Gabriel Eilertsen

Nithesh Chandher Karthikeyan, Jonas Unger, Gabriel Eilertsen

Data availability remains a critical bottleneck in many deep learning applications. Large-scale datasets are often expensive to collect, curate and annotate, which can limit the scalability and applicability of supervised learning methods. In this work, we evaluate the classification performance of models trained on synthetic image datasets produced by generative deep learning. In particular, we use latent diffusion models conditioned on learned representations from DINOv2, DINOv3, and CLIP. Our results demonstrates that this representation-conditioned formulation significantly outperforms class-conditioned generation by a large margin (+10.76 p.p. top-1 accuracy on ImageNet100), by improving sample quality and mode coverage. Furthermore, by scaling the size of the synthetic dataset, we are able to outperform a classifier trained on the real data (+2.0 p.p top-1 accuracy). We also demonstrate how generated images can be used for augmentation purposes, outperforming classical augmentation methods, and how the conditioning space can be used for sample filtering to further improve training value. Collectively, these findings highlight that representation-conditioned diffusion models provide a promising approach for augmenting, complementing, or potentially replacing real-world datasets in large-scale visual learning tasks.

Nithesh Chandher Karthikeyan, Jonas Unger, Gabriel Eilertsen

Diffusion models have emerged as powerful tools for high-quality image generation and editing, but guiding these models to produce specific outputs remains a challenge. Conventional approaches rely on conditioning mechanisms, such as text prompts or semantic maps, which require extensively annotated datasets. In this preliminary work, we explore diffusion models conditioned on representations from a pre-trained self-supervised model. The self-conditioning mechanism not only improves the quality of unconditional image generation, but also provides a representation space that can be used to control the generation. We explore this conditioning space by identifying directions of variations, and demonstrate promising properties in terms of smoothness and disentanglement.

Rym Jaroudi, Lukáš Malý, Gabriel Eilertsen et al.

This paper presents the Standalone Neural ODE (sNODE), a continuous-depth neural ODE model capable of describing a full deep neural network. This uses a novel nonlinear conjugate gradient (NCG) descent optimization scheme for training, where the Sobolev gradient can be incorporated to improve smoothness of model weights. We also present a general formulation of the neural sensitivity problem and show how it is used in the NCG training. The sensitivity analysis provides a reliable measure of uncertainty propagation throughout a network, and can be used to study model robustness and to generate adversarial attacks. Our evaluations demonstrate that our novel formulations lead to increased robustness and performance as compared to ResNet models, and that it opens up for new opportunities for designing and developing machine learning with improved explainability.

Tahereh Dehdarirad, Michael Felsberg, Gabriel Eilertsen et al.

Variational autoencoders (VAEs) often struggle to represent non-commutative structure in learned latent spaces. Symmetry-aware VAEs commonly address this issue by enforcing commutativity through algebraic regularization, which is appropriate for commutative transformation groups but can suppress meaningful non-commutative structure when it is intrinsic to the data. We argue that non-commutativity should instead be explicitly diagnosed and reflected in reconstruction behavior. We introduce a Lie Group VAE framework that combines geometric and algebraic perspectives on uncertainty while separating discrete generative factors from continuous geometric transformations. In a first phase, the model is trained without structural constraints while algebraic non-commutativity is measured through finite Baker-Campbell-Hausdorff deviations and decoder order sensitivity is measured through reconstruction order-swap tests. These diagnostics reveal a scale mismatch between latent non-commutativity and reconstruction behavior under unconstrained training. In a second phase, we introduce a deformation-stability constraint with a data-driven calibration constant that aligns decoder sensitivity with algebraic non-commutativity. We evaluate the framework on dSprites, 3DShapes, 3DCars, and CelebA against generic and symmetry-aware baselines, including beta-VAE, CLG-VAE, and CFASL. Across synthetic benchmarks, the method improves reconstruction quality and yields decoder-level behavior more consistent with latent non-commutative structure. Qualitative analyses show clearer order-dependent latent compositions and more stable reconstructions. On CelebA, the model yields more faithful reconstructions and factor-specific latent traversals than CFASL, while also exhibiting meaningful order-dependent interactions between learned latent directions.

