Ashkan Khakzar

Ario Sadafi, Oleksandra Adonkina, Ashkan Khakzar et al.

Explainability is a key requirement for computer-aided diagnosis systems in clinical decision-making. Multiple instance learning with attention pooling provides instance-level explainability, however for many clinical applications a deeper, pixel-level explanation is desirable, but missing so far. In this work, we investigate the use of four attribution methods to explain a multiple instance learning models: GradCAM, Layer-Wise Relevance Propagation (LRP), Information Bottleneck Attribution (IBA), and InputIBA. With this collection of methods, we can derive pixel-level explanations on for the task of diagnosing blood cancer from patients' blood smears. We study two datasets of acute myeloid leukemia with over 100 000 single cell images and observe how each attribution method performs on the multiple instance learning architecture focusing on different properties of the white blood single cells. Additionally, we compare attribution maps with the annotations of a medical expert to see how the model's decision-making differs from the human standard. Our study addresses the challenge of implementing pixel-level explainability in multiple instance learning models and provides insights for clinicians to better understand and trust decisions from computer-aided diagnosis systems.

Matan Atad, Vitalii Dmytrenko, Yitong Li et al.

Deep learning models used in medical image analysis are prone to raising reliability concerns due to their black-box nature. To shed light on these black-box models, previous works predominantly focus on identifying the contribution of input features to the diagnosis, i.e., feature attribution. In this work, we explore counterfactual explanations to identify what patterns the models rely on for diagnosis. Specifically, we investigate the effect of changing features within chest X-rays on the classifier's output to understand its decision mechanism. We leverage a StyleGAN-based approach (StyleEx) to create counterfactual explanations for chest X-rays by manipulating specific latent directions in their latent space. In addition, we propose EigenFind to significantly reduce the computation time of generated explanations. We clinically evaluate the relevancy of our counterfactual explanations with the help of radiologists. Our code is publicly available.

Ashkan Khakzar, Pedram Khorsandi, Rozhin Nobahari et al.

It is a mystery which input features contribute to a neural network's output. Various explanation (feature attribution) methods are proposed in the literature to shed light on the problem. One peculiar observation is that these explanations (attributions) point to different features as being important. The phenomenon raises the question, which explanation to trust? We propose a framework for evaluating the explanations using the neural network model itself. The framework leverages the network to generate input features that impose a particular behavior on the output. Using the generated features, we devise controlled experimental setups to evaluate whether an explanation method conforms to an axiom. Thus we propose an empirical framework for axiomatic evaluation of explanation methods. We evaluate well-known and promising explanation solutions using the proposed framework. The framework provides a toolset to reveal properties and drawbacks within existing and future explanation solutions.

Jindong Gu, Xiaojun Jia, Pau de Jorge et al.

The emergence of Deep Neural Networks (DNNs) has revolutionized various domains by enabling the resolution of complex tasks spanning image recognition, natural language processing, and scientific problem-solving. However, this progress has also brought to light a concerning vulnerability: adversarial examples. These crafted inputs, imperceptible to humans, can manipulate machine learning models into making erroneous predictions, raising concerns for safety-critical applications. An intriguing property of this phenomenon is the transferability of adversarial examples, where perturbations crafted for one model can deceive another, often with a different architecture. This intriguing property enables black-box attacks which circumvents the need for detailed knowledge of the target model. This survey explores the landscape of the adversarial transferability of adversarial examples. We categorize existing methodologies to enhance adversarial transferability and discuss the fundamental principles guiding each approach. While the predominant body of research primarily concentrates on image classification, we also extend our discussion to encompass other vision tasks and beyond. Challenges and opportunities are discussed, highlighting the importance of fortifying DNNs against adversarial vulnerabilities in an evolving landscape.

Ashkan Khakzar, Yawei Li, Yang Zhang et al.

