Titus J. Brinker

Tirtha Chanda, Katja Hauser, Sarah Hobelsberger et al.

Although artificial intelligence (AI) systems have been shown to improve the accuracy of initial melanoma diagnosis, the lack of transparency in how these systems identify melanoma poses severe obstacles to user acceptance. Explainable artificial intelligence (XAI) methods can help to increase transparency, but most XAI methods are unable to produce precisely located domain-specific explanations, making the explanations difficult to interpret. Moreover, the impact of XAI methods on dermatologists has not yet been evaluated. Extending on two existing classifiers, we developed an XAI system that produces text and region based explanations that are easily interpretable by dermatologists alongside its differential diagnoses of melanomas and nevi. To evaluate this system, we conducted a three-part reader study to assess its impact on clinicians' diagnostic accuracy, confidence, and trust in the XAI-support. We showed that our XAI's explanations were highly aligned with clinicians' explanations and that both the clinicians' trust in the support system and their confidence in their diagnoses were significantly increased when using our XAI compared to using a conventional AI system. The clinicians' diagnostic accuracy was numerically, albeit not significantly, increased. This work demonstrates that clinicians are willing to adopt such an XAI system, motivating their future use in the clinic.

Achim Hekler, Roman C. Maron, Sarah Haggenmüller et al.

Background: Convolutional neural network (CNN)-based melanoma classifiers face several challenges that limit their usefulness in clinical practice. Objective: To investigate the impact of multiple real-world dermoscopic views of a single lesion of interest on a CNN-based melanoma classifier. Methods: This study evaluated 656 suspected melanoma lesions. Classifier performance was measured using area under the receiver operating characteristic curve (AUROC), expected calibration error (ECE) and maximum confidence change (MCC) for (I) a single-view scenario, (II) a multiview scenario using multiple artificially modified images per lesion and (III) a multiview scenario with multiple real-world images per lesion. Results: The multiview approach with real-world images significantly increased the AUROC from 0.905 (95% CI, 0.879-0.929) in the single-view approach to 0.930 (95% CI, 0.909-0.951). ECE and MCC also improved significantly from 0.131 (95% CI, 0.105-0.159) to 0.072 (95% CI: 0.052-0.093) and from 0.149 (95% CI, 0.125-0.171) to 0.115 (95% CI: 0.099-0.131), respectively. Comparing multiview real-world to artificially modified images showed comparable diagnostic accuracy and uncertainty estimation, but significantly worse robustness for the latter. Conclusion: Using multiple real-world images is an inexpensive method to positively impact the performance of a CNN-based melanoma classifier.

Katharina Fogelberg, Sireesha Chamarthi, Roman C. Maron et al.

The limited ability of Convolutional Neural Networks to generalize to images from previously unseen domains is a major limitation, in particular, for safety-critical clinical tasks such as dermoscopic skin cancer classification. In order to translate CNN-based applications into the clinic, it is essential that they are able to adapt to domain shifts. Such new conditions can arise through the use of different image acquisition systems or varying lighting conditions. In dermoscopy, shifts can also occur as a change in patient age or occurence of rare lesion localizations (e.g. palms). These are not prominently represented in most training datasets and can therefore lead to a decrease in performance. In order to verify the generalizability of classification models in real world clinical settings it is crucial to have access to data which mimics such domain shifts. To our knowledge no dermoscopic image dataset exists where such domain shifts are properly described and quantified. We therefore grouped publicly available images from ISIC archive based on their metadata (e.g. acquisition location, lesion localization, patient age) to generate meaningful domains. To verify that these domains are in fact distinct, we used multiple quantification measures to estimate the presence and intensity of domain shifts. Additionally, we analyzed the performance on these domains with and without an unsupervised domain adaptation technique. We observed that in most of our grouped domains, domain shifts in fact exist. Based on our results, we believe these datasets to be helpful for testing the generalization capabilities of dermoscopic skin cancer classifiers.

