Deep-learning techniques can improve the eficiency of medical diagnosis while challenging human experts' accuracy. However, the rationale behind these classifier's decisions is largely opaque, which is dangerous in sensitive applications such as healthcare. Case-based explanations explain the decision process behind these mechanisms by exemplifying similar cases using previous studies from other patients. Yet, these may contain personally identifiable information, which makes them impossible to share without violating patients' privacy rights. Previous works have used GANs to generate anonymous case-based explanations, which had limited visual quality. We solve this issue by employing a latent difusion model in a three-step procedure: generating a catalogue of synthetic images, removing the images that closely resemble existing patients, and using this anonymous catalogue during an explanation retrieval process. We evaluate the proposed method on the MIMIC-CXR-JPG dataset and achieve explanations that simultaneously have high visual quality, are anonymous, and retain their explanatory value.
CEUR ceur-ws.org
Various medical imaging techniques such as X-ray or MRI are important to detect and diagnose multiple medical conditions. These imaging modalities provide clinicians insights into patients’ health, facilitating accurate and timely diagnoses. In recent years, Deep Learning techniques have shown the capability to rival human performance in diverse diagnosis tasks while being more eficient. However, this leap brings a large challenge – the inherent lack of interpretability in Deep Learning models.
While these models exhibit remarkable diagnostic capabilities, their decision-making processes remain largely opaque, posing a barrier to understanding the rationale behind their predictions. This lack of interpretability raises legitimate concerns, especially in critical medical decisions where transparency is necessary for establishing trust.
One way to better understand deep-learning reasoning is through case-based explanations,
which explain by example, closely mimicking the rationale of human professionals. Using
these mechanisms, a doctor would have access to a prediction and a set of image explanations
for each machine-made diagnosis, allowing the professional to compare the current situation
with previous historical examples. Ideally, these explanations should be shareable between
medical professionals, doctors, and patients or even between hospitals, which is particularly
important to build a diverse catalogue of explanations. Yet, this is not possible since medical
images contain personally identifiable data, which may allow for the re-identification of patients;
therefore, they are protected under strict regulations such as the GDPR [
The generation of anonymous case-based explanations using GAN architectures has been
previously explored [
This work combines three distinct research areas: anonymization techniques, difusion models and explainability. Although some works connect the bridge between two of these topics, the combination of all three is still largely unexplored.
Difusion Models Recently, difusion models such as DDPM [ 7] have emerged as an alternative to GANs to generate high-quality images while avoiding common dificulties of GANs, such as mode collapse. The difusion method consists of two distinct processes. During the forward process, Gaussian noise is iteratively added to a training image. In contrast, in the reverse process, a Deep-learning model, typically a U-Net [8], starting from pure noise, predicts what noise was added at each step and iteratively removes it, generating a new image.
Since the difusion process is computationally expensive to scale to higher resolutions, instead
of performing the difusion process on the image space, Latent Difusion Models (LDM) [
Anonymization Visual anonymization methods can generally be divided into classical and
machine-learning-based. Within the classical methods, two of the most common anonymization
techniques include blurring the image or K-Same [11], which overlays images, therefore
obtaining K-Anonymity. Yet, these methods require aggressive image modifications, making
them dificult to analyse visually. Recently, there has been a focus on generating synthetic
images using machine learning methods. These methods can be divided into two sub-groups:
diferential-privacy methods [ 12, 13], which provide strong privacy guarantees at the cost of
having a reduced image quality, struggling to be scaled beyond 32 × 32 pixel images due to
their computational complexity. The other methods present a more ad-hoc approach; instead of
providing strong statistical guarantees, they typically employ identity verification networks [
Case-based Explanations Case-based explanations are a post hoc mechanism which justifies
a model’s decisions by retrieving cases or examples from the training data. This method is
analogue to human reasoning, making them easy to interpret. While certain image-retrieval
scenarios employ similarity metrics such as the Euclidean distance or the Structural similarity
index measure (SSIM) [14], which analyses an image as a whole, for case-based explanations,
the focus is set on small localized features which would be overlooked by said measures. One
way of retrieving relevant images is by comparing the features obtained by a task-specific
Deep-Learning classifier [ 15]. Additionally, Montenegro et al. [
Our methodology, summarized in Figure 1, follows the anonymization protocol defined by
Packhäuser et al. [
We train a latent difusion model which uses the Medfusion [ 10] architecture for image generation. Using it, we sample synthetic images to build a synthetic dataset that will be anonymized.