One challenging property lurking in medical datasets is the imbalanced data distribution, where the frequency of the samples between the different classes is not balanced. Training a model on an imbalanced dataset can introduce unique challenges to the learning problem where a model is biased towards the highly frequent class. Many methods are proposed to tackle the distributional differences and the imbalanced problem. However, the impact of these approaches on the learned features is not well studied. In this paper, we look deeper into the internal units of neural networks to observe how handling data imbalance affects the learned features. We study several popular cost-sensitive approaches for handling data imbalance and analyze the feature maps of the convolutional neural networks from multiple perspectives: analyzing the alignment of salient features with pathologies and analyzing the pathology-related concepts encoded by the networks. Our study reveals differences and insights regarding the trained models that are not reflected by quantitative metrics such as AUROC and AP and show up only by looking at the models through a lens.

Matthias Keicher, Kamilia Zaripova, Tobias Czempiel et al.

The automation of chest X-ray reporting has garnered significant interest due to the time-consuming nature of the task. However, the clinical accuracy of free-text reports has proven challenging to quantify using natural language processing metrics, given the complexity of medical information, the variety of writing styles, and the potential for typos and inconsistencies. Structured reporting and standardized reports, on the other hand, can provide consistency and formalize the evaluation of clinical correctness. However, high-quality annotations for structured reporting are scarce. Therefore, we propose a method to predict clinical findings defined by sentences in structured reporting templates, which can be used to fill such templates. The approach involves training a contrastive language-image model using chest X-rays and related free-text radiological reports, then creating textual prompts for each structured finding and optimizing a classifier to predict clinical findings in the medical image. Results show that even with limited image-level annotations for training, the method can accomplish the structured reporting tasks of severity assessment of cardiomegaly and localizing pathologies in chest X-rays.

Yang Zhang, Yawei Li, Hannah Brown et al.

Feature attribution explains neural network outputs by identifying relevant input features. The attribution has to be faithful, meaning that the attributed features must mirror the input features that influence the output. One recent trend to test faithfulness is to fit a model on designed data with known relevant features and then compare attributions with ground truth input features.This idea assumes that the model learns to use all and only these designed features, for which there is no guarantee. In this paper, we solve this issue by designing the network and manually setting its weights, along with designing data. The setup, AttributionLab, serves as a sanity check for faithfulness: If an attribution method is not faithful in a controlled environment, it can be unreliable in the wild. The environment is also a laboratory for controlled experiments by which we can analyze attribution methods and suggest improvements.

Yawei Li, Yang Zhang, Kenji Kawaguchi et al.

Feature attribution methods attempt to explain neural network predictions by identifying relevant features. However, establishing a cohesive framework for assessing feature attribution remains a challenge. There are several views through which we can evaluate attributions. One principal lens is to observe the effect of perturbing attributed features on the model's behavior (i.e., faithfulness). While providing useful insights, existing faithfulness evaluations suffer from shortcomings that we reveal in this paper. In this work, we propose two new perspectives within the faithfulness paradigm that reveal intuitive properties: soundness and completeness. Soundness assesses the degree to which attributed features are truly predictive features, while completeness examines how well the resulting attribution reveals all the predictive features. The two perspectives are based on a firm mathematical foundation and provide quantitative metrics that are computable through efficient algorithms. We apply these metrics to mainstream attribution methods, offering a novel lens through which to analyze and compare feature attribution methods.

Adam Davies, Ashkan Khakzar

Artificial neural networks have long been understood as "black boxes": though we know their computation graphs and learned parameters, the knowledge encoded by these weights and functions they perform are not inherently interpretable. As such, from the early days of deep learning, there have been efforts to explain these models' behavior and understand them internally; and recently, mechanistic interpretability (MI) has emerged as a distinct research area studying the features and implicit algorithms learned by foundation models such as large language models. In this work, we aim to ground MI in the context of cognitive science, which has long struggled with analogous questions in studying and explaining the behavior of "black box" intelligent systems like the human brain. We leverage several important ideas and developments in the history of cognitive science to disentangle divergent objectives in MI and indicate a clear path forward. First, we argue that current methods are ripe to facilitate a transition in deep learning interpretation echoing the "cognitive revolution" in 20th-century psychology that shifted the study of human psychology from pure behaviorism toward mental representations and processing. Second, we propose a taxonomy mirroring key parallels in computational neuroscience to describe two broad categories of MI research, semantic interpretation (what latent representations are learned and used) and algorithmic interpretation (what operations are performed over representations) to elucidate their divergent goals and objects of study. Finally, we elaborate the parallels and distinctions between various approaches in both categories, analyze the respective strengths and weaknesses of representative works, clarify underlying assumptions, outline key challenges, and discuss the possibility of unifying these modes of interpretation under a common framework.

