BIOIMAGING 2026 Abstracts


Full Papers
Paper Nr: 136
Title:

A Sector-Based Radiomic Approach to Characterize Carotid Plaque Calcification Dynamics from Computer Tomography Angiography

Authors:

Marina De Santis, Aikaterini Tziotziou, Daniel Bos, Luca Mainardi, Valentina D. A. Corino, Ali C. Akyildiz and Anna Corti

Abstract: Understanding the role of calcification changes in vulnerable carotid plaques is crucial to clarify the course of atherosclerosis and to prevent cerebrovascular events. This study proposes a novel sector-based radiomics approach to investigate the relationship of baseline local radiomic features with morphological evolution of plaque calcifications over time. Eighteen patients underwent multi-detector computational tomography angiography (MDCTA), and vessel sectorization was performed along the reconstructed lumen centerline. Radiomic features were extracted sector by sector, and their association with longitudinal changes in calcium thickness over 2-year follow-up was assessed using odds ratio (OR). Selected textural features related to structural heterogeneity were significantly associated with calcium progression (OR range: 7.54-9.08), with the highest odds observed for ngtdm_Complexity (OR = 9.08, 95% CI 6.9-11.8). These findings support the potential of radiomics as a non-invasive biomarker of plaque development.

Paper Nr: 184
Title:

Identifying Statistical Predictors of Borderline Personality Disorder Using fMRI Data

Authors:

Spoorthi Kalkunte, Shubha V. Hegde, Shreya S. Adiga, Riya Jha and Gowri Srinivasa

Abstract: Borderline Personality Disorder (BPD) is a complex and debilitating mental health condition that affects a significant portion of the population. The diagnosis of borderline personality disorder is typically based on subjective interpretation of patient responses and behavior. This study aims to create an objective diagnostic pipeline to supplement prevalent diagnostic methods with analysis of fMRI data. We use publicly available data that contains anatomical and functional MRI images from 36 subjects, including 20 patients, and 16 healthy controls, collected during the Cyberball social exclusion task. Machine learning models have been trained for classification, and first-level and second-level General Linear Model analyses are performed to find crucial predictors. A classification model with the Welch method for feature extraction is seen to yield the most promising results for objective classification of BPD. Upon increasing exclusion, higher activation is observed in patients in the temporal gyrus, thalamus, and the right postcentral gyrus, and hypoactivation in the right superior parietal gyrus and precentral gyrus. Increased functional connectivity is also observed between the amygdala and the insular cortex as well as the frontal pole, and between the amygdala and parahippocampal gyrus. This study seeks to deepen our comprehension of BPD and potentially enhance diagnostic procedures.

Paper Nr: 196
Title:

Towards Data Quality-Aware Dataset Distillation in Bioimaging

Authors:

Bárbara Capelo, Maria Russo, André Carreiro, Hugo Gamboa and Duarte Folgado

Abstract: The success of modern machine learning depends on large, diverse, and annotated datasets, yet in healthcare, such data remain fragmented across institutions and challenging to share due to strict privacy and governance regulations. Dataset distillation offers a potential solution by synthesizing a small set of representative samples that encapsulate the information of the original dataset, serving as compact and effective drop-in replacements. However, current approaches largely assess dataset distillation performance through downstream task accuracy, neglecting the intrinsic quality of the synthesized data. This limitation is especially relevant in bioimaging, where fidelity and diversity directly impact clinical reliability and trust. In this work, we present a preliminary study exploring the role of data quality in guiding and evaluating dataset distillation. We complement traditional task-based evaluation with intrinsic measures of fidelity and diversity to better characterize the overall quality of distilled datasets. Using a blood cell microscopy benchmark, we assess how different initialization strategies affect the quality and effectiveness of the resulting distilled datasets. Our findings provide early evidence that explicit quality measures can offer complementary insights beyond model performance. This feasibility study represents an initial step toward quality-aware dataset distillation, supporting the development of more reliable synthetic datasets for medical imaging applications.

Paper Nr: 214
Title:

From Scan to Plan: Organ-Specific Deep Learning Networks for Brachytherapy

Authors:

Nirmalya Gayen, Aditya Kumar, Bhuneshwar Singh Netam, Nithin Shivashankar, Kirthi A. S. Koushik and Vijay Natarajan

Abstract: The accurate identification of anatomical structures within volumetric data derived from medical scanning devices, such as CT and MRI, is a significant aspect of clinical workflows in radiology and oncology treatment planning. With the advance of AI and high-performance computing, many methods and tools have been developed over the past decade and a half. However, end-to-end integration of solutions with existing workflows and practices remains a challenge. Here, we focus on the need for segmentation of anatomical structures whose delineations are clinically defined by a combination of anatomy, function, and treatment planning. Existing deep learning approaches for segmentation often struggle to effectively differentiate closely placed organs, such as the bladder, rectum, and sigmoid colon. We propose an efficient and robust, organ-specific segmentation pipeline based on tailored 2D U-Net models, coupled with anatomy-guided preprocessing and geometric postprocessing algorithms. We validate our method by a user study involving trained radiation oncologists, demonstrating high segmentation accuracy and significant reductions in contouring time. The results show that our approach produces consistently accurate contours that closely match expert delineations, with minimal corrections needed in clinical practice. This work highlights the benefits of deep learning integration in brachytherapy, enabling quicker planning and improved consistency through clinically validated organ segmentation.