We must identify if each synthetic image closely resembles any real patient in the training set. To do so, first, a patient retrieval network consists of a Siamese Neural Network (SNN) [ 16] based on the ResNet-50 [17] architecture, which takes as input two images and, using a contrastive loss function, approximates embeddings so that the embeddings from the same patients are grouped together.
The verification model follows the same architecture, but instead of using a contrastive loss, this network compares both input images and classifies whether or not they belong to the same patient. For both models, the training data consists of positive training pairs obtained based on the patient identity information of a dataset and randomized negative pairs.
Considering the synthetic dataset, the verification model, and the retrieval network, we have all the components required to obtain an anonymous dataset. The anonymization procedure we employ can be broken down into three separate steps: 1. Compute training identity embeddings: We obtain an identity embedding for each training set image using the identity retrieval network. These embeddings are stored in a KD-tree [18], used for more eficient queries using a nearest-neighbour strategy later. 2. Search nearest training image: For each synthetic sample, we retrieve the top-1 most similar training sample. In this step, we compute the identity embedding of each image and perform a lookup of the previously mentioned KD-tree. Each real-synthetic image pair is stored in a list used in the next step. 3. Remove non-anonymous image pairs: For each real-synthetic image pair, we use the identity verification network to obtain a prediction of whether or not both images belong to the same patient. If this likelihood exceeds a predefined threshold, the synthetic image is deemed non-anonymous and removed from the synthetic dataset.
We aim to retrieve explanations similar in task-related features and not merely structurally similar. We employ a DenseNet121 [19] classifier and use the feature vector present in the last layer, which contains 1024 features, to compare images. The intuition behind this method is that the classifier will represent the most important diagnosis characteristics in its feature space.
During inference, an image is passed through the classifier, and we use the features to obtain the most similar features from a catalogue of synthetic images. The catalogue size will vary depending on the application. In a real-case scenario, we want to have as many images as possible to capture the most variability and obtain images that are as relevant as possible.
Data We perform experiments on the MIMIC-CXR-JPG dataset [
Image Generation We generate MIMIC-CXR-JPG images using the Medfusion [10]
architecture, the latent embeddings are encoded using a VAE, which accepts 3 × 512 × 512 pixel images
and has 8 embedding channels. The difusion model uses DDIM [ 21], has label embedding and
time embeddings of size 1024 and employs a scaled linear noise scheduler with = 1000 , which
varies from Β0 = 0.002, Β = 0.02. Using a U-Net [8] architecture, we train the model for a
maximum of 1001 epochs with early stopping with a patience of 30 epochs. We generate 1000
synthetic images for both datasets, with a balanced split between positive and negative cases.
Retrieval and Verification For the retrieval phase, similarly to Packhäuser et al. [
Explanation Retrieval For explanation retrieval, we employ a DenseNet121 [19] network trained using the Adam optimizer with a learning rate of 1 −3 up to 10 epochs with an early stopping criteria and patience = 3 . To search for similar images, we use the last feature vector of the model, located before the dense layer and sized 1 × 1024.
(FID) [23], which compares the distribution of real and generated images based on the deepest layer of an Inception V3 network. To evaluate the image utility for classification tasks, we compare a classifier trained on synthetic data with a classifier trained on real data in Table 2.