Yifan Ding, Arturas Aleksandraus, Amirhossein Ahmadian et al.

Out-of-distribution (OOD) detection is critical for ensuring the reliability of deep learning systems, particularly in safety-critical applications. Likelihood-based deep generative models have historically faced criticism for their unsatisfactory performance in OOD detection, often assigning higher likelihood to OOD data than in-distribution samples when applied to image data. In this work, we demonstrate that likelihood is not inherently flawed. Rather, several properties in the images space prohibit likelihood as a valid detection score. Given a sufficiently good likelihood estimator, specifically using the probability flow formulation of a diffusion model, we show that likelihood-based methods can still perform on par with state-of-the-art methods when applied in the representation space of pre-trained encoders. The code of our work can be found at $\href{https://github.com/limchaos/Likelihood-OOD.git}{\texttt{https://github.com/limchaos/Likelihood-OOD.git}}$.

Milda Pocevičiūtė, Gabriel Eilertsen, Stina Garvin et al.

Multiple-instance learning (MIL) is an attractive approach for digital pathology applications as it reduces the costs related to data collection and labelling. However, it is not clear how sensitive MIL is to clinically realistic domain shifts, i.e., differences in data distribution that could negatively affect performance, and if already existing metrics for detecting domain shifts work well with these algorithms. We trained an attention-based MIL algorithm to classify whether a whole-slide image of a lymph node contains breast tumour metastases. The algorithm was evaluated on data from a hospital in a different country and various subsets of this data that correspond to different levels of domain shift. Our contributions include showing that MIL for digital pathology is affected by clinically realistic differences in data, evaluating which features from a MIL model are most suitable for detecting changes in performance, and proposing an unsupervised metric named Fréchet Domain Distance (FDD) for quantification of domain shifts. Shift measure performance was evaluated through the mean Pearson correlation to change in classification performance, where FDD achieved 0.70 on 10-fold cross-validation models. The baselines included Deep ensemble, Difference of Confidence, and Representation shift which resulted in 0.45, -0.29, and 0.56 mean Pearson correlation, respectively. FDD could be a valuable tool for care providers and vendors who need to verify if a MIL system is likely to perform reliably when implemented at a new site, without requiring any additional annotations from pathologists.

Chao Wang, Francesco Banterle, Bin Ren et al.

This paper presents a comprehensive review of the AIM 2025 Challenge on Inverse Tone Mapping (ITM). The challenge aimed to push forward the development of effective ITM algorithms for HDR image reconstruction from single LDR inputs, focusing on perceptual fidelity and numerical consistency. A total of \textbf{67} participants submitted \textbf{319} valid results, from which the best five teams were selected for detailed analysis. This report consolidates their methodologies and performance, with the lowest PU21-PSNR among the top entries reaching 29.22 dB. The analysis highlights innovative strategies for enhancing HDR reconstruction quality and establishes strong benchmarks to guide future research in inverse tone mapping.

Shreyas Shivakumara, Gabriel Eilertsen, Karljohan Lundin Palmerius

Neural Radiance Fields (NeRFs) have demonstrated significant potential in synthesizing novel viewpoints. Evaluating the NeRF-generated outputs, however, remains a challenge due to the unique artifacts they exhibit, and no individual metric performs well across all datasets. We hypothesize that combining two successful metrics, Deep Image Structure and Texture Similarity (DISTS) and Video Multi-Method Assessment Fusion (VMAF), based on different perceptual methods, can overcome the limitations of individual metrics and achieve improved correlation with subjective quality scores. We experiment with two normalization strategies for the individual metrics and two fusion strategies to evaluate their impact on the resulting correlation with the subjective scores. The proposed pipeline is tested on two distinct datasets, Synthetic and Outdoor, and its performance is evaluated across three different configurations. We present a detailed analysis comparing the correlation coefficients of fusion methods and individual scores with subjective scores to demonstrate the robustness and generalizability of the fusion metrics.

