In the image-only comparison, the GAN-ResNet50+XGBoost ( $n=4$ ) model outperformed both the senior and junior clinicians in terms of per-image accuracy and F1-score.
Limitations Observed: Despite high accuracy, the model exhibited a false-negative rate of 18.0% for malignant tumors on the retrospective test set, highlighting the risk of missed diagnoses.

Key Contributions

Hybrid Learning Strategy: Demonstrates the efficacy of combining CNN feature extraction with classical machine learning classifiers (specifically XGBoost) in a limited-data medical imaging setting.
Synthetic Data Evaluation: Provides empirical evidence that synthetic data augmentation can improve classification performance in eyelid tumor diagnosis, while emphasizing that the choice of generative model (GAN vs. Diffusion) and the augmentation ratio ( $n$ ) must be tuned based on downstream task performance rather than just image quality metrics.
Prospective Validation: Includes a prospective test cohort, offering a more rigorous evaluation than purely retrospective studies.

Significance and Claims
The authors claim that SY-CNN+ offers a viable approach for eyelid tumor classification in data-scarce environments. The study supports the use of synthetic augmentation and hybrid CNN-based learning to improve diagnostic accuracy where specialist resources are limited.

However, the paper maintains a modest tone regarding clinical deployment. The authors explicitly state that the model should not be used as a standalone diagnostic tool due to the observed false-negative rate for malignant tumors. Instead, they propose its near-term role as a clinical decision support system to help identify suspicious lesions for further specialist review. The authors acknowledge that larger, multi-center cohorts, higher-resolution models, and broader clinician comparisons are necessary before the system can be considered for routine clinical use.