The ImageCLEFmed GAN 2025 task aims to detect whether the synthesized medical images contain "fingerprints" from the training data. In this paper, we adopted a similarity classification method based on image enhancement and deep learning models to determine which real images were used in the training process by comparing the similarity between real and synthetic images. We preprocessed the images using multiple image enhancement techniques (such as Gaussian filtering, Hessian matrix, Laplacian operator, and bilateral filtering). Then, we used convolutional neural networks (CNN) and ResNet50 models to extract image features and calculate the similarity between images. Through experiments on the validation set, our similarity classification method achieved excellent accuracy and F1 score performance. In the submitted results, our best F1 score was 0.633.The best kappa score is -0.016.It proves that the method used can efectively distinguish between "used" and "unused" images. Our experimental results show that it is possible to successfully identify the real images used to generate synthetic images through image enhancement and deep learning models.Our code is available at https://github.com/Qqiiiii/ImageCLEF.git.
In the field of medical image analysis, deep learning models[
Although generative models can significantly improve data diversity when generating synthetic
images, synthetic images may inadvertently expose sensitive information in the training data during the
process of learning data distribution, thus bringing the risk of privacy leakage.[
In this context, ImageCLEFmed GAN 2025[
Our method first performs multiple image enhancement processes on the image, including Gaussian
ifltering, Hessian matrix, Laplacian operator, and bilateral filtering to enhance the details and features
of the image.[
Experimental results on the validation set show that our similarity classification method achieves excellent performance in both accuracy and F1 score, with the best F1 score being 0.633. This proves that the adopted method can efectively identify real images used for training synthetic images and has great potential for application in privacy protection and synthetic image analysis.
This paper is organized as follows: Section 2 introduces the task and the dataset, Section 3 elaborates on our proposed method, Section 4 presents and discusses the experimental results, and finally, Section 5 summarizes the contributions of this paper and proposes future research directions.
The task introduced in ImageCLEFmed GAN 2025 aims to study whether specific real images are used to train the generative model to generate synthetic biomedical images. Participants are required to annotate each real image in the test set to indicate whether it was used to generate the corresponding synthetic image. Specifically, participants need to annotate each real image as "used" (1) or "not used" (0) to determine whether it participated in the training process of generating the image. This subtask aims to detect "fingerprints" in synthetic images, that is, to identify whether there are real image features in the generated image that can be traced back to the training data. The data of the training set and test set are shown in Table 1.
This task focuses on the potential privacy leakage and data security issues in the process of generating synthetic images. It explores whether the generative model can generate synthetic images that are highly similar to real patient images, which may lead to the leakage of training data. Participants need to analyze the test image dataset and evaluate whether certain real images are used in the training process of the generative model. To this end, the task’s dataset includes real and synthetic images.
Gaussian filter enhancement is a common image smoothing method that can efectively reduce the image’s noise and retain the image’s edge information. In the experiment, we applied Gaussian blur to smooth the details of the image and enhanced the image details through histogram equalization. This enhancement method helps the model better identify the key information in the image during image processing.
Laplacian operator enhancement is an edge detection method that uses second-order derivatives to identify edge information in an image. In the experiment, we used the Laplacian operator to extract edge regions in the image. This edge information plays an important role in subsequent feature extraction, helping the model focus on the details in the image, thereby improving the accuracy of similarity calculation.
The Hessian matrix enhancement method captures the local curvature information of the image by calculating the second-order derivative of the image, which is particularly suitable for edge and detail enhancement. We use the Hessian matrix to calculate the edge area of the image and further enhance the details in the image. Through this method, we can enhance the high-frequency information in the image, making the details in the image more prominent, thereby improving the quality of subsequent feature extraction.
Bilateral filtering is a smoothing method that can efectively preserve image edges, especially for images with complex textures. In our experiments, bilateral filtering is used to smooth areas in the image while keeping the edges sharp. By enhancing the details of the image, bilateral filtering improves the image’s visual efect and makes the image’s key information more obvious.
These four image enhancement methods enhance the image details from diferent perspectives and processing methods. Gaussian filtering focuses on denoising and preprocessing, the Laplacian operator emphasizes the edge information of the image, the Hessian matrix enhances the structural features of the image through curvature analysis, and bilateral filtering improves the details of the image through edge-preserving smoothing. These enhancement methods cooperate with each other in the feature extraction process, which can efectively improve the quality of the image and help the subsequent feature extraction and similarity calculation to be more accurate. The original generated image and the image after the four enhancement methods are shown in Figure 1. The comparison of image enhancement methods is shown in Table 2.
Feature extraction is a key step in our task. By extracting efective features from images, we can calculate the similarity between images and perform subsequent classification. In this task, we used convolutional neural networks (CNN) and ResNet50 models to automatically extract image features, taking advantage of the deep learning network’s ability to extract deep patterns in images.