Martin J. Hetz, Tabea-Clara Bucher, Titus J. Brinker

The variation in histologic staining between different medical centers is one of the most profound challenges in the field of computer-aided diagnosis. The appearance disparity of pathological whole slide images causes algorithms to become less reliable, which in turn impedes the wide-spread applicability of downstream tasks like cancer diagnosis. Furthermore, different stainings lead to biases in the training which in case of domain shifts negatively affect the test performance. Therefore, in this paper we propose MultiStain-CycleGAN, a multi-domain approach to stain normalization based on CycleGAN. Our modifications to CycleGAN allow us to normalize images of different origins without retraining or using different models. We perform an extensive evaluation of our method using various metrics and compare it to commonly used methods that are multi-domain capable. First, we evaluate how well our method fools a domain classifier that tries to assign a medical center to an image. Then, we test our normalization on the tumor classification performance of a downstream classifier. Furthermore, we evaluate the image quality of the normalized images using the Structural similarity index and the ability to reduce the domain shift using the Fréchet inception distance. We show that our method proves to be multi-domain capable, provides the highest image quality among the compared methods, and can most reliably fool the domain classifier while keeping the tumor classifier performance high. By reducing the domain influence, biases in the data can be removed on the one hand and the origin of the whole slide image can be disguised on the other, thus enhancing patient data privacy.

Hendrik A. Mehrtens, Alexander Kurz, Tabea-Clara Bucher et al.

In the past years, deep learning has seen an increase in usage in the domain of histopathological applications. However, while these approaches have shown great potential, in high-risk environments deep learning models need to be able to judge their uncertainty and be able to reject inputs when there is a significant chance of misclassification. In this work, we conduct a rigorous evaluation of the most commonly used uncertainty and robustness methods for the classification of Whole Slide Images, with a focus on the task of selective classification, where the model should reject the classification in situations in which it is uncertain. We conduct our experiments on tile-level under the aspects of domain shift and label noise, as well as on slide-level. In our experiments, we compare Deep Ensembles, Monte-Carlo Dropout, Stochastic Variational Inference, Test-Time Data Augmentation as well as ensembles of the latter approaches. We observe that ensembles of methods generally lead to better uncertainty estimates as well as an increased robustness towards domain shifts and label noise, while contrary to results from classical computer vision benchmarks no systematic gain of the other methods can be shown. Across methods, a rejection of the most uncertain samples reliably leads to a significant increase in classification accuracy on both in-distribution as well as out-of-distribution data. Furthermore, we conduct experiments comparing these methods under varying conditions of label noise. Lastly, we publish our code framework to facilitate further research on uncertainty estimation on histopathological data.

Jan Clusmann, Dyke Ferber, Isabella C. Wiest et al.

Vision-language artificial intelligence models (VLMs) possess medical knowledge and can be employed in healthcare in numerous ways, including as image interpreters, virtual scribes, and general decision support systems. However, here, we demonstrate that current VLMs applied to medical tasks exhibit a fundamental security flaw: they can be attacked by prompt injection attacks, which can be used to output harmful information just by interacting with the VLM, without any access to its parameters. We performed a quantitative study to evaluate the vulnerabilities to these attacks in four state of the art VLMs which have been proposed to be of utility in healthcare: Claude 3 Opus, Claude 3.5 Sonnet, Reka Core, and GPT-4o. Using a set of N=297 attacks, we show that all of these models are susceptible. Specifically, we show that embedding sub-visual prompts in medical imaging data can cause the model to provide harmful output, and that these prompts are non-obvious to human observers. Thus, our study demonstrates a key vulnerability in medical VLMs which should be mitigated before widespread clinical adoption.

Christoph Wies, Lucas Schneider, Sarah Haggenmueller et al.

Pathologists routinely use immunohistochemical (IHC)-stained tissue slides against MelanA in addition to hematoxylin and eosin (H&E)-stained slides to improve their accuracy in diagnosing melanomas. The use of diagnostic Deep Learning (DL)-based support systems for automated examination of tissue morphology and cellular composition has been well studied in standard H&E-stained tissue slides. In contrast, there are few studies that analyze IHC slides using DL. Therefore, we investigated the separate and joint performance of ResNets trained on MelanA and corresponding H&E-stained slides. The MelanA classifier achieved an area under receiver operating characteristics curve (AUROC) of 0.82 and 0.74 on out of distribution (OOD)-datasets, similar to the H&E-based benchmark classification of 0.81 and 0.75, respectively. A combined classifier using MelanA and H&E achieved AUROCs of 0.85 and 0.81 on the OOD datasets. DL MelanA-based assistance systems show the same performance as the benchmark H&E classification and may be improved by multi stain classification to assist pathologists in their clinical routine.

Tirtha Chanda, Sarah Haggenmueller, Tabea-Clara Bucher et al.