The classifier trained on synthetic data still achieves acceptable performance even though it has been trained on a comparatively small dataset. This indicates that the signal required to diagnose Cardiomegaly is still present in the new images. model, we report the Area under the Receiver operating characteristic (AUC), precision (P), recall (R) and F1-score. For the retrieval task, we track the R-precision, a common information retrieval metric which is calculated as shown in Equation 1 using the number of relevant images retrieved ( ) within the first images, where is the total number of relevant images existing , shown in Equation 2, is the mean of the average precision ) for each query.
R-Precision =
1 =1
94.27
From the original 1,000 synthetic images, we use the anonymization procedure proposed in Section 3.3 to remove all synthetic images similar to their closest real counterpart based on the predictions made by the verification system. Using this process, 161 images were removed. In pairs are more similar than those within the lower ranges. ySntheic set, sorted by the predicted likelihood of belonging to the same patient. The images on the left are less likely to contain identifiable patient data, while images on the right are highly likely.
Similar to previous studies [15], the explanation ranking is evaluated with the help of a radiologist. For this purpose, we perform two diferent experiments. For the first experiment, we used 5 test cases, each containing a test image we aim to diagnose and 10 synthetic catalogue images from which we will retrieve explanations. The second experiment difers by using a catalogue of 5 synthetic images and 5 real images to compare the utility of both image types. All the images were randomly sampled from the test set while ensuring a balanced class distribution. The catalogue is limited to 10 images due to the time-consuming nature of evaluating the images.
To evaluate ranking performance, we use a metric commonly used in ranking tasks, the normalized Discounted Cumulative Gain (nDCG ) [24] metric (Equation 3) where is the number of retrieved images. The relevance values for each image ( ) vary from 5.5, the most similar image, to 1.0, the least similar example, with a 0.5 step between each relevance value. And IDCG is the ideal DCG value.
The results of the ranking evaluation are shown in Figure 3. We can notice that our CNNbased method, on average, outperforms its SSIM counterpart. An example of the retrieval test we used to evaluate the solution is shown in Figure 4. Interestingly, the CNN-based retrieval mechanism obtained the same Top-3 images as the expert-based choices, although in a diferent order.
Additionally, a radiologist accessed each image belonging to the test cases. In Table 5, we see that the generative model can conditionally generate images with a specific diagnosis while maintaining a similar class agreement compared to real images. We can also note that synthetic and real images are similarly relevant, supporting the fact that the generated images can be used as alternatives to real ones.
Finally, this method still presents some limitations. There is no upper bound on the number nDCG =
DCG IDCG (3)
DCG = ∑
2rel − 1 log2( + 1) (4) 4
7 G
G C
C D
D n
n setT CNN
SSIM
CNN
SSIM
CNN
SSIM b-asdExpert NC-based ISM-based method. The color outline around each retrieved image indicates whether the test and catalogue images share the same class (green) or not (red). The ranking annotated on the top-right corner of each image indicates the ground truth based on expert rating. of synthetic images that will be removed, leading to a waste of computing power on generating images that will efectively be deleted. In future work, the anonymization mechanism could be integrated into the generative model to combat this issue. Another key issue is the empirical nature of this anonymization procedure, making it unproven. Unfortunately, the current best method to provide anonymization, with strong guarantees, is through Diferential Privacy, which struggles to scale beyond low-resolution images, making it inefective in generating explanations which are interpretable by humans.
We proposed a method to generate visually anonymous case-based explanations that retain their utility and realism by levering latent difusion models. This solution demonstrates that the generated images are empirically unlikely to be traced back to the original patients, allowing for the visual explanation of classifier decisions without exposing patient data. This is a step towards integrating trustworthy and interpretable classifiers in the medical domain while preserving patient privacy.
This work is financed by National Funds through the FCT - Fundação para a Ciência e a Tecnologia, I.P. (Portuguese Foundation for Science and Technology) within the project CAGING, with reference 2022.10486.PTDC (DOI 10.54499/2022.10486.PTDC). [7] A. Nichol, P. Dhariwal, Improved Denoising Difusion Probabilistic Models, 2021. doi: 10.