Constantin Venhoff, Ashkan Khakzar, Sonia Joseph et al.

Training vision language models (VLMs) aims to align visual representations from a vision encoder with the textual representations of a pretrained large language model (LLM). However, many VLMs exhibit reduced factual recall performance compared to their LLM backbones, raising the question of how effective multimodal fine-tuning is at extending existing mechanisms within the LLM to visual inputs. We argue that factual recall based on visual inputs requires VLMs to solve a two-hop problem: (1) forming entity representations from visual inputs, and (2) recalling associated factual knowledge based on these entity representations. By benchmarking 14 VLMs with various architectures (LLaVA, Native, Cross-Attention), sizes (7B-124B parameters), and training setups on factual recall tasks against their original LLM backbone models, we find that 11 of 14 models exhibit factual recall degradation. We select three models with high and two models with low performance degradation, and use attribution patching, activation patching, and probing to show that degraded VLMs struggle to use the existing factual recall circuit of their LLM backbone, because they resolve the first hop too late in the computation. In contrast, high-performing VLMs resolve entity representations early enough to reuse the existing factual recall mechanism. Finally, we demonstrate two methods to recover performance: patching entity representations from the LLM backbone into the VLM, and prompting with chain-of-thought reasoning. Our results highlight that the speed of early entity resolution critically determines how effective VLMs are in using preexisting LLM mechanisms. More broadly, our work illustrates how mechanistic analysis can explain and unveil systematic failures in multimodal alignment.

Lachin Naghashyar, Hunar Batra, Ashkan Khakzar et al.

Contemporary Vision-Language Models (VLMs) achieve strong performance on a wide range of tasks by pairing a vision encoder with a pre-trained language model, fine-tuned for visual-text inputs. Yet despite these gains, it remains unclear how language backbone representations adapt during multimodal training and when vision-specific capabilities emerge. In this work, we present the first mechanistic analysis of VLM adaptation. Using stage-wise model diffing, a technique that isolates representational changes introduced during multimodal fine-tuning, we reveal how a language model learns to "see". We first identify vision-preferring features that emerge or reorient during fine-tuning. We then show that a selective subset of these features reliably encodes spatial relations, revealed through controlled shifts to spatial prompts. Finally, we trace the causal activation of these features to a small group of attention heads. Our findings show that stage-wise model diffing reveals when and where spatially grounded multimodal features arise. It also provides a clearer view of modality fusion by showing how visual grounding reshapes features that were previously text-only. This methodology enhances the interpretability of multimodal training and provides a foundation for understanding and refining how pretrained language models acquire vision-grounded capabilities.

Runtao Liu, Ashkan Khakzar, Jindong Gu et al.

With the ability to generate high-quality images, text-to-image (T2I) models can be exploited for creating inappropriate content. To prevent misuse, existing safety measures are either based on text blacklists, which can be easily circumvented, or harmful content classification, requiring large datasets for training and offering low flexibility. Hence, we propose Latent Guard, a framework designed to improve safety measures in text-to-image generation. Inspired by blacklist-based approaches, Latent Guard learns a latent space on top of the T2I model's text encoder, where it is possible to check the presence of harmful concepts in the input text embeddings. Our proposed framework is composed of a data generation pipeline specific to the task using large language models, ad-hoc architectural components, and a contrastive learning strategy to benefit from the generated data. The effectiveness of our method is verified on three datasets and against four baselines. Code and data will be shared at https://latentguard.github.io/.

