This paper describes KDE Lab approach in ImageCLEF GANs 2024. ImageCLEFmedical GANs 2024 task consists of two sub tasks, Identify training data fingerprints task and Detect Generative models'fingerprints task. Identify training data fingerprints task aims to examine the existing hypothesis that GANs are generating medical images that contain certain "fingerprints" of the real images used for generative network training. Detect Generative models'fingerprints task aims to investigate the hypothesis that generative models imprint distinctive "fingerprints" onto the generated images. We studied Identify training data fingerprints task in this research. For experiment, we attempted a methods that is approach using Multiple pre-training models and the results were compared. Also, we attempted a methods that is approach using super-resolution technique. Finally, the experiment was a failure and the results could not be compared. The competition result had an accuracy of 0.484.
In this research, we use some of the images given in the ‘Identify training data “fingerprints” Tasks’ of the ImageCLEFmedical GANs 2024. The image data are axial slices of 3D CT images of about 8000 pulmonary TB patients. These images are stored in the form of 256x256 pixel 8-bit/pixel PNG images. From that image data, we used a total of 10,200 images, 10,000 images generated from GAN model 1, which is not revealed, and 100 images each annotated as having been used/not used for training the image generation for that model. The 10,000 images generated from GAN model 2 that were not revealed and the 100 images each annotated as used/not used for training the image generation of that model were not used this time.
The images in the dataset are grayscale. Therefore, we attempted to colorize them using (tensorlfow.image.grayscale_to_rgb). To augment the data, horizontal inversion and ImageGenerator were applied as follows.
• zoom_range : 0.97-1.03 • rotation_range : 10 • height_shift_range : 0.05 • width_shift_range : 0.05 – Scaling is performed randomly within a scaling factor range of 0.97 1.03. – Randomly rotate the image in the range of -10 10 degrees. – Vertical translation in the range of x 0.05 pixels of the original image height.
– Left-right translation within a range of x 0.05 pixels of the original image’s width.
The default input channel is (224,224), so it was resized from (256,256) to (224,224). The horizontal inversion is performed only on the real image; 24 images are added in ImageGenerator for the image before and after horizontal inversion, respectively. The number of images can be reduced from 1 real image to 1 original image + 1 flipped image + 48 images converted by ImageGenerator = 50 images, which means that the data can be padded up to 10,000 images in total (1:1) when 200 real images are used. We thought that super-resolution technique would allow us to obtain more features that are not present in the real images but are prominent only in the generated images, so we performed super-resolution technique on the real images using ESRGAN [7] and then performed the same operation.
We attempted CNN approach with three diferent CNN that employed ResNet50, MobileNetV2 and
DenseNet-121. These CNNs used models pre-trained in imagenet(Method1-4). The input of the structure
is set to (
prediction_Negative
prediction_Positive Actual_Negative Actual_Positive
TN(True Negative) FN(False Negative)
FP(False Positive) TP(True Positive)
Accuracy, Precision, Recall, and F1 values were used as evaluation indices. They can be calculated
from the confusion_matrix of sklearn. The calculation formula is as follows(
+ + + +
+
+
2 · ·
·
(
The results obtained from the Method 1-4 approach were submitted and returned scores are shown in Table 3-5. Methods 2-4 failed to submit; Method 1 received a accuracy_score of 0.48425,precision_score of 0.480433, recall_score of 0.44825,f1_score of 0.4542. Failed of score was due to the fact that there were many of the same predicted results and that the method of submission was incorrect. The Method 1 values of Accuracy, Precision, Recall, and F1 were 0.484, 0.476, 0.317, 0.380 for Dataset1, 0.481, 0.484, 0.579, 0.527 for Dataset2.
This paper describes an approach to the training data fingerprint identification task in the ImageCLEF2024 GAN. We attempted to use multiple pre-training models and super-resolution techniques in our approach. However, this time, we were unable to produce correct values and could not compare the results for each model or image data with and without super-imaging. Therefore, it was not possible to identify features that are not present in the real images but are prominent only in the generated images. As a result of the competition, the CNN approach with EficientNetV2 achieved an accuracy of 0.484.
A part of this research was carried out with the support of the Grant for Toyohashi Heart Center Smart Hospital Joint Research Course and the Grant-in-Aid for Scientific Research (C) (issue numbers 22K12149 and 22K12040). [5] K. He, X. Zhang, S. Ren, J. Sun, Deep Residual Learning for Image Recognition, in: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016, pp. 770–778. [6] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, L. Fei-Fei, ImageNet: A large-scale hierarchical image database, in: 2009 IEEE Conference on Computer Vision and Pattern Recognition (CVPR 2009), 2009, pp. 248–255. doi:10.1109/CVPR.2009.5206848. [7] X. Wang, K. Yu, S. Wu, J. Gu, Y. Liu, C. Dong, C. C. Loy, Y. Qiao, X. Tang, Esrgan: Enhanced super-resolution generative adversarial networks, 2018. arXiv:1809.00219.