48550/arXiv.2102.09672. arXiv:2102.09672. [8] O. Ronneberger, P. Fischer, T. Brox, U-net: Convolutional networks for biomedical image segmentation, CoRR (2015). arXiv:1505.04597.
[9] D. P. Kingma, M. Welling, Auto-encoding variational bayes, 2022. arXiv:1312.6114. [10] G. Müller-Franzes, J. M. Niehues, F. Khader, S. T. Arasteh, C. Haarburger, C. Kuhl, T. Wang, T. Han, T. Nolte, S. Nebelung, J. N. Kather, D. Truhn, A multimodal comparison of latent denoising difusion probabilistic models and generative adversarial networks for medical image synthesis, Scientific Reports 13 (2023) 12098. doi: 10.1038/s41598-023-39278-0. [11] R. Gross, E. Airoldi, B. Malin, L. Sweeney, Integrating utility into face de-identification, in: G. Danezis, D. Martin (Eds.), Privacy Enhancing Technologies, Springer Berlin Heidelberg, Berlin, Heidelberg, 2006, pp. 227–242. [12] T. Dockhorn, T. Cao, A. Vahdat, K. Kreis, Diferentially Private Difusion Models, Transactions on Machine Learning Research (2023). URL: https://openreview.net/forum?id= ZPpQk7FJXF. [13] Z. Chu, J. He, D. Peng, X. Zhang, N. Zhu, Diferentially private denoise difusion probability models, IEEE Access 11 (2023) 108033–108040. doi:10.1109/ACCESS.2023.3315592. [14] Z. Wang, A. Bovik, H. Sheikh, E. Simoncelli, Image quality assessment: from error visibility to structural similarity, IEEE Transactions on Image Processing 13 (2004) 600–612. [15] W. Silva, A. Poellinger, J. S. Cardoso, M. Reyes, Interpretability-guided content-based medical image retrieval, in: A. L. Martel, P. Abolmaesumi, D. Stoyanov, D. Mateus, M. A. Zuluaga, S. K. Zhou, D. Racoceanu, L. Joskowicz (Eds.), Medical Image Computing and Computer Assisted Intervention – MICCAI 2020, Springer International Publishing, Cham, 2020, pp. 305–314. [16] J. Bromley, I. Guyon, Y. LeCun, E. Säckinger, R. Shah, Signature verification using a “siamese” time delay neural network, in: Proceedings of the 6th International Conference on Neural Information Processing Systems, NIPS’93, Morgan Kaufmann Publishers Inc., San Francisco, CA, USA, 1993, p. 737–744. [17] K. He, X. Zhang, S. Ren, J. Sun, Deep residual learning for image recognition, CoRR abs/1512.03385 (2015). arXiv:1512.03385. [18] J. L. Bentley, Multidimensional binary search trees used for associative searching, Commun.
ACM 18 (1975) 509–517. doi:10.1145/361002.361007. [19] G. Huang, Z. Liu, L. Van Der Maaten, K. Q. Weinberger, Densely connected convolutional networks, in: 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), IEEE, 2017, pp. 2261–2269. doi:10.1109/CVPR.2017.243. [20] E. Puddy, C. Hill, Interpretation of the chest radiograph, Continuing Education in
Anaesthesia Critical Care & Pain 7 (2007) 71–75. doi:10.1093/bjaceaccp/mkm014. [21] J. Song, C. Meng, S. Ermon, Denoising difusion implicit models, CoRR abs/2010.02502 (2020). arXiv:2010.02502. [22] D. P. Kingma, J. Ba, Adam: A method for stochastic optimization, 2017. arXiv:1412.6980. [23] M. Heusel, H. Ramsauer, T. Unterthiner, B. Nessler, S. Hochreiter, Gans trained by a two time-scale update rule converge to a local nash equilibrium, 2018. arXiv:1706.08500. [24] K. Fernandes, J. S. Cardoso, Hypothesis transfer learning based on structural model similarity, Neural Computing and Applications 31 (2019) 3417–3430.