Marc Aubreville

An initial application of Confocal laser microendoscopy (CLE) in basal cell carcinoma (BCC) was promising, yet data in this field are still limited. Tumor vasculature, which can be clearly visualized by CLE, differs significantly from that of normal tissue, but has not yet been systematically analyzed as a target of CLE in BCC. By systematically assessing vascular features, we aim to define diagnostic criteria that support a more standardized and objective use of CLE in BCC. CLE imaging was performed in nine patients with histologically confirmed facial BCC. Forty artifact-free vascular sequences (20 BCC, 20 healthy skin) were analyzed using five predefined criteria independently rated by three blinded reviewers. Diagnostic accuracy, sensitivity, specificity, and interrater agreement were calculated. Normal skin showed regular, well-defined, round vessels of uniform size, whereas BCC demonstrated heterogeneous, dilated, and elongated vessels with indistinct borders and variable flow. Mean sensitivity and specificity were 91.7 ± 4.7% and 86.7 ± 6.2%, respectively, with substantial interrater agreement (κ = 0.77). Vascular features in CLE sequences supported the differentiation between BCC and healthy skin. Incorporating vascular criteria into CLE-based scoring systems appears promising for a more standardized differentiation of BCC from healthy skin. Further studies with larger patient cohorts are warranted to validate these preliminary findings and assess their clinical applicability. Not applicable, as this study is a retrospective analysis.

Cognitive biases are systematic patterns of error in human thinking. While artificial intelligence (AI)-based decision support systems offer great potential to enhance diagnostic accuracy and efficiency in pathology, collaboration between medical professionals and AI can also introduce or amplify these biases. This review aims to identify cognitive biases and their modes of manifestation in human-computer interaction (HCI) within pathology and to extrapolate potential manifestations of biases not yet explored in this domain but reported in HCIs across other medical specialties. We conducted a structured literature review (19 August 2025) across the ACM, IEEE, and PubMed databases. Studies were eligible if they operationalized cognitive biases during expert–machine interaction in diagnostic decision-making using qualitative or quantitative methods, including review articles citing primary studies that met these criteria. The final corpus comprised 24 studies (8 primary studies and 16 review articles). From these review articles, 18 additional primary studies were extracted and used in their place solely for analysis purposes. A narrative synthesis of the identified primary research revealed 12 cognitive biases reported in AI-assisted medical decision-making, with only one study originating from pathology. For each bias, definitions and hypothetical pathology-specific examples were derived. This review is intended as a primer for the veterinary and human pathology communities and strives to contribute to the safe and effective integration of AI into diagnostic practice.

Automated measurements of anisokaryosis in canine cutaneous mast cell tumors (ccMCTs) have been shown to be predictive of survival, but questions remain regarding the intratumoral distribution of anisokaryosis. Whole-slide images of 96 ccMCTs were analyzed with a deep learning-based segmentation algorithm to quantify anisokaryosis using the standard deviation (SD) of the nuclear area. In 35/96 cases, >5% of the non-overlapping 256 × 256 µm ² regions were hotspots (SD ≥11.5 µm ² ). Regions selected by 7 pathologists within these 35 cases matched hotspots in 32% of the instances. Outcome analysis (tumor-related death) based on single tumor regions yielded an area under the curve (AUC) of 0.901 for pathologist-selected hotspots, falling between random region selection (AUC: 0.862) and 90th-percentile targeted selection (AUC: 0.956). Whole-slide analysis of the hotspot proportion predicted survival with an AUC of 0.956, with 20% of hotspots as a prognostically meaningful threshold. Whereas pathologists-selected tumor regions are prognostically meaningful for nuclear morphometry, whole-slide analysis may provide additional prognostic information.

Deep learning-based automated image analysis (DL-AIA) has been shown to outperform trained pathologists in tasks related to feature quantification. Related to these capabilities, the use of DL-AIA tools is currently extending from proof-of-principle studies to routine applications, including the evaluation of patient samples (diagnostic pathology), regulatory safety assessment (toxicologic pathology), and recurrent research tasks. To ensure that DL-AIA applications are safe and reliable, it is critical to conduct a thorough and objective generalization performance assessment to evaluate an algorithm’s ability to accurately predict patterns of interest and possibly evaluate model robustness (ie, the algorithm’s capacity to maintain predictive accuracy on images from different sources). In this article, we review the practices for performance assessment in veterinary pathology publications by which 2 main approaches were identified: (1) exclusive visual performance assessment (ie, eyeballing algorithmic predictions) plus validation of the model’s application utilizing secondary performance indices and (2) statistical performance assessment (alongside other methods), which requires creation of a test set with ground truth labels that is a hold-out from model development. This article compares the strengths and weaknesses of statistical and visual performance assessment methods. Furthermore, we discuss relevant considerations for rigorous statistical performance evaluation including metric selection, test data set image composition, ground truth label quality, resampling methods such as bootstrapping, statistical comparison of multiple models, and evaluation of model stability. It is our conclusion that visual and statistical evaluation have complementary strengths and a combined approach provides the greatest insight into the DL model’s performance and sources of error.

Whereas reproducibility of studies is a prerequisite for trustworthy deep learning (DL) in veterinary histopathology and microscopy, the actual degree of methodological transparency that exists in the literature remains uncertain. We performed a Preferred Reporting Items for Systematic Reviews and Meta-Analyses-guided systematic review to quantify the degree to which supervised DL and supervised machine learning studies report reproducibility-critical details. Using a veterinary-journal-restricted Boolean search executed in PubMed and Scopus, we screened 180 unique records and included 50 primary research articles for full-text analysis. Based on a recently published guideline for the development of DL models in veterinary pathology, we extracted information for each study across 5 dimensions: (1) study and task characterization, (2) data transparency, (3) experimental design and data-leakage control, (4) model and training details, and (5) performance evaluation and reporting. Among the included studies, private data sets predominated, with 90% of studies relying on private data. Sharing of code was uncommon (3%). Key training details such as augmentation and hyperparameters were often incompletely reported; augmentation was not reported in 56% of studies, and key hyperparameters were absent in 40% of studies. It was often not clear whether patient-level stratification (necessary to avoid data leakage) was performed. In summary, these results highlight major deficits in the reporting of details and experimental design necessary for reproducing DL results in veterinary histopathology. This review provides a practical baseline and reporting roadmap to support more transparent and reproducible research in veterinary computational pathology.

Supervised deep learning (DL) receives great interest for automated analysis of microscopic images with an increasing body of literature supporting its potential. The development and testing of those DL models rely heavily on the availability of high-quality, large-scale data sets. However, creating such data sets is a complex and resource-intensive process, often hindered by challenges such as time constraints, domain variability, and risks of bias in image collection and label creation. This review provides a comprehensive guide to the critical steps in data set creation, including (1) image acquisition, (2) selection of annotation software, and (3) annotation creation. For image acquisition, besides ensuring a sufficiently large number, it is important to address sources of image variability (domain shifts), such as those related to slide preparation and digitization, that could lead to algorithmic errors if not adequately represented in the training data. For annotations, key quality criteria are the 3 “C”s: correctness, consistency, and completeness. For mitigation of annotation bias of a single annotator, this review explores advanced annotation methods (eg, computer-assisted annotations). To support data set creators, a standard operating procedure is provided as supplemental material, summarizing all important considerations for data set creation. Furthermore, this article underscores the importance of open data sets in driving innovation and enhancing reproducibility of DL research. By addressing the challenges and offering practical recommendations, this review aims to advance the creation and availability of high-quality, large-scale data sets, ultimately contributing to the development of generalizable and robust DL models for pathology applications.