Yang Zhang, Er Jin, Yanfei Dong et al.

Diffusion models have achieved impressive advancements in various vision tasks. However, these gains often rely on increasing model size, which escalates computational complexity and memory demands, complicating deployment, raising inference costs, and causing environmental impact. While some studies have explored pruning techniques to improve the memory efficiency of diffusion models, most existing methods require extensive retraining to retain the model performance. Retraining a modern large diffusion model is extremely costly and resource-intensive, which limits the practicality of these methods. In this work, we achieve low-cost diffusion pruning without retraining by proposing a model-agnostic structural pruning framework for diffusion models that learns a differentiable mask to sparsify the model. To ensure effective pruning that preserves the quality of the final denoised latent, we design a novel end-to-end pruning objective that spans the entire diffusion process. As end-to-end pruning is memory-intensive, we further propose time step gradient checkpointing, a technique that significantly reduces memory usage during optimization, enabling end-to-end pruning within a limited memory budget. Results on state-of-the-art U-Net diffusion models SDXL and diffusion transformers (FLUX) demonstrate that our method can effectively prune up to 20% parameters with minimal perceptible performance degradation, and notably, without the need for model retraining. We also showcase that our method can still prune on top of time step distilled diffusion models.

Runtao Liu, I Chieh Chen, Jindong Gu et al.

Text-to-image (T2I) models are widespread, but their limited safety guardrails expose end users to harmful content and potentially allow for model misuse. Current safety measures are typically limited to text-based filtering or concept removal strategies, able to remove just a few concepts from the model's generative capabilities. In this work, we introduce AlignGuard, a method for safety alignment of T2I models. We enable the application of Direct Preference Optimization (DPO) for safety purposes in T2I models by synthetically generating a dataset of harmful and safe image-text pairs, which we call CoProV2. Using a custom DPO strategy and this dataset, we train safety experts, in the form of low-rank adaptation (LoRA) matrices, able to guide the generation process away from specific safety-related concepts. Then, we merge the experts into a single LoRA using a novel merging strategy for optimal scaling performance. This expert-based approach enables scalability, allowing us to remove 7x more harmful concepts from T2I models compared to baselines. AlignGuard consistently outperforms the state-of-the-art on many benchmarks and establishes new practices for safety alignment in T2I networks. Code and data will be shared at https://safetydpo.github.io/.

Xingyi Yang, Constantin Venhoff, Ashkan Khakzar et al.

Neurons in large language models often exhibit \emph{polysemanticity}, simultaneously encoding multiple unrelated concepts and obscuring interpretability. Instead of relying on post-hoc methods, we present \textbf{MoE-X}, a Mixture-of-Experts (MoE) language model designed to be \emph{intrinsically} interpretable. Our approach is motivated by the observation that, in language models, wider networks with sparse activations are more likely to capture interpretable factors. However, directly training such large sparse networks is computationally prohibitive. MoE architectures offer a scalable alternative by activating only a subset of experts for any given input, inherently aligning with interpretability objectives. In MoE-X, we establish this connection by rewriting the MoE layer as an equivalent sparse, large MLP. This approach enables efficient scaling of the hidden size while maintaining sparsity. To further enhance interpretability, we enforce sparse activation within each expert and redesign the routing mechanism to prioritize experts with the highest activation sparsity. These designs ensure that only the most salient features are routed and processed by the experts. We evaluate MoE-X on chess and natural language tasks, showing that it achieves performance comparable to dense models while significantly improving interpretability. MoE-X achieves a perplexity better than GPT-2, with interpretability surpassing even sparse autoencoder (SAE)-based approaches.

Constantin Venhoff, Ashkan Khakzar, Sonia Joseph et al.