Artificial intelligence (AI) systems have substantially improved dermatologists' diagnostic accuracy for melanoma, with explainable AI (XAI) systems further enhancing clinicians' confidence and trust in AI-driven decisions. Despite these advancements, there remains a critical need for objective evaluation of how dermatologists engage with both AI and XAI tools. In this study, 76 dermatologists participated in a reader study, diagnosing 16 dermoscopic images of melanomas and nevi using an XAI system that provides detailed, domain-specific explanations. Eye-tracking technology was employed to assess their interactions. Diagnostic performance was compared with that of a standard AI system lacking explanatory features. Our findings reveal that XAI systems improved balanced diagnostic accuracy by 2.8 percentage points relative to standard AI. Moreover, diagnostic disagreements with AI/XAI systems and complex lesions were associated with elevated cognitive load, as evidenced by increased ocular fixations. These insights have significant implications for clinical practice, the design of AI tools for visual tasks, and the broader development of XAI in medical diagnostics.

Sireesha Chamarthi, Katharina Fogelberg, Roman C. Maron et al.

The potential of deep neural networks in skin lesion classification has already been demonstrated to be on-par if not superior to the dermatologists diagnosis. However, the performance of these models usually deteriorates when the test data differs significantly from the training data (i.e. domain shift). This concerning limitation for models intended to be used in real-world skin lesion classification tasks poses a risk to patients. For example, different image acquisition systems or previously unseen anatomical sites on the patient can suffice to cause such domain shifts. Mitigating the negative effect of such shifts is therefore crucial, but developing effective methods to address domain shift has proven to be challenging. In this study, we carry out an in-depth analysis of eight different unsupervised domain adaptation methods to analyze their effectiveness in improving generalization for dermoscopic datasets. To ensure robustness of our findings, we test each method on a total of ten distinct datasets, thereby covering a variety of possible domain shifts. In addition, we investigated which factors in the domain shifted datasets have an impact on the effectiveness of domain adaptation methods. Our findings show that all of the eight domain adaptation methods result in improved AUPRC for the majority of analyzed datasets. Altogether, these results indicate that unsupervised domain adaptations generally lead to performance improvements for the binary melanoma-nevus classification task regardless of the nature of the domain shift. However, small or heavily imbalanced datasets lead to a reduced conformity of the results due to the influence of these factors on the methods performance.

Lukas Heinlein, Roman C. Maron, Achim Hekler et al.

Early detection of melanoma, a potentially lethal type of skin cancer with high prevalence worldwide, improves patient prognosis. In retrospective studies, artificial intelligence (AI) has proven to be helpful for enhancing melanoma detection. However, there are few prospective studies confirming these promising results. Existing studies are limited by low sample sizes, too homogenous datasets, or lack of inclusion of rare melanoma subtypes, preventing a fair and thorough evaluation of AI and its generalizability, a crucial aspect for its application in the clinical setting. Therefore, we assessed 'All Data are Ext' (ADAE), an established open-source ensemble algorithm for detecting melanomas, by comparing its diagnostic accuracy to that of dermatologists on a prospectively collected, external, heterogeneous test set comprising eight distinct hospitals, four different camera setups, rare melanoma subtypes, and special anatomical sites. We advanced the algorithm with real test-time augmentation (R-TTA, i.e. providing real photographs of lesions taken from multiple angles and averaging the predictions), and evaluated its generalization capabilities. Overall, the AI showed higher balanced accuracy than dermatologists (0.798, 95% confidence interval (CI) 0.779-0.814 vs. 0.781, 95% CI 0.760-0.802; p<0.001), obtaining a higher sensitivity (0.921, 95% CI 0.900- 0.942 vs. 0.734, 95% CI 0.701-0.770; p<0.001) at the cost of a lower specificity (0.673, 95% CI 0.641-0.702 vs. 0.828, 95% CI 0.804-0.852; p<0.001). As the algorithm exhibited a significant performance advantage on our heterogeneous dataset exclusively comprising melanoma-suspicious lesions, AI may offer the potential to support dermatologists particularly in diagnosing challenging cases.

Hendrik Mehrtens, Tabea Bucher, Titus J. Brinker

Reliable uncertainty estimation is one of the major challenges for medical classification tasks. While many approaches have been proposed, recently the statistical framework of conformal predictions has gained a lot of attention, due to its ability to provide provable calibration guarantees. Nonetheless, the application of conformal predictions in safety-critical areas such as medicine comes with pitfalls, limitations and assumptions that practitioners need to be aware of. We demonstrate through examples from dermatology and histopathology that conformal predictions are unreliable under distributional shifts in input and label variables. Additionally, conformal predictions should not be used for selecting predictions to improve accuracy and are not reliable for subsets of the data, such as individual classes or patient attributes. Moreover, in classification settings with a small number of classes, which are common in medical image classification tasks, conformal predictions have limited practical value.