Convolutional Neural Network (CNN) is a powerful deep learning model widely used in computer vision tasks, especially image classification, object detection, and image segmentation. In this task, CNN is used to automatically extract features from images.
CNN extracts features from images through multiple layers of convolution and pooling operations. The convolution layer extracts diferent features from the image, such as edges, corners, textures, etc., by performing convolution operations with local areas of the image. The pooling layer reduces the spatial size of the image and retains important feature information by performing dimensionality reduction on the convolution results. During the feature extraction process, CNN can gradually extract more advanced features through convolution and pooling operations at each layer, thereby capturing complex patterns and details in the image. We use the optimized CNN model to extract features from the image and use these features for subsequent similarity calculations and classification. The model consists of multiple convolutional, activation, and pooling layers. Through optimized structure and training methods, it can eficiently extract low-level and high-level features of the image.
The advantage of CNN is that it can automatically learn features. CNN can automatically learn image features through the back-propagation algorithm without manually designing feature extractors. It can also connect locally and share weights. The convolution operation of CNN greatly reduces the number of parameters by locally connecting and sharing weights, allowing the model to better handle complex images. It is also translation invariant. The convolution operation is invariant to the translation of the image, which means that CNN can recognize the same objects in the image regardless of their position.
ResNet50 is a deep convolutional neural network, which is designed based on the concept of "Residual Learning" and uses residual blocks to solve the gradient vanishing problem in deep networks. Due to its depth and efective structure, ResNet50 performs very well in image classification and feature extraction tasks, especially when processing large-scale image datasets.
ResNet50 consists of a 50-layer deep convolutional neural network, which mainly uses residual blocks to enhance the network’s expressiveness. Each residual block contains several convolutional layers but adds "skip connections" that allow signals to jump over certain layers directly, thereby avoiding the gradient vanishing problem that may occur in traditional deep networks. Through this residual learning mechanism, ResNet50 can train deeper networks and capture more complex features. In the feature extraction process, ResNet50 gradually learns the complex patterns in the image through the previous multi-layer convolution operations and finally outputs a set of high-dimensional features through the fully connected layer, which can efectively represent the visual information of the image. In our study, we used a custom ResNet50 model, which was trained on the dataset and can extract high-level features of the image.
The image features extracted by CNN and ResNet50 are all high-dimensional feature vectors, which play an important role in similarity calculation. We use these high-dimensional feature vectors as input for subsequent similarity calculations and judge their similarity by comparing the feature vector diferences between diferent images.
In the 2025 ImageCLEFmed GAN task, the choice of similarity calculation method is crucial to evaluate the similarity between generated images and real images. We used several common similarity calculation methods, including cosine similarity, structural similarity index (SSIM), and Jaccard similarity. These methods are widely used in image processing, machine learning, information retrieval, and other fields. They can efectively measure the similarity between image features and provide strong support for subsequent classification tasks.
Cosine similarity is commonly used when calculating image similarity, particularly in content-based image retrieval (CBIR) systems. Cosine similarity assesses the similarity between two vectors by measuring the cosine of the angle between them. The core idea is that the closer the directions of the two vectors, the more similar they are, regardless of their magnitudes. Calculating cosine similarity involves converting each image into a vector form. This typically entails flattening the pixel values of the image or features extracted from the image (such as color histograms, texture descriptors, shape features, etc.) into a one-dimensional vector. The cosine similarity value ranges from -1 to 1, where 1 indicates identical directions (very similar), 0 indicates orthogonality (no similarity), and -1 indicates completely opposite directions. Cosine similarity focuses on directional similarity, ignoring magnitude. In some cases, two images might be very similar in terms of certain feature ratios, but the absolute diferences in actual pixel values could be significant.
cos( ) =
A · B ‖A‖‖B‖ = √︁∑︀=1 ()2 ×
Where A and B are two vectors. A · B represents the dot product of vectors A and B. ‖A‖ and ‖B‖ represent the magnitudes of vectors A and B.
Structural Similarity (SSIM) is a more intuitive and efective method for calculating image similarity. SSIM considers images’ luminance, contrast, and structural information, allowing it to more accurately reflect the human visual system’s perception of image quality. SSIM first calculates the luminance diference between two images. The luminance comparison is achieved by calculating the mean values of the images, which reflects the overall brightness levels of the images. Next, SSIM calculates the contrast diference. The contrast comparison is achieved by calculating the standard deviation of the images; the greater the standard deviation, the higher the image contrast. Finally, SSIM compares the structural information of the two images. This step is achieved by calculating the covariance of the images. Covariance reflects the linear relationship between the images’ pixels, capturing the images’ structural characteristics.