Effective multimodal reasoning depends on the alignment of visual and linguistic representations, yet the mechanisms by which vision-language models (VLMs) achieve this alignment remain poorly understood. Following the LiMBeR framework, we deliberately maintain a frozen large language model (LLM) and a frozen vision transformer (ViT), connected solely by training a linear adapter during visual instruction tuning. By keeping the language model frozen, we ensure it maintains its original language representations without adaptation to visual data. Consequently, the linear adapter must map visual features directly into the LLM's existing representational space rather than allowing the language model to develop specialized visual understanding through fine-tuning. Our experimental design uniquely enables the use of pre-trained sparse autoencoders (SAEs) of the LLM as analytical probes. These SAEs remain perfectly aligned with the unchanged language model and serve as a snapshot of the learned language feature-representations. Through systematic analysis of SAE reconstruction error, sparsity patterns, and feature SAE descriptions, we reveal the layer-wise progression through which visual representations gradually align with language feature representations, converging in middle-to-later layers. This suggests a fundamental misalignment between ViT outputs and early LLM layers, raising important questions about whether current adapter-based architectures optimally facilitate cross-modal representation learning.

Yang Zhang, Er Jin, Yanfei Dong et al.

Recent advances in generative models have demonstrated remarkable capabilities in producing high-quality images, but their reliance on large-scale unlabeled data has raised significant safety and copyright concerns. Efforts to address these issues by erasing unwanted concepts have shown promise. However, many existing erasure methods involve excessive modifications that compromise the overall utility of the model. In this work, we address these issues by formulating a novel minimalist concept erasure objective based \emph{only} on the distributional distance of final generation outputs. Building on our formulation, we derive a tractable loss for differentiable optimization that leverages backpropagation through all generation steps in an end-to-end manner. We also conduct extensive analysis to show theoretical connections with other models and methods. To improve the robustness of the erasure, we incorporate neuron masking as an alternative to model fine-tuning. Empirical evaluations on state-of-the-art flow-matching models demonstrate that our method robustly erases concepts without degrading overall model performance, paving the way for safer and more responsible generative models.

Farnoush Rezaei Jafari, Oliver Eberle, Ashkan Khakzar et al.

Activation patching is a standard method in mechanistic interpretability for localizing the components of a model responsible for specific behaviors, but it is computationally expensive to apply at scale. Attribution patching offers a faster, gradient-based approximation, yet suffers from noise and reduced reliability in deep, highly non-linear networks. In this work, we introduce Relevance Patching (RelP), which replaces the local gradients in attribution patching with propagation coefficients derived from Layer-wise Relevance Propagation (LRP). LRP propagates the network's output backward through the layers, redistributing relevance to lower-level components according to local propagation rules that ensure properties such as relevance conservation or improved signal-to-noise ratio. Like attribution patching, RelP requires only two forward passes and one backward pass, maintaining computational efficiency while improving faithfulness. We validate RelP across a range of models and tasks, showing that it more accurately approximates activation patching than standard attribution patching, particularly when analyzing residual stream and MLP outputs in the Indirect Object Identification (IOI) task. For instance, for MLP outputs in GPT-2 Large, attribution patching achieves a Pearson correlation of 0.006, whereas RelP reaches 0.956, highlighting the improvement offered by RelP. Additionally, we compare the faithfulness of sparse feature circuits identified by RelP and Integrated Gradients (IG), showing that RelP achieves comparable faithfulness without the extra computational cost associated with IG.

Oishi Deb, Anjun Hu, Ashkan Khakzar et al.

We propose a training-free method, Articulate3D, to pose a 3D asset through language control. Despite advances in vision and language models, this task remains surprisingly challenging. To achieve this goal, we decompose the problem into two steps. We modify a powerful image-generator to create target images conditioned on the input image and a text instruction. We then align the mesh to the target images through a multi-view pose optimisation step. In detail, we introduce a self-attention rewiring mechanism (RSActrl) that decouples the source structure from pose within an image generative model, allowing it to maintain a consistent structure across varying poses. We observed that differentiable rendering is an unreliable signal for articulation optimisation; instead, we use keypoints to establish correspondences between input and target images. The effectiveness of Articulate3D is demonstrated across a diverse range of 3D objects and free-form text prompts, successfully manipulating poses while maintaining the original identity of the mesh. Quantitative evaluations and a comparative user study, in which our method was preferred over 85\% of the time, confirm its superiority over existing approaches. Project page:https://odeb1.github.io/articulate3d_page_deb/

Arshia Hemmat, Adam Davies, Tom A. Lamb et al.