Martin J. Hetz, Carina Nogueira Garcia, Sarah Haggenmüller et al.

Clinical dermatology necessitates precision and innovation for efficient diagnosis and treatment of various skin conditions. This paper introduces the development of a cutting-edge hyperspectral dermatoscope (the Hyperscope) tailored for human skin analysis. We detail the requirements to such a device and the design considerations, from optical configurations to sensor selection, necessary to capture a wide spectral range with high fidelity. Preliminary results from 15 individuals and 160 recorded skin images demonstrate the potential of the Hyperscope in identifying and characterizing various skin conditions, offering a promising avenue for non-invasive skin evaluation and a platform for future research in dermatology-related hyperspectral imaging.

Gesa Mittmann, Sara Laiouar-Pedari, Hendrik A. Mehrtens et al.

The aggressiveness of prostate cancer, the most common cancer in men worldwide, is primarily assessed based on histopathological data using the Gleason scoring system. While artificial intelligence (AI) has shown promise in accurately predicting Gleason scores, these predictions often lack inherent explainability, potentially leading to distrust in human-machine interactions. To address this issue, we introduce a novel dataset of 1,015 tissue microarray core images, annotated by an international group of 54 pathologists. The annotations provide detailed localized pattern descriptions for Gleason grading in line with international guidelines. Utilizing this dataset, we develop an inherently explainable AI system based on a U-Net architecture that provides predictions leveraging pathologists' terminology. This approach circumvents post-hoc explainability methods while maintaining or exceeding the performance of methods trained directly for Gleason pattern segmentation (Dice score: 0.713 $\pm$ 0.003 trained on explanations vs. 0.691 $\pm$ 0.010 trained on Gleason patterns). By employing soft labels during training, we capture the intrinsic uncertainty in the data, yielding strong results in Gleason pattern segmentation even in the context of high interobserver variability. With the release of this dataset, we aim to encourage further research into segmentation in medical tasks with high levels of subjectivity and to advance the understanding of pathologists' reasoning processes.

Martin J. Hetz, Nicolas Carl, Sarah Haggenmüller et al.

Large Language Models (LLMs) are revolutionizing medical Question-Answering (medQA) through extensive use of medical literature. However, their performance is often hampered by outdated training data and a lack of explainability, which limits clinical applicability. This study aimed to create and assess UroBot, a urology-specialized chatbot, by comparing it with state-of-the-art models and the performance of urologists on urological board questions, ensuring full clinician-verifiability. UroBot was developed using OpenAI's GPT-3.5, GPT-4, and GPT-4o models, employing retrieval-augmented generation (RAG) and the latest 2023 guidelines from the European Association of Urology (EAU). The evaluation included ten runs of 200 European Board of Urology (EBU) In-Service Assessment (ISA) questions, with performance assessed by the mean Rate of Correct Answers (RoCA). UroBot-4o achieved an average RoCA of 88.4%, surpassing GPT-4o by 10.8%, with a score of 77.6%. It was also clinician-verifiable and exhibited the highest run agreement as indicated by Fleiss' Kappa (k = 0.979). By comparison, the average performance of urologists on board questions, as reported in the literature, is 68.7%. UroBot's clinician-verifiable nature and superior accuracy compared to both existing models and urologists on board questions highlight its potential for clinical integration. The study also provides the necessary code and instructions for further development of UroBot.

Alexander Kurz, Hendrik A. Mehrtens, Tabea-Clara Bucher et al.

Deep Neural Networks have shown promising classification performance when predicting certain biomarkers from Whole Slide Images in digital pathology. However, the calibration of the networks' output probabilities is often not evaluated. Communicating uncertainty by providing reliable confidence scores is of high relevance in the medical context. In this work, we compare three neural network architectures that combine feature representations on patch-level to a slide-level prediction with respect to their classification performance and evaluate their calibration. As slide-level classification task, we choose the prediction of Microsatellite Instability from Colorectal Cancer tissue sections. We observe that Transformers lead to good results in terms of classification performance and calibration. When evaluating the classification performance on a separate dataset, we observe that Transformers generalize best. The investigation of reliability diagrams provides additional insights to the Expected Calibration Error metric and we observe that especially Transformers push the output probabilities to extreme values, which results in overconfident predictions.