Handcrafted radiomic descriptors are widely used in biomedical image analysis, but classic radiomics pipelines often suffer from high feature redundancy and an underdeveloped, weakly principled feature-selection practice, which together can impair generalization and limit model interpretability. To address this, we introduce ReSHAPe (Redundancy-Reduced SHAP-based Evaluation), a two-stage, model-aware feature selection pipeline that makes SHAP-driven selection practical for radiomics. ReSHAPe first performs redundancy pruning by removing highly correlated features using Spearman rank correlation, retaining within each correlated group the descriptor with the lower absolute skewness. It then applies SHAP-based global importance to rank the remaining features and iteratively select a compact subset; an ensemble variant aggregates SHAP rankings across multiple classifiers to promote consensus and interoperability. We evaluate ReSHAPe on three MedMNIST v2 subsets (BreastMNIST, PneumoniaMNIST, BloodMNIST) using 285 handcrafted features and five well-known classifiers (SVM, Decision Tree, Random Forest, Extra Trees, XGBoost), comparing against univariate filters (ANOVA F-test, mutual information), SHAP-only selection, correlation-based filtering, and full-feature baselines. Across datasets, ReSHAPe preserves performance while drastically reducing dimensionality; on radiomic tasks, it is consistently competitive with SHAP-only selection f-measure weighted values differences typically lower than 0.03, and it remains effective in the non-radiomic multiclass setting (maximum decrease of f-measure weighted value lower than 0.04). Finally, the correlation pre-filtering stage markedly reduces SHAP overhead, which would otherwise require 200 additional model training/evaluation steps when applied directly to the full feature space.

Incomplete neuroimaging data remains a major challenge in Alzheimer’s disease diagnosis, as many patients undergo only a subset of recommended imaging protocols. This work addresses this limitation by proposing a generative transformer-based framework designed to support multimodal analysis in the presence of missing modalities. We systematically investigate multimodal performance and fairness within a unified foundation model framework for Alzheimer’s disease classification while introducing a generative approach that combines structural MRI, DTI, and PET data and leverages ControlNet-based diffusion models to synthesize anatomically consistent surrogate modalities when data are unavailable. These synthetic images are used exclusively as a training-time augmentation strategy for incomplete-modality settings, rather than as replacements for clinical acquisitions. Vision transformers adapted via Low-Rank Adaptation are employed for efficient feature extraction, while clinical variables are integrated through a dedicated projection module. Experimental results show that a transformer-based fusion head can improve over simple aggregation strategies in some complex multimodal settings, achieving an F1-score of in multiclass classification when combined with generative augmentation and clinical data. However, these benefits are not uniform since strong unimodal volumetric PET baselines remain superior in the best-case binary setting, and the effect of generative augmentation is strongly configuration-dependent, with some settings benefiting and others degrading substantially under non-selective synthetic augmentation.

Can the integration of vision and language representations advance artificial intelligence methods for automated white blood cell (WBC) analysis across heterogeneous clinical conditions? Motivated by this question, we present WBC-CLIP, a dual-encoder framework that enhances WBC classification and analysis by combining image data with rich textual descriptions derived from quantitative morphological features. Our method leverages multiple large language models to convert numerical and categorical cell attributes into diverse, semantically enriched textual descriptions. These captions are jointly embedded with their corresponding WBC images using a contrastive learning strategy inspired by the CLIP architecture, enabling the model to learn stable and meaningful cross-modal associations. We evaluate WBC-CLIP through zero-shot classification and image–text retrieval tasks across both in-distribution and out-of-distribution datasets. The framework advances automated WBC analysis while providing improved explainability by explicitly grounding visual representations in morphology-aware textual descriptors, addressing key challenges in computer-aided diagnostics.

This survey paper provides an overview of different feature types used in radiomics research and their applications across various medical imaging modalities and disease domains. The paper delves into the key aspects of the radiomics workflow, including data engineering techniques for image acquisition, preprocessing, fusion, and segmentation. It then presents a comprehensive review of the most commonly employed feature categories in radiomics, such as shape-based, first-order statistical, second-order texture, and transform-based features. The paper also discusses the emerging role of deep learning features extracted using convolutional neural networks, recurrent neural networks, and transformers. The analysis of feature usage trends across different anatomical regions and imaging modalities offers valuable insights that can guide the optimization of feature engineering strategies in future radiomics research. The survey concludes by highlighting several opportunities for further advancement in the field, including the need for larger multi-center datasets, multi-modal data fusion, self-supervised learning, and the development of efficient embedded models for on-device deployment.

White blood cell (WBC) subtype classification is a critical step in monitoring an individual’s health. However, it remains a challenging task due to the significant morphological variability of WBCs and the domain shift introduced by differing acquisition protocols across hospitals. Numerous approaches have been proposed to mitigate domain shift, including supervised and unsupervised domain adaptation, as well as domain generalisation. These methods, however, require a suitable amount of representative target images, even if unlabelled, or a suitable amount of images from multiple sources, which may not be feasible due to privacy regulations. In this study, we explore an alternative paradigm, known as Distributed Collaborative Machine Learning (DCML), which consists of exploiting images from different sources in a privacy-preserving setup. Although DCML methods seem well suited to this application, to the best of our knowledge, they have not been used for this task or to address the above-mentioned issues. However, we argue that DCML deserves further consideration in medical images as a potential alternative solution against domain shift in a privacy-preserving setup. To substantiate our view, we consider three DCML methods: early and late fusion and federated learning approaches, each offering distinct trade-offs in terms of training constraints, computational overhead and communications costs. We then conduct an extensive, cross-dataset experimental evaluation on four benchmark datasets and provide evidence that even simple implementations of DCML methods can effectively mitigate domain shift in WBC classification tasks.