SSIM(, ) =
(2 x · y + 1) (2 xy + 2) ( x2 + y2 + 1) ( x2 + y2 + 2)
Where and are corresponding blocks of the two images. and are the mean values of image blocks and . and are the standard deviations of image blocks and . is the covariance of image blocks and .
Jaccard Similarity is another commonly used method for measuring image similarity. Jaccard Similarity is a common method for measuring the similarity between two sets, especially for evaluating the overlap between objects or regions in an image. In image processing, Jaccard Similarity is often used to compare the similarity between two binary images, that is, to quantify their similarity by comparing the image’s feature regions (such as edges, targets, etc.). In image processing, Jaccard Similarity is often used to compare the features of an image (such as edges, textures, colors, etc.) or the overlapping parts of the pixel level of a binary image. The value of Jaccard Similarity is between 0 and 1, and the larger the value, the more similar the two images are. For this task, we use Jaccard Similarity to evaluate the similarity between synthetic and real images. We first convert the real and synthetic images into high-dimensional feature vectors through a feature extraction model (ResNet50). Then, we obtain the binary representation of the salient regions in the image by binarizing these feature vectors. We set a threshold. When the Jaccard similarity exceeds the threshold, the real image is considered to have participated in the generation process. Otherwise, the image is considered to have not been used. This way, we can label the real image as "used" or "unused". The advantage of Jaccard similarity is that it can intuitively reflect the degree of overlap of image features, especially for images or features after binarization. Therefore, when processing synthetic images, Jaccard similarity can efectively capture the common parts between images and help us identify similar areas between generated and real images. In addition, Jaccard similarity is not afected by the scale or specific size of the image, so it has a certain robustness when processing images of diferent sizes.
Jaccard(, ) = | ∩ | | ∪ | where and represent the sets of feature regions (such as edges, textures, or salient regions) of the two images, | ∩ | is the number of common elements (intersection) between the sets, and | ∪ | is the total number of elements in the union of the two sets.
We use the following evaluation indicators to evaluate the model’s performance: Kappa coeficient, accuracy, precision, recall, and F1 score. Since the Kappa coeficient can efectively measure the consistency between the model prediction and the true label, especially in the case of class imbalance, we use the Kappa coeficient as the main evaluation indicator. At the same time, the F1 score is used to provide a balanced evaluation of precision and recall. The definitions of these metrics are as follows:
We used two diferent models (CNN and ResNet50) and three similarity calculation methods (cosine similarity, SSIM, and Jaccard similarity) to conduct experiments, and the experimental results of diferent models and similarity calculation methods are shown in Table 3. We pay special attention to the Kappa coeficient because it can comprehensively evaluate the classification consistency of the model.
Our team submitted 5 results, with the best kappa value of -0.016 and the best F1 score of 0.633.The results show that the Kappa coeficient of the CNN model is low, indicating that the model’s predictions are less consistent with the actual labels. When using SSIM similarity, although the Kappa coeficient has improved, it still shows that the model has greater uncertainty when processing data. This may be related to the dataset’s deviation or the model’s overfitting. The ResNet50 model has the smallest Kappa coeficient in all experiments, especially under the Jaccard similarity calculation. The Kappa value is close to 0, indicating that its classification results are less consistent with the actual labels. Cosine similarity performs well on the CNN model. Although the Kappa coeficient is low, it performs well regarding recall, indicating that this method can efectively detect "used" images. SSIM similarity is better than cosine similarity, especially on the CNN model, which can balance precision and recall and obtain more stable results. SSIM considers the image’s structural information and may be more suitable for such tasks. Jaccard similarity has the worst efect on ResNet50, with an F1 score of 0.448 and a negative Kappa coeficient, indicating that this method may not be suitable for such tasks.
This paper proposes a method that combines image enhancement with deep learning models to detect the "fingerprint" of training data in synthetic biomedical images. We evaluate the similarity between real and synthetic images using cosine similarity, SSIM, and Jaccard similarity. Experimental results show that the CNN model performs best when combined with SSIM similarity calculation. Although the performance of the ResNet50 model is weak, it still has the potential for optimization. Future research can optimize the similarity calculation method, further improve the feature extraction ability of the deep learning model, and enhance the performance and robustness of the model. This study provides an efective method for the privacy protection of synthetic images and has good application prospects.
During the preparation of this work Chat-GPT-4o and Grammarly were used to check grammar and spelling. After using this tool, the author reviewed and edited the content as needed and takes full responsibility for the publication’s content. [11] P. Ma, H. Yuan, Y. Chen, H. Chen, G. Weng, Y. Liu, A laplace operator-based active contour model with improved image edge detection performance, Digital Signal Processing 151 (2024) 104550. [12] N. S. Awarayi, F. Twum, J. B. Hayfron-Acquah, K. Owusu-Agyemang, A bilateral filtering-based image enhancement for alzheimer disease classification using cnn, Plos one 19 (2024) e0302358.