Abstract
As in school, one teacher to cover all subjects is insufficient to distill equally robust information to a student. Hence, each subject is taught by a highly specialised teacher. Following a similar philosophy, we propose a multiple specialized teacher framework to distill knowledge to a student network. In our approach, directed at face recognition use cases, we train four teachers on one specific ethnicity, leading to four highly specialized and biased teachers. Our strategy learns a project of these four teachers into a common space and distill that information to a student network. Our results highlighted increased performance and reduced bias for all our experiments. In addition, we further show that having biased/specialized teachers is crucial by showing that our approach achieves better results than when knowledge is distilled from four teachers trained on balanced datasets. Our approach represents a step forward to the understanding of the importance of ethnicity-specific features.

Abstract
Algorithmic bias in healthcare mirrors existing data biases. However, the factors driving unfairness are not always known. Medical devices capture significant amounts of data but are prone to errors; for instance, pulse oximeters overestimate the arterial oxygen saturation of darker-skinned individuals, leading to worse outcomes. The impact of this bias in machine learning (ML) models remains unclear. This study addresses the technical challenges of quantifying the impact of medical device bias in downstream ML. Our experiments compare a perfect world, without pulse oximetry bias, using SaO(2) (blood-gas), to the actual world, with biased measurements, using SpO(2) (pulse oximetry). Under this counterfactual design, two models are trained with identical data, features, and settings, except for the method of measuring oxygen saturation: models using SaO(2) are a control and models using SpO(2) a treatment. The blood-gas oximetry linked dataset was a suitable testbed, containing 163,396 nearly-simultaneous SpO(2) - SaO(2) paired measurements, aligned with a wide array of clinical features and outcomes. We studied three classification tasks: in-hospital mortality, respiratory SOFA score in the next 24 h, and SOFA score increase by two points. Models using SaO(2) instead of SpO(2) generally showed better performance. Patients with overestimation of O-2 by pulse oximetry of >= 3% had significant decreases in mortality prediction recall, from 0.63 to 0.59, P < 0.001. This mirrors clinical processes where biased pulse oximetry readings provide clinicians with false reassurance of patients' oxygen levels. A similar degradation happened in ML models, with pulse oximetry biases leading to more false negatives in predicting adverse outcomes.

Abstract
Nuclear-derived morphological features and biomarkers provide relevant insights regarding the tumour microenvironment, while also allowing diagnosis and prognosis in specific cancer types. However, manually annotating nuclei from the gigapixel Haematoxylin and Eosin (H&E)-stained Whole Slide Images (WSIs) is a laborious and costly task, meaning automated algorithms for cell nuclei instance segmentation and classification could alleviate the workload of pathologists and clinical researchers and at the same time facilitate the automatic extraction of clinically interpretable features for artificial intelligence (AI) tools. But due to high intra- and inter-class variability of nuclei morphological and chromatic features, as well as H&Estains susceptibility to artefacts, state-of-the-art algorithms cannot correctly detect and classify instances with the necessary performance. In this work, we hypothesize context and attention inductive biases in artificial neural networks (ANNs) could increase the performance and generalization of algorithms for cell nuclei instance segmentation and classification. To understand the advantages, use-cases, and limitations of context and attention-based mechanisms in instance segmentation and classification, we start by reviewing works in computer vision and medical imaging. We then conduct a thorough survey on context and attention methods for cell nuclei instance segmentation and classification from H&E-stained microscopy imaging, while providing a comprehensive discussion of the challenges being tackled with context and attention. Besides, we illustrate some limitations of current approaches and present ideas for future research. As a case study, we extend both a general (Mask-RCNN) and a customized (HoVer-Net) instance segmentation and classification methods with context- and attention-based mechanisms and perform a comparative analysis on a multicentre dataset for colon nuclei identification and counting. Although pathologists rely on context at multiple levels while paying attention to specific Regions of Interest (RoIs) when analysing and annotating WSIs, our findings suggest translating that domain knowledge into algorithm design is no trivial task, but to fully exploit these mechanisms in ANNs, the scientific understanding of these methods should first be addressed.

Abstract
In many real-world prediction tasks, the class labels contain information about the relative order between the labels that are not captured by commonly used loss functions such as multicategory cross-entropy. In ordinal regression, many works have incorporated ordinality into models and loss functions by promoting unimodality of the probability output. However, current approaches are based on heuristics, particularly nonparametric ones, which are still insufficiently explored in the literature. We analyze the set of unimodal distributions in the probability simplex, establishing fundamental properties and giving new perspectives to understand the ordinal regression problem. Two contributions are then proposed to incorporate the preference for unimodal distributions into the predictive model: 1) UnimodalNet, a new architecture that by construction ensures the output is a unimodal distribution, and 2) Wasserstein regularization, a new loss term that relies on the notion of projection in a set to promote unimodality. Experiments show that the new architecture achieves top performance, while the proposed new loss term is very competitive while maintaining high unimodality.

Abstract
Deep learning in computational pathology (CPath) has rapidly advanced in recent years. Research has primarily focused on enhancing accuracy and interpretability across various histology image analysis tasks, from tile-level to slide-level foundation models and novel multiple instance learning (MIL) strategies. However, it is equally important for models to provide well-calibrated confidence estimates. Due to factors such as dataset bias, overfitting, and limited training data, existing models tend to be overly confident on test sets. Promising solutions to address this issue include temperature scaling, a post-hoc method that adjusts logits using a single scalar value. However, the role of calibration in CPath is yet to be clarified. In this study, we evaluate temperature scaling and linear temperature scaling for CPath tasks, analyzing their impact on recalibration in both in-domain and out-of-domain distributions. The results show the limitations of current probability calibration techniques and motivate future work. © 2025 IEEE.

Abstract
Immunohistochemistry (IHC) is crucial for the clinical categorisation of breast cancer cases. Deep generative models may offer a cost-effective alternative by virtually generating IHC images from hematoxylin and eosin samples. This review explores the state-of-the-art in virtual staining for breast cancer biomarkers (HER2, PgR, ER and Ki-67) and benchmarks several models on public datasets. It serves as a resource for researchers and clinicians interested in applying or developing virtual staining techniques.