Despite the importance of shape perception in human vision, early neural image classifiers relied less on shape information for object recognition than other (often spurious) features. While recent research suggests that current large Vision-Language Models (VLMs) exhibit more reliance on shape, we find them to still be seriously limited in this regard. To quantify such limitations, we introduce IllusionBench, a dataset that challenges current cutting-edge VLMs to decipher shape information when the shape is represented by an arrangement of visual elements in a scene. Our extensive evaluations reveal that, while these shapes are easily detectable by human annotators, current VLMs struggle to recognize them, indicating important avenues for future work in developing more robust visual perception systems. The full dataset and codebase are available at: \url{https://arshiahemmat.github.io/illusionbench/}

Razieh Rezaei, Masoud Jalili Sabet, Jindong Gu et al.

Visual prompting infuses visual information into the input image to adapt models toward specific predictions and tasks. Recently, manually crafted markers such as red circles are shown to guide the model to attend to a target region on the image. However, these markers only work on models trained with data containing those markers. Moreover, finding these prompts requires guesswork or prior knowledge of the domain on which the model is trained. This work circumvents manual design constraints by proposing to learn the visual prompts for guiding the attention of vision transformers. The learned visual prompt, added to any input image would redirect the attention of the pre-trained vision transformer to its spatial location on the image. Specifically, the prompt is learned in a self-supervised manner without requiring annotations and without fine-tuning the vision transformer. Our experiments demonstrate the effectiveness of the proposed optimization-based visual prompting strategy across various pre-trained vision encoders.

Razieh Rezaei, Alireza Dizaji, Ashkan Khakzar et al.

Neural networks are increasingly finding their way into the realm of graphs and modeling relationships between features. Concurrently graph neural network explanation approaches are being invented to uncover relationships between the nodes of the graphs. However, there is a disparity between the existing attribution methods, and it is unclear which attribution to trust. Therefore research has introduced evaluation experiments that assess them from different perspectives. In this work, we assess attribution methods from a perspective not previously explored in the graph domain: retraining. The core idea is to retrain the network on important (or not important) relationships as identified by the attributions and evaluate how networks can generalize based on these relationships. We reformulate the retraining framework to sidestep issues lurking in the previous formulation and propose guidelines for correct analysis. We run our analysis on four state-of-the-art GNN attribution methods and five synthetic and real-world graph classification datasets. The analysis reveals that attributions perform variably depending on the dataset and the network. Most importantly, we observe that the famous GNNExplainer performs similarly to an arbitrary designation of edge importance. The study concludes that the retraining evaluation cannot be used as a generalized benchmark and recommends it as a toolset to evaluate attributions on a specifically addressed network, dataset, and sparsity.

Paul Engstler, Matthias Keicher, David Schinz et al.

Do black-box neural network models learn clinically relevant features for fracture diagnosis? The answer not only establishes reliability quenches scientific curiosity but also leads to explainable and verbose findings that can assist the radiologists in the final and increase trust. This work identifies the concepts networks use for vertebral fracture diagnosis in CT images. This is achieved by associating concepts to neurons highly correlated with a specific diagnosis in the dataset. The concepts are either associated with neurons by radiologists pre-hoc or are visualized during a specific prediction and left for the user's interpretation. We evaluate which concepts lead to correct diagnosis and which concepts lead to false positives. The proposed frameworks and analysis pave the way for reliable and explainable vertebral fracture diagnosis.

Yang Zhang, Ashkan Khakzar, Yawei Li et al.

