Conformal prediction is a model-agnostic approach to generating prediction sets that cover the true class with a high probability. Although its prediction set size is expected to capture aleatoric uncertainty, there is a lack of evidence regarding its effectiveness. The literature presents that prediction set size can upper-bound aleatoric uncertainty or that prediction sets are larger for difficult instances and smaller for easy ones, but a validation of this attribute of conformal predictors is missing. This work investigates how effectively conformal predictors quantify aleatoric uncertainty, specifically the inherent ambiguity in datasets caused by overlapping classes. We perform this by measuring the correlation between prediction set sizes and the number of distinct labels assigned by human annotators per instance. We further assess the similarity between prediction sets and human-provided annotations. We use three conformal prediction approaches to generate prediction sets for eight deep learning models trained on four datasets. The datasets contain annotations from multiple human annotators (ranging from five to fifty participants) per instance, enabling the identification of class overlap. We show that the vast majority of the conformal prediction outputs show a very weak to weak correlation with human annotations, with only a few showing moderate correlation. These findings underscore the necessity of critically reassessing the prediction sets generated using conformal predictors. While they can provide a higher coverage of the true classes, their capability in capturing aleatoric uncertainty and generating sets that align with human annotations remains limited.

Datasets with large heterogeneity in data density are prone to result in visualizations with unutilized whitespace and overplotting. Sequential data, such as in genomics or electronic health records, often has such density distributions, containing long empty stretches between data-rich regions. For these data types, the exact distance is not necessarily the central feature, and plotting along a linear axis causes occlusion of the actual datapoints. In this work, we evaluate content-driven axis compression for minimizing whitespace and reducing occlusion while maintaining distance information. The axis is compressed based on the density of the data, and it requires minimal a priori knowledge of the data distribution. We present a framework for reasoning about the benefits and drawbacks of using axis compression. The framework includes reasoning about the separability of items and clusters, suggests careful consideration of how the position visual channel is utilized, and highlights the influence of the usage context for which the visualization is designed. We conducted a study to evaluate the strengths and weaknesses of axis compression and found that it leads to improved item visibility while maintaining high accuracy, also for tasks targeted at understanding the data distribution.

Multi-label learning often requires identifying all relevant labels for training instances, but collecting full label annotations is costly and labor-intensive. In many datasets, only a single positive label is annotated per training instance, despite the presence of multiple relevant labels. This setting, known as single-positive multi-label learning (SPMLL), presents a significant challenge due to its extreme form of partial supervision. We propose a model-agnostic approach to SPMLL that draws on conformal prediction to produce calibrated set-valued outputs, enabling reliable multi-label predictions at test time. Our method bridges the supervision gap between single-label training and multi-label evaluation without relying on label distribution assumptions. We evaluate our approach on 12 benchmark datasets, demonstrating consistent improvements over existing baselines and practical applicability.

Abstract Comparing gene organization across genomic sequences reveals insights into evolutionary and functional diversity among different organisms and varieties. Performing this task across many sequences, such as from a pangenome, is challenging because of the scale, the density of information, and the inherent variation. Often, analyses are centered on a genomic region of interest—a locus that might be associated with a trait or contain genes within the same family or biological pathway. Within these regions, researchers examine the conservation of gene order and orientation across organisms and assess sequence similarity, along with other gene content features such as gene size, to find biological variations or potential errors in the data. Automated methods in comparative genomics struggle to identify meaningful patterns due to varying and often unknown features of interest, leaving manual, time‐intensive, and scalability‐challenged visualization as the primary alternative. To address these challenges, we present a multiscale design for studying gene organization within pangenomes, developed in close collaboration with domain experts. Our tool, Multipla, enables users to explore organization at multiple levels of detail in a decluttered manner through layout abstractions, semantic zooming, and layouts with flexible distance definitions and feature selections, combining the advantages of manual and automated methods used in practice. We evaluate the design of Multipla through two pangenomic use cases and conclude with lessons learned from designing multiscale views for pangenomic locus analysis.