Yifan Ding, Xixi Liu, Jonas Unger et al.

Out-of-distribution (OOD) detection is essential for the safe deployment of machine learning models. Recent advances have explored improved classification losses and representation learning strategies to enhance OOD detection. However, these methods are often tailored to specific post-hoc detection techniques, limiting their generalizability. In this work, we identify a critical issue in Logit Normalization (LogitNorm), which inhibits its effectiveness in improving certain post-hoc OOD detection methods. To address this, we propose Extended Logit Normalization ($\textbf{ELogitNorm}$), a novel hyperparameter-free formulation that significantly benefits a wide range of post-hoc detection methods. By incorporating feature distance-awareness to LogitNorm, $\textbf{ELogitNorm}$ shows more robust OOD separability and in-distribution (ID) confidence calibration than its predecessor. Extensive experiments across standard benchmarks demonstrate that our approach outperforms state-of-the-art training-time methods in OOD detection while maintaining strong ID classification accuracy.

Axel Gödrich, Daniel König, Gabriel Eilertsen et al.

Cameras digitize real-world scenes as pixel intensity values with a limited value range given by the available bits per pixel (bpp). High Dynamic Range (HDR) cameras capture those luminance values in higher resolution through an increase in the number of bpp. Most displays, however, are limited to 8 bpp. Naive HDR compression methods lead to a loss of the rich information contained in those HDR images. In this paper, tone mapping algorithms for thermal infrared images with 16 bpp are investigated that can preserve this information. An optimized multi-scale Retinex algorithm sets the baseline. This algorithm is then approximated with a deep learning approach based on the popular U-Net architecture. The remaining noise in the images after tone mapping is reduced implicitly by utilizing a self-supervised deep learning approach that can be jointly trained with the tone mapping approach in a multi-task learning scheme. Further discussions are provided on denoising and deflickering for thermal infrared video enhancement in the context of tone mapping. Extensive experiments on the public FLIR ADAS Dataset prove the effectiveness of our proposed method in comparison with the state-of-the-art.

George Baravdish, Gabriel Eilertsen, Rym Jaroudi et al.

The inverse problem of supervised reconstruction of depth-variable (time-dependent) parameters in a neural ordinary differential equation (NODE) is considered, that means finding the weights of a residual network with time continuous layers. The NODE is treated as an isolated entity describing the full network as opposed to earlier research, which embedded it between pre- and post-appended layers trained by conventional methods. The proposed parameter reconstruction is done for a general first order differential equation by minimizing a cost functional covering a variety of loss functions and penalty terms. A nonlinear conjugate gradient method (NCG) is derived for the minimization. Mathematical properties are stated for the differential equation and the cost functional. The adjoint problem needed is derived together with a sensitivity problem. The sensitivity problem can estimate changes in the network output under perturbation of the trained parameters. To preserve smoothness during the iterations the Sobolev gradient is calculated and incorporated. As a proof-of-concept, numerical results are included for a NODE and two synthetic datasets, and compared with standard gradient approaches (not based on NODEs). The results show that the proposed method works well for deep learning with infinite numbers of layers, and has built-in stability and smoothness.

Milda Pocevičiūtė, Gabriel Eilertsen, Sofia Jarkman et al.