One principal approach for illuminating a black-box neural network is feature attribution, i.e. identifying the importance of input features for the network's prediction. The predictive information of features is recently proposed as a proxy for the measure of their importance. So far, the predictive information is only identified for latent features by placing an information bottleneck within the network. We propose a method to identify features with predictive information in the input domain. The method results in fine-grained identification of input features' information and is agnostic to network architecture. The core idea of our method is leveraging a bottleneck on the input that only lets input features associated with predictive latent features pass through. We compare our method with several feature attribution methods using mainstream feature attribution evaluation experiments. The code is publicly available.

Ashkan Khakzar, Sabrina Musatian, Jonas Buchberger et al.

Convolutional neural networks are showing promise in the automatic diagnosis of thoracic pathologies on chest x-rays. Their black-box nature has sparked many recent works to explain the prediction via input feature attribution methods (aka saliency methods). However, input feature attribution methods merely identify the importance of input regions for the prediction and lack semantic interpretation of model behavior. In this work, we first identify the semantics associated with internal units (feature maps) of the network. We proceed to investigate the following questions; Does a regression model that is only trained with COVID-19 severity scores implicitly learn visual patterns associated with thoracic pathologies? Does a network that is trained on weakly labeled data (e.g. healthy, unhealthy) implicitly learn pathologies? Moreover, we investigate the effect of pretraining and data imbalance on the interpretability of learned features. In addition to the analysis, we propose semantic attribution to semantically explain each prediction. We present our findings using publicly available chest pathologies (CheXpert, NIH ChestX-ray8) and COVID-19 datasets (BrixIA, and COVID-19 chest X-ray segmentation dataset). The Code is publicly available.

Ashkan Khakzar, Yang Zhang, Wejdene Mansour et al.

Neural networks have demonstrated remarkable performance in classification and regression tasks on chest X-rays. In order to establish trust in the clinical routine, the networks' prediction mechanism needs to be interpretable. One principal approach to interpretation is feature attribution. Feature attribution methods identify the importance of input features for the output prediction. Building on Information Bottleneck Attribution (IBA) method, for each prediction we identify the chest X-ray regions that have high mutual information with the network's output. Original IBA identifies input regions that have sufficient predictive information. We propose Inverse IBA to identify all informative regions. Thus all predictive cues for pathologies are highlighted on the X-rays, a desirable property for chest X-ray diagnosis. Moreover, we propose Regression IBA for explaining regression models. Using Regression IBA we observe that a model trained on cumulative severity score labels implicitly learns the severity of different X-ray regions. Finally, we propose Multi-layer IBA to generate higher resolution and more detailed attribution/saliency maps. We evaluate our methods using both human-centric (ground-truth-based) interpretability metrics, and human-independent feature importance metrics on NIH Chest X-ray8 and BrixIA datasets. The Code is publicly available.

Ashkan Khakzar, Soroosh Baselizadeh, Saurabh Khanduja et al.

Is critical input information encoded in specific sparse pathways within the neural network? In this work, we discuss the problem of identifying these critical pathways and subsequently leverage them for interpreting the network's response to an input. The pruning objective -- selecting the smallest group of neurons for which the response remains equivalent to the original network -- has been previously proposed for identifying critical pathways. We demonstrate that sparse pathways derived from pruning do not necessarily encode critical input information. To ensure sparse pathways include critical fragments of the encoded input information, we propose pathway selection via neurons' contribution to the response. We proceed to explain how critical pathways can reveal critical input features. We prove that pathways selected via neuron contribution are locally linear (in an L2-ball), a property that we use for proposing a feature attribution method: "pathway gradient". We validate our interpretation method using mainstream evaluation experiments. The validation of pathway gradient interpretation method further confirms that selected pathways using neuron contributions correspond to critical input features. The code is publicly available.

Seong Tae Kim, Leili Goli, Magdalini Paschali et al.