Genomics is at the core of precision medicine, and there are high expectations on genomics-enabled improvement of patient outcomes in the years to come. Around the world, initiatives to increase the use of DNA sequencing in clinical routine are being deployed, such as the use of broad panels in the standard care for oncology patients. Such a development comes at the cost of increased demands on throughput in genomic data analysis. In this paper, we use the task of copy number variant (CNV) analysis as a context for exploring visualization concepts for clinical genomics. CNV calls are generated algorithmically, but time-consuming manual intervention is needed to separate relevant findings from irrelevant ones in the resulting large call candidate lists. We present a visualization environment, named Copycat, to support this review task in a clinical scenario. Key components are a scatter-glyph plot replacing the traditional list visualization, and a glyph representation designed for at-a-glance relevance assessments. Moreover, we present results from a formative evaluation of the prototype by domain specialists, from which we elicit insights to guide both prototype improvements and visualization for clinical genomics in general.

Abstract Visual representations based on circular shapes are frequently used in visualization applications. One example are circos plots within bioinformatics, which bend graphs into a wheel of information with connective lines running through the center like spokes. The results are aesthetically appealing and impressive visualizations that fit long data sequences into a small quadratic space. However, the authors' experiences are that when asked, a visualization researcher would generally advise against making visualizations with radial layouts. Upon reviewing the literature we found that there is evidence that circular layouts are preferable in some cases, but we found no clear evidence for what layout is preferable for matrices and connective data in particular, which both are common data types in circos plots. In this work, we thus performed a user study to compare circular and linear layouts. The tasks are inspired by genomics data, but our results generalize to many other application areas, involving comparison and connective data. To build the prototype we utilized Gosling, a grammar for visualizing genomics data. We contribute empirical evidence on the suitedness of linear versus circular layouts, adding to the specific and general knowledge concerning perception of circular graphs. In addition, we contribute a case study evaluation of the grammar Gosling as a rapid prototyping language, confirming its utility and providing guidance on suitable areas for future development.

Designing useful human-AI interaction for clinical workflows remains challenging despite the impressive performance of recent AI models. One specific difficulty is a lack of successful examples demonstrating how to achieve safe and efficient workflows while mitigating AI imperfections. In this paper, we present an interactive AI-powered visual search tool that supports pathologists in cancer assessments. Our evaluation with six pathologists demonstrates that it can 1) reduce time needed with maintained quality, 2) build user trust progressively, and 3) learn and improve from use. We describe our iterative design process, model development, and key features. Through interviews, design choices are related to the overall user experience. Implications for future human-AI interaction design are discussed with respect to trust, explanations, learning from use, and collaboration strategies.

The high capacity of neural networks allows fitting models to data with high precision, but makes generalization to unseen data a challenge. If a domain shift exists, i.e. differences in image statistics between training and test data, care needs to be taken to ensure reliable deployment in real-world scenarios. In digital pathology, domain shift can be manifested in differences between whole-slide images, introduced by for example differences in acquisition pipeline - between medical centers or over time. In order to harness the great potential presented by deep learning in histopathology, and ensure consistent model behavior, we need a deeper understanding of domain shift and its consequences, such that a model's predictions on new data can be trusted. This work focuses on the internal representation learned by trained convolutional neural networks, and shows how this can be used to formulate a novel measure - the representation shift - for quantifying the magnitude of model-specific domain shift. We perform a study on domain shift in tumor classification of hematoxylin and eosin stained images, by considering different datasets, models, and techniques for preparing data in order to reduce the domain shift. The results show how the proposed measure has a high correlation with drop in performance when testing a model across a large number of different types of domain shifts, and how it improves on existing techniques for measuring data shift and uncertainty. The proposed 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. We see techniques for measuring, understanding and overcoming the domain shift as a crucial step towards reliable use of deep learning in the future clinical pathology applications.