Radiomics is an innovative discipline in medical imaging that uses advanced quantitative feature extraction from radiological images to provide a non-invasive method of interpreting the intricate biological panorama of diseases. This discipline takes advantage of the unique characteristics of medical imaging, where radiation or ultrasound combines with biological tissues, to reveal disease features and important biomarkers that are invisible to the human eye. Radiomics plays a crucial role in healthcare, spanning disease diagnosis, prognosis, recurrences, treatment response assessment, and personalized medicine. Radiomics uses a systematic approach that includes image preprocessing, segmentation, feature extraction, feature selection, classification, and evaluation. This survey attempts to shed light on the crucial roles that feature selection and classification play in discovering important biomarkers and forecasting disease directions despite the challenges posed by high dimensionality (i.e., when the data contains a huge number of features). By analyzing 47 relevant research articles, this study has provided several insights into the key techniques used across different stages of the radiology workflow. The findings indicate that 27 articles utilized the SVM classifier, while 23 of the surveyed studies used the LASSO feature selection approach. This demonstrates how these particular methodologies have been widely used in Radiomics research. The assessment did, however, also point out areas that require more research, such as evaluating the stability of feature selection and classification algorithms and adopting novel approaches like ensemble and hybrid selection methods. Additionally, we examine some of the challenges and emerging subfields within the field of radiomics.

This study addresses the critical challenge of cell nuclei segmentation in histopathological image analysis, which is essential for cancer diagnosis and prognosis. Traditional segmentation methods often struggle with complexities such as overlapping nuclei and variations in shape and size. Recent advancements in machine learning and deep learning, particularly convolutional neural networks, have improved segmentation accuracy but face limitations due to the need for large annotated datasets, often constrained by privacy regulations in the medical field. This research explores federated learning as a solution, enabling collaborative model training across institutions without sharing sensitive patient data. By leveraging diverse datasets while maintaining privacy, federated learning enhances model robustness and generalization capabilities. The study evaluates various federated learning techniques on in- and out-of-domain test sets, aiming to improve the reliability of segmentation models in clinical settings. The findings suggest that federated learning techniques can effectively address the challenges of data scarcity and domain shift, paving the way for more accurate and widely applicable segmentation algorithms in computational pathology.

Within the contemporary dairy industry, the effective monitoring of cheese ripeness constitutes a critical yet challenging task. This paper proposes the first public dataset encompassing images of cheese wheels that depict various products at distinct stages of ripening and introduces an innovative hybrid approach, integrating machine learning and computer vision techniques to automate the detection of cheese ripeness. By leveraging deep learning and shallow learning techniques, the proposed method endeavors to overcome the limitations associated with conventional assessment methodologies. It aims to provide automation, precision, and consistency in the evaluation of cheese ripeness, delving into a hierarchical classification for the simultaneous classification of distinct cheese types and ripeness levels and presenting a comprehensive solution to enhance the efficiency of the cheese production process. By employing a lightweight hierarchical feature aggregation methodology, this investigation navigates the intricate landscape of preprocessing steps, feature selection, and diverse classifiers. We report a noteworthy achievement, attaining a best F-measure score of 0.991 through the merging of features extracted from EfficientNet and DarkNet-53, opening the field to concretely address the complexity inherent in cheese quality assessment.

The rapid increase of the elderly population and new advances in pervasive computing technologies allow innovative tools and applications to support independent living for frail people and identify early symptoms of health problems, including neurodegenerative disorders. Among several studies reported in the literature, monitoring locomotion traces to detect symp-toms of cognitive impairment has gained increasing attention. Therefore, in this work, we propose a novel technique for the recognition of locomotion patterns related to cognitive decline based on sensor data acquired in smart homes. In particular, we introduce a vision-based method to graphically represent indoor trajectories with random rotation, using different handcrafted features designed for image analysis tasks and combined with features extracted directly from spatio-temporal sequences of movements. Experiments on a real-world dataset acquired in a smart-home test-bed show that the proposed approach achieves promising results.

Mobile sensor-based activity recognition is a growing research field with important applications areas, such as healthcare and well-being. Data collected from multiple sensors, including smartphone sensors that are now ubiquitous, can be exploited to build predictive models capable of recognizing actions and activities performed by humans in their daily life. This involves several processing steps, from the cleansing of raw data to the extraction of suitable features and the induction of proper classifiers. Dimensionality reduction techniques may also be important for the efficiency and the exploitability of the induced models, especially when dealing with multi-sensor data leading to high-dimensional feature vectors. In such a scenario, feature selection algorithms can be very useful to identify and retain only the most informative and predictive features. However, little research has so far investigated which selection approaches may be most appropriate in sensor-based activity recognition tasks. To give a contribution in this direction, our paper compares the performance of different feature selection methods, both univariate and multivariate, on a public domain benchmark containing smartphone sensor data. We performed a comprehensive evaluation considering the extent to which each method effectively identifies the most predictive features and the overall stability of the selection process, i.e., its robustness to changes in the input data. Our results give interesting insight on which methods may be most suited in this domain, showing that it is possible to significantly reduce the data dimensionality without compromising the activity recognition performance.

Image analysis is an important tool for several application fields, like biology and especially botany. Analysis of seed fossils can provide important information about their evolution, on agriculture origin, on plants domestication and knowledge of diets in ancient times. The aim of this work is to make the analysis process simple for biologists, by obtaining all the features needed for botanist user through a unique framework, that is still not available at the moment. We propose an ImageJ plugin able to extract morphological, textural and color features from seeds images in order to use them for classification. The experimental results have confirmed the goodness and correctness of the extracted features, making the proposed framework easily extendable to other application domains.