Deep learning (DL) has shown great potential in digital pathology applications. The robustness of a diagnostic DL-based solution is essential for safe clinical deployment. In this work we evaluate if adding uncertainty estimates for DL predictions in digital pathology could result in increased value for the clinical applications, by boosting the general predictive performance or by detecting mispredictions. We compare the effectiveness of model-integrated methods (MC dropout and Deep ensembles) with a model-agnostic approach (Test time augmentation, TTA). Moreover, four uncertainty metrics are compared. Our experiments focus on two domain shift scenarios: a shift to a different medical center and to an underrepresented subtype of cancer. Our results show that uncertainty estimates increase reliability by reducing a model's sensitivity to classification threshold selection as well as by detecting between 70\% and 90\% of the mispredictions done by the model. Overall, the deep ensembles method achieved the best performance closely followed by TTA.

Karin Stacke, Jonas Unger, Claes Lundström et al.

Unsupervised learning has made substantial progress over the last few years, especially by means of contrastive self-supervised learning. The dominating dataset for benchmarking self-supervised learning has been ImageNet, for which recent methods are approaching the performance achieved by fully supervised training. The ImageNet dataset is however largely object-centric, and it is not clear yet what potential those methods have on widely different datasets and tasks that are not object-centric, such as in digital pathology. While self-supervised learning has started to be explored within this area with encouraging results, there is reason to look closer at how this setting differs from natural images and ImageNet. In this paper we make an in-depth analysis of contrastive learning for histopathology, pin-pointing how the contrastive objective will behave differently due to the characteristics of histopathology data. We bring forward a number of considerations, such as view generation for the contrastive objective and hyper-parameter tuning. In a large battery of experiments, we analyze how the downstream performance in tissue classification will be affected by these considerations. The results point to how contrastive learning can reduce the annotation effort within digital pathology, but that the specific dataset characteristics need to be considered. To take full advantage of the contrastive learning objective, different calibrations of view generation and hyper-parameters are required. Our results pave the way for realizing the full potential of self-supervised learning for histopathology applications.

Apostolia Tsirikoglou, Karin Stacke, Gabriel Eilertsen et al.

The scarcity of labeled data is a major bottleneck for developing accurate and robust deep learning-based models for histopathology applications. The problem is notably prominent for the task of metastasis detection in lymph nodes, due to the tissue's low tumor-to-non-tumor ratio, resulting in labor- and time-intensive annotation processes for the pathologists. This work explores alternatives on how to augment the training data for colon carcinoma metastasis detection when there is limited or no representation of the target domain. Through an exhaustive study of cross-validated experiments with limited training data availability, we evaluate both an inter-organ approach utilizing already available data for other tissues, and an intra-organ approach, utilizing the primary tumor. Both these approaches result in little to no extra annotation effort. Our results show that these data augmentation strategies can be an efficient way of increasing accuracy on metastasis detection, but fore-most increase robustness.

Gabriel Eilertsen, Saghi Hajisharif, Param Hanji et al.

Single-image high dynamic range (SI-HDR) reconstruction has recently emerged as a problem well-suited for deep learning methods. Each successive technique demonstrates an improvement over existing methods by reporting higher image quality scores. This paper, however, highlights that such improvements in objective metrics do not necessarily translate to visually superior images. The first problem is the use of disparate evaluation conditions in terms of data and metric parameters, calling for a standardized protocol to make it possible to compare between papers. The second problem, which forms the main focus of this paper, is the inherent difficulty in evaluating SI-HDR reconstructions since certain aspects of the reconstruction problem dominate objective differences, thereby introducing a bias. Here, we reproduce a typical evaluation using existing as well as simulated SI-HDR methods to demonstrate how different aspects of the problem affect objective quality metrics. Surprisingly, we found that methods that do not even reconstruct HDR information can compete with state-of-the-art deep learning methods. We show how such results are not representative of the perceived quality and that SI-HDR reconstruction needs better evaluation protocols.

Gabriel Eilertsen, Apostolia Tsirikoglou, Claes Lundström et al.