We present a visualization application that enables effective interactive visual analysis of large-scale 3D histopathology, that is, high-resolution 3D microscopy data of human tissue. Clinical work flows and research based on pathology have, until now, largely been dominated by 2D imaging. As we will show in the paper, studying volumetric histology data will open up novel and useful opportunities for both research and clinical practice. Our starting point is the current lack of appropriate visualization tools in histopathology, which has been a limiting factor in the uptake of digital pathology. Visualization of 3D histology data does pose difficult challenges in several aspects. The full-color datasets are dense and large in scale, on the order of 100,000 × 100,000× 100 voxels. This entails serious demands on both rendering performance and user experience design. Despite this, our developed application supports interactive study of 3D histology datasets at native resolution. Our application is based on tailoring and tuning of existing methods, system integration work, as well as a careful study of domain specific demands emanating from a close participatory design process with domain experts as team members. Results from a user evaluation employing the tool demonstrate a strong agreement among the 14 participating pathologists that 3D histopathology will be a valuable and enabling tool for their work.

In this paper, we tackle the challenge of effective Transfer Function (TF) design for Direct Volume Rendering (DVR) of full-color datasets. We propose a novel TF design toolbox based on color similarity which is used to adjust opacity as well as replacing colors. We show that both CIE L*u*v* chromaticity and the chroma component of YCbCr are equally suited as underlying color space for the TF widgets. In order to maximize the area utilized in the TF editor, we renormalize the color space based on the histogram of the dataset. Thereby, colors representing a higher share of the dataset are depicted more prominently, thus providing a higher sensitivity for fine-tuning TF widgets. The applicability of our TF design toolbox is demonstrated by volume ray casting challenging full-color volume data including the visible male cryosection dataset and examples from 3D histology.

Medical imaging plays a central role in a vast range of healthcare practices. The usefulness of 3D visualizations has been demonstrated for many types of treatment planning. Nevertheless, full access to 3D renderings outside of the radiology department is still scarce even for many image-centric specialties. Our work stems from the hypothesis that this under-utilization is partly due to existing visualization systems not taking the prerequisites of this application domain fully into account. We have developed a medical visualization table intended to better fit the clinical reality. The overall design goals were two-fold: similarity to a real physical situation and a very low learning threshold. This paper describes the development of the visualization table with focus on key design decisions. The developed features include two novel interaction components for touch tables. A user study including five orthopedic surgeons demonstrates that the system is appropriate and useful for this application domain.

In many applications of Direct Volume Rendering (DVR) the importance of a certain material or feature is highly dependent on its relative spatial location. For instance, in the medical diagnostic procedure, the patient's symptoms often lead to specification of features, tissues and organs of particular interest. One such example is pockets of gas which, if found inside the body at abnormal locations, are a crucial part of a diagnostic visualization. This paper presents an approach that enhances DVR transfer function design with spatial localization based on user specified material dependencies. Semantic expressions are used to define conditions based on relations between different materials, such as only render iodine uptake when close to liver. The underlying methods rely on estimations of material distributions which are acquired by weighing local neighborhoods of the data against approximations of material likelihood functions. This information is encoded and used to influence rendering according to the user's specifications. The result is improved focus on important features by allowing the user to suppress spatially less-important data. In line with requirements from actual clinical DVR practice, the methods do not require explicit material segmentation that would be impossible or prohibitively time-consuming to achieve in most real cases. The scheme scales well to higher dimensions which accounts for multi-dimensional transfer functions and multivariate data. Dual-Energy Computed Tomography, an important new modality in radiology, is used to demonstrate this scalability. In several examples we show significantly improved focus on clinically important aspects in the rendered images.

