Michael Götz

Pancreatic ductal adenocarcinoma (PDAC) subtyping typically relies on immunohistochemistry (IHC) staining for critical markers like HNF1A and KRT81, a labor‐intensive manual staining process that introduces variability. Virtual staining methods offer promising alternatives by generating synthetic IHC images from routine hematoxylin and eosin (H&E) slides. However, most current approaches evaluate success by image quality measures rather than assessing diagnostically relevant features. Here, we introduce a novel cycleGAN framework utilizing a contrastive‐inspired approach trained on semipaired datasets derived from consecutive tissue sections. Our method significantly enhances PDAC subtyping accuracy based on synthetic IHC images generated from standard H&E inputs, improving the classification F1‐score from 0.66 to 0.77 for KRT81 and from 0.61 to 0.73 for HNF1A, compared with classification directly on H&E images. This approach also substantially outperforms baseline CycleGAN models. These results underscore the clinical potential of contrastive virtual staining to streamline PDAC diagnostics and improve their robustness. © 2025 The Author(s). The Journal of Pathology published by John Wiley & Sons Ltd on behalf of The Pathological Society of Great Britain and Ireland.

Background: Postoperative pulmonary complications (POPC) are common after general anaesthesia and are a major cause of increased morbidity and mortality in surgical patients. However, prevention and treatment methods for POPC that are considered effective tie up human and technical resources. Therefore, the planned research project aims to create a prediction model that enables the reliable identification of high-risk patients immediately after surgery based on a tailored machine learning algorithm. Methods: This clinical cohort study will follow the TRIPOD statement for multivariable prediction model development. Development of the prognostic model will require 512 patients undergoing elective surgery under general anaesthesia. Besides the collection of perioperative routine data, standardised lung sonography will be performed postoperatively in the recovery room on each patient. During the postoperative course, patients will be examined in a structured manner on postoperative days 1,3 and 7 to detect POPC. The endpoints determined in this way, together with the clinical and imaging data collected, are then used to train a machine learning model based on neural networks and ensemble methods to predict POPC in the early postoperative phase. Discussion: In the perioperative setting, detecting POPC before they become clinically manifest is desirable. This would ensure optimal patient care and resource allocation and help initiate adequate patient treatment after being transferred from the recovery room to the ward. A reliable prediction algorithm based on machine learning holds great potential to improve postoperative outcomes. Trial registration: ClinicalTrials.gov ID: NCT05789953 (29th of March 2023)

Explainable artificial intelligence (xAI) is becoming increasingly important as the need for understanding the model’s reasoning grows when applying them in high-risk areas. This is especially crucial in the field of medicine, where decision support systems are utilised to make diagnoses or to determine appropriate therapies. Here it is essential to provide intuitive and comprehensive explanations to evaluate the system’s correctness. To meet this need, we have developed Proto-Caps, an intrinsically explainable model for image classification. It explains its decisions by providing visual prototypes that resemble specific appearance features. These characteristics are predefined by humans, which on the one hand makes them understandable and on the other hand leads to the model basing its decision on the same features as the human expert. On two public datasets, this method shows better performance compared to existing explainable approaches, despite the additive explainability modality through the visual prototypes. In addition to the performance evaluations, we conducted an analysis of truthfulness by examining the joint information between the target prediction and its explanation output. This was done in order to ensure that the explanation actually reasons the target classification. Through extensive hyperparameter studies, we also found optimal model settings, providing a starting point for further research. Our work emphasises the prospects of combining xAI approaches for greater explainability and demonstrates that incorporating explainability does not necessarily lead to a loss of performance.

Due to the sensitive nature of medicine, it is particularly important and highly demanded that AI methods are explainable. This need has been recognised and there is great research interest in xAI solutions with medical applications. However, there is a lack of user-centred evaluation regarding the actual impact of the explanations. We evaluate attribute- and prototype-based explanations with the Proto-Caps model. This xAI model reasons the target classification with human-defined visual features of the target object in the form of scores and attribute-specific prototypes. The model thus provides a multimodal explanation that is intuitively understandable to humans thanks to predefined attributes. A user study involving six radiologists shows that the explanations are subjectivly perceived as helpful, as they reflect their decision-making process. The results of the model are considered a second opinion that radiologists can discuss using the model's explanations. However, it was shown that the inclusion and increased magnitude of model explanations objectively can increase confidence in the model's predictions when the model is incorrect. We can conclude that attribute scores and visual prototypes enhance confidence in the model. However, additional development and repeated user studies are needed to tailor the explanation to the respective use case.