Insufficient training data is a major bottleneck for most deep learning practices, not least in medical imaging where data is difficult to collect and publicly available datasets are scarce due to ethics and privacy. This work investigates the use of synthetic images, created by generative adversarial networks (GANs), as the only source of training data. We demonstrate that for this application, it is of great importance to make use of multiple GANs to improve the diversity of the generated data, i.e. to sufficiently cover the data distribution. While a single GAN can generate seemingly diverse image content, training on this data in most cases lead to severe over-fitting. We test the impact of ensembled GANs on synthetic 2D data as well as common image datasets (SVHN and CIFAR-10), and using both DCGANs and progressively growing GANs. As a specific use case, we focus on synthesizing digital pathology patches to provide anonymized training data.

Milda Pocevičiūtė, Gabriel Eilertsen, Claes Lundström

Machine learning (ML) algorithms are optimized for the distribution represented by the training data. For outlier data, they often deliver predictions with equal confidence, even though these should not be trusted. In order to deploy ML-based digital pathology solutions in clinical practice, effective methods for detecting anomalous data are crucial to avoid incorrect decisions in the outlier scenario. We propose a new unsupervised learning approach for anomaly detection in histopathology data based on generative adversarial networks (GANs). Compared to the existing GAN-based methods that have been used in medical imaging, the proposed approach improves significantly on performance for pathology data. Our results indicate that histopathology imagery is substantially more complex than the data targeted by the previous methods. This complexity requires not only a more advanced GAN architecture but also an appropriate anomaly metric to capture the quality of the reconstructed images.

Milda Pocevičiūtė, Gabriel Eilertsen, Claes Lundström

Artificial intelligence (AI) has shown great promise for diagnostic imaging assessments. However, the application of AI to support medical diagnostics in clinical routine comes with many challenges. The algorithms should have high prediction accuracy but also be transparent, understandable and reliable. Thus, explainable artificial intelligence (XAI) is highly relevant for this domain. We present a survey on XAI within digital pathology, a medical imaging sub-discipline with particular characteristics and needs. The review includes several contributions. Firstly, we give a thorough overview of current XAI techniques of potential relevance for deep learning methods in pathology imaging, and categorise them from three different aspects. In doing so, we incorporate uncertainty estimation methods as an integral part of the XAI landscape. We also connect the technical methods to the specific prerequisites in digital pathology and present findings to guide future research efforts. The survey is intended for both technical researchers and medical professionals, one of the objectives being to establish a common ground for cross-disciplinary discussions.

Apostolia Tsirikoglou, Karin Stacke, Gabriel Eilertsen et al.

Digitization of histopathology slides has led to several advances, from easy data sharing and collaborations to the development of digital diagnostic tools. Deep learning (DL) methods for classification and detection have shown great potential, but often require large amounts of training data that are hard to collect, and annotate. For many cancer types, the scarceness of data creates barriers for training DL models. One such scenario relates to detecting tumor metastasis in lymph node tissue, where the low ratio of tumor to non-tumor cells makes the diagnostic task hard and time-consuming. DL-based tools can allow faster diagnosis, with potentially increased quality. Unfortunately, due to the sparsity of tumor cells, annotating this type of data demands a high level of effort from pathologists. Using weak annotations from slide-level images have shown great potential, but demand access to a substantial amount of data as well. In this study, we investigate mitigation strategies for limited data access scenarios. Particularly, we address whether it is possible to exploit mutual structure between tissues to develop general techniques, wherein data from one type of cancer in a particular tissue could have diagnostic value for other cancers in other tissues. Our case is exemplified by a DL model for metastatic colon cancer detection in lymph nodes. Could such a model be trained with little or even no lymph node data? As alternative data sources, we investigate 1) tumor cells taken from the primary colon tumor tissue, and 2) cancer data from a different organ (breast), either as is or transformed to the target domain (colon) using Cycle-GANs. We show that the suggested approaches make it possible to detect cancer metastasis with no or very little lymph node data, opening up for the possibility that existing, annotated histopathology data could generalize to other domains.

Gabriel Eilertsen, Daniel Jönsson, Timo Ropinski et al.