Direct Volume Rendering has proved to be an effective visualization method for medical data sets and has reached wide-spread clinical use. The diagnostic exploration, in essence, corresponds to a tissue classification task, which is often complex and time-consuming. Moreover, a major problem is the lack of information on the uncertainty of the classification, which can have dramatic consequences for the diagnosis. In this paper this problem is addressed by proposing animation methods to convey uncertainty in the rendering. The foundation is a probabilistic Transfer Function model which allows for direct user interaction with the classification. The rendering is animated by sampling the probability domain over time, which results in varying appearance for uncertain regions. A particularly promising application of this technique is a "sensitivity lens" applied to focus regions in the data set. The methods have been evaluated by radiologists in a study simulating the clinical task of stenosis assessment, in which the animation technique is shown to outperform traditional rendering in terms of assessment accuracy.

We present a direct interblock interpolation technique that enables direct volume rendering of blocked, multiresolution volumes. The proposed method smoothly interpolates between blocks of arbitrary block-wise level-of-detail (LOD) without sample replication or padding. This permits extreme changes in resolution across block boundaries and removes the interblock dependency for the LOD creation process. In addition the full data reduction from the LOD selection can be maintained throughout the rendering pipeline. Our rendering pipeline employs a flat block subdivision followed by a transfer function based adaptive LOD scheme. We demonstrate the effectiveness of our method by rendering volumes of the order of gigabytes using consumer graphics cards on desktop PC systems.

The high complexity of Transfer Function (TF) design is a major obstacle to widespread routine use of Direct Volume Rendering, particularly in the case of medical imaging. Both manual and automatic TF design schemes would benefit greatly from a fast and simple method for detection of tissue value ranges. To this end, we introduce the a-histogram, an enhancement that amplifies ranges corresponding to spatially coherent materials. The properties of the a-histogram have been explored for synthetic data sets and then successfully used to detect vessels in 20 Magnetic Resonance angiographies, proving the potential of this approach as a fast and simple technique for histogram enhancement in general and for TF construction in particular.

This paper presents a procedure for virtual autopsies based on interactive 3D visualizations of large scale, high resolution data from CT-scans of human cadavers. The procedure is described using examples from forensic medicine and the added value and future potential of virtual autopsies is shown from a medical and forensic perspective. Based on the technical demands of the procedure state-of-the-art volume rendering techniques are applied and refined to enable real-time, full body virtual autopsies involving gigabyte sized data on standard GPUs. The techniques applied include transfer function based data reduction using level-of-detail selection and multi-resolution rendering techniques. The paper also describes a data management component for large, out-of-core data sets and an extension to the GPU-based raycaster for efficient dual TF rendering. Detailed benchmarks of the pipeline are presented using data sets from forensic cases.

Direct Volume Rendering (DVR) is of increasing diagnostic value in the analysis of data sets captured using the latest medical imaging modalities. The deployment of DVR in everyday clinical work, however, has so far been limited. One contributing factor is that current Transfer Function (TF) models can encode only a small fraction of the user's domain knowledge. In this paper, we use histograms of local neighborhoods to capture tissue characteristics. This allows domain knowledge on spatial relations in the data set to be integrated into the TF. As a first example, we introduce Partial Range Histograms in an automatic tissue detection scheme and present its effectiveness in a clinical evaluation. We then use local histogram analysis to perform a classification where the tissue-type certainty is treated as a second TF dimension. The result is an enhanced rendering where tissues with overlapping intensity ranges can be discerned without requiring the user to explicitly define a complex, multidimensional TF.

Direct Volume Rendering (DVR) is known to be of diagnostic value in the analysis of medical data sets. However, its deployment in everyday clinical use has so far been limited. Two major challenges are that the current methods for Transfer Function (TF) construction are too complex and that the tissue separation abilities of the TF need to be extended. In this paper we propose the use of histogram analysis in local neighborhoods to address both these conflicting problems. To reduce TF construction difficulty, we introduce Partial Range Histograms in an automatic tissue detection scheme, which in connection with Adaptive Trapezoids enable efficient TF design. To separate tissues with overlapping intensity ranges, we propose a fuzzy classification based on local histograms as a second TF dimension. This increases the power of the TF, while retaining intuitive presentation and interaction.