Chest computed tomography (CT) has played an essential diagnostic role in assessing patients with COVID-19 by showing disease-specific image features such as ground-glass opacity and consolidation. Image segmentation methods have proven to help quantify the disease burden and even help predict the outcome. The availability of longitudinal CT series may also result in an efficient and effective method to reliably assess the progression of COVID-19, monitor the healing process and the response to different therapeutic strategies. In this paper, we propose a new framework to identify infection at a voxel level (identification of healthy lung, consolidation, and ground-glass opacity) and visualize the progression of COVID-19 using sequential low-dose non-contrast CT scans. In particular, we devise a longitudinal segmentation network that utilizes the reference scan information to improve the performance of disease identification. Experimental results on a clinical longitudinal dataset collected in our institution show the effectiveness of the proposed method compared to the static deep neural networks for disease quantification.

Ashkan Khakzar, Soroosh Baselizadeh, Nassir Navab

Saliency methods interpret the prediction of a neural network by showing the importance of input elements for that prediction. A popular family of saliency methods utilize gradient information. In this work, we empirically show that two approaches for handling the gradient information, namely positive aggregation, and positive propagation, break these methods. Though these methods reflect visually salient information in the input, they do not explain the model prediction anymore as the generated saliency maps are insensitive to the predicted output and are insensitive to model parameter randomization. Specifically for methods that aggregate the gradients of a chosen layer such as GradCAM++ and FullGrad, exclusively aggregating positive gradients is detrimental. We further support this by proposing several variants of aggregation methods with positive handling of gradient information. For methods that backpropagate gradient information such as LRP, RectGrad, and Guided Backpropagation, we show the destructive effect of exclusively propagating positive gradient information.

Stefan Denner, Ashkan Khakzar, Moiz Sajid et al.

Segmentation of Multiple Sclerosis (MS) lesions in longitudinal brain MR scans is performed for monitoring the progression of MS lesions. We hypothesize that the spatio-temporal cues in longitudinal data can aid the segmentation algorithm. Therefore, we propose a multi-task learning approach by defining an auxiliary self-supervised task of deformable registration between two time-points to guide the neural network toward learning from spatio-temporal changes. We show the efficacy of our method on a clinical dataset comprised of 70 patients with one follow-up study for each patient. Our results show that spatio-temporal information in longitudinal data is a beneficial cue for improving segmentation. We improve the result of current state-of-the-art by 2.6% in terms of overall score (p<0.05). Code is publicly available.

Ashkan Khakzar, Soroosh Baselizadeh, Saurabh Khanduja et al.

Attributing the output of a neural network to the contribution of given input elements is a way of shedding light on the black-box nature of neural networks. Due to the complexity of current network architectures, current gradient-based attribution methods provide very noisy or coarse results. We propose to prune a neural network for a given single input to keep only neurons that highly contribute to the prediction. We show that by input-specific pruning, network gradients change from reflecting local (noisy) importance information to global importance. Our proposed method is efficient and generates fine-grained attribution maps. We further provide a theoretical justification of the pruning approach relating it to perturbations and validate it through a novel experimental setup. Our method is evaluated by multiple benchmarks: sanity checks, pixel perturbation, and Remove-and-Retrain (ROAR). These benchmarks evaluate the method from different perspectives and our method performs better than other methods across all evaluations.

Ashkan Khakzar, Shadi Albarqouni, Nassir Navab

Neural networks are proven to be remarkably successful for classification and diagnosis in medical applications. However, the ambiguity in the decision-making process and the interpretability of the learned features is a matter of concern. In this work, we propose a method for improving the feature interpretability of neural network classifiers. Initially, we propose a baseline convolutional neural network with state of the art performance in terms of accuracy and weakly supervised localization. Subsequently, the loss is modified to integrate robustness to adversarial examples into the training process. In this work, feature interpretability is quantified via evaluating the weakly supervised localization using the ground truth bounding boxes. Interpretability is also visually assessed using class activation maps and saliency maps. The method is applied to NIH ChestX-ray14, the largest publicly available chest x-rays dataset. We demonstrate that the adversarially robust optimization paradigm improves feature interpretability both quantitatively and visually.