This paper presents an empirical study on the weights of neural networks, where we interpret each model as a point in a high-dimensional space -- the neural weight space. To explore the complex structure of this space, we sample from a diverse selection of training variations (dataset, optimization procedure, architecture, etc.) of neural network classifiers, and train a large number of models to represent the weight space. Then, we use a machine learning approach for analyzing and extracting information from this space. Most centrally, we train a number of novel deep meta-classifiers with the objective of classifying different properties of the training setup by identifying their footprints in the weight space. Thus, the meta-classifiers probe for patterns induced by hyper-parameters, so that we can quantify how much, where, and when these are encoded through the optimization process. This provides a novel and complementary view for explainable AI, and we show how meta-classifiers can reveal a great deal of information about the training setup and optimization, by only considering a small subset of randomly selected consecutive weights. To promote further research on the weight space, we release the neural weight space (NWS) dataset -- a collection of 320K weight snapshots from 16K individually trained deep neural networks.

Karin Stacke, Gabriel Eilertsen, Jonas Unger et al.

Domain shift is a significant problem in histopathology. There can be large differences in data characteristics of whole-slide images between medical centers and scanners, making generalization of deep learning to unseen data difficult. To gain a better understanding of the problem, we present a study on convolutional neural networks trained for tumor classification of H&E stained whole-slide images. We analyze how augmentation and normalization strategies affect performance and learned representations, and what features a trained model respond to. Most centrally, we present a novel measure for evaluating the distance between domains in the context of the learned representation of a particular model. This measure can reveal how sensitive a model is to domain variations, and can be used to detect new data that a model will have problems generalizing to. The results show how learning is heavily influenced by the preparation of training data, and that the latent representation used to do classification is sensitive to changes in data distribution, especially when training without augmentation or normalization.

Gabriel Eilertsen, Rafał K. Mantiuk, Jonas Unger

Convolutional neural networks (CNNs) can model complicated non-linear relations between images. However, they are notoriously sensitive to small changes in the input. Most CNNs trained to describe image-to-image mappings generate temporally unstable results when applied to video sequences, leading to flickering artifacts and other inconsistencies over time. In order to use CNNs for video material, previous methods have relied on estimating dense frame-to-frame motion information (optical flow) in the training and/or the inference phase, or by exploring recurrent learning structures. We take a different approach to the problem, posing temporal stability as a regularization of the cost function. The regularization is formulated to account for different types of motion that can occur between frames, so that temporally stable CNNs can be trained without the need for video material or expensive motion estimation. The training can be performed as a fine-tuning operation, without architectural modifications of the CNN. Our evaluation shows that the training strategy leads to large improvements in temporal smoothness. Moreover, for small datasets the regularization can help in boosting the generalization performance to a much larger extent than what is possible with naïve augmentation strategies.

Gabriel Eilertsen, Joel Kronander, Gyorgy Denes et al.

Camera sensors can only capture a limited range of luminance simultaneously, and in order to create high dynamic range (HDR) images a set of different exposures are typically combined. In this paper we address the problem of predicting information that have been lost in saturated image areas, in order to enable HDR reconstruction from a single exposure. We show that this problem is well-suited for deep learning algorithms, and propose a deep convolutional neural network (CNN) that is specifically designed taking into account the challenges in predicting HDR values. To train the CNN we gather a large dataset of HDR images, which we augment by simulating sensor saturation for a range of cameras. To further boost robustness, we pre-train the CNN on a simulated HDR dataset created from a subset of the MIT Places database. We demonstrate that our approach can reconstruct high-resolution visually convincing HDR results in a wide range of situations, and that it generalizes well to reconstruction of images captured with arbitrary and low-end cameras that use unknown camera response functions and post-processing. Furthermore, we compare to existing methods for HDR expansion, and show high quality results also for image based lighting. Finally, we evaluate the results in a subjective experiment performed on an HDR display. This shows that the reconstructed HDR images are visually convincing, with large improvements as compared to existing methods.