Paper Nr: 236
Title:

PCA-Based Identification of Soft Tissue Regions with Irradiation-Induced CT-Density Changes for Image-Guided Radiotherapy

Authors:

Greta Karpavičienė, Algimantas Kriščiukaitis, Reda Čerapaitė-Trušinskienė, Robertas Petrolis, Diana Meilutytė-Lukauskienė and Renata Paukštaitienė

Abstract: Patient positioning, together with physical changes in soft tissues during the course of radiotherapy, significantly affects the irradiation of targeted tissues and adjacent structures. This leads to a high risk of under-dosage of the tumour and/or over-dosage of critical normal structures during the late sessions of treatment. Accurate and timely identification of irradiation-affected tissue regions is highly valuable for adaptive radiotherapy planning. We propose a method for the identification and evaluation of specific computed tomography (CT) attenuation changes that can reveal the affected tissue regions. The search for correlated CT attenuation changes in the tumour and surrounding tissues, based on principal component analysis of series of intensity values in each fixed voxel, can reveal the actual three-dimensional region of irradiation-affected tissues for radiotherapy control and replanning.

Paper Nr: 278
Title:

Partial Volume Correction with/out Spillover Correction in Postsurgical SPECT/CT Thyroid Imaging

Authors:

Elena Ttofi, Theodoros Leontiou, Costas Kyriacou and Yiannis Parpottas

Abstract: SPECT imaging is essential in evaluating differentiated thyroid cancer, particularly in post-thyroidectomy patients. Individualized dosimetry and therapy require accurate volume calculation and uptake quantification, both affected by partial volume (PV) and spillover (SO) effects. This study assessed the individual and combined impact of SO and PV corrections in SPECT/CT imaging for accurate calculation of small volumes and uptake values. Scatter- and attenuation-corrected I-123 and I-131 SPECT/CT images were acquired from a neck–thyroid phantom containing small remnants (0.5–10 mL) and administered with diagnostic activities, with or without background radiation. Two custom algorithms were utilized to correct for SO and PV. For smaller volumes (0.5–1.5 mL), the % differences between calculated and actual volumes were 5.3±0.6 % and 7.0±0.6 % for the I-123 and I-131 data, respectively, when both corrections were applied successively. Applying only SO, the % differences were increased to 11.3±0.7 % and 12.2±0.9 %, respectively, while PVC alone produced overestimated volumes. Higher background radiation did not affect volume accuracy after applying SO or combined corrections. Uptake reduction within the target volume was proportional to the volume reduction. Successive application of SO and partial PV corrections is recommended for accurate assessment of small volumes and uptake in postsurgical thyroid SPECT/CT imaging.

Paper Nr: 282
Title:

A Framework for Assessing and Optimising Data Sufficiency in Ultrasound Tongue Imaging

Authors:

Saja Al Ani, Joanne Cleland and Ahmed Zoha

Abstract: Deep learning (DL) applied to ultrasound tongue imaging (UTI) for speech-disorder assessment is limited by the cost and scarcity of expert-annotated data. Each ultrasound frame requires labelling by trained speech and language therapists (SLTs), making large-scale dataset construction expensive and time-consuming. This paper presents a cost-aware framework that integrates statistical power-curve modelling with active learning (AL) to optimise dataset size and annotation efficiency. Power-curve analysis quantifies the relationship between dataset size and classification performance, identifying a point of diminishing returns beyond which additional annotation yields minimal improvement. Experiments on paediatric UTI data from 28 participants showed that performance stabilised when approximately 65–70% of the available training data were used, reaching an asymptotic accuracy of around 90%. Building on this, uncertainty-based AL further reduced annotation requirements by prioritising informative samples, achieving comparable performance with only 40% of labelled data. A combined cost analysis demonstrated that integrating dataset-size optimisation with selective annotation reduced total data-collection and annotation costs by approximately 45% relative to full annotation, without compromising performance. The proposed framework provides a quantitative and reproducible methodology for planning data collection and annotation in clinical speech-imaging research, supporting more scalable and resource-efficient development of DL-based UTI analysis systems.

Paper Nr: 412
Title:

HistoKit: Fast and Accurate Tissue and Artifact Detection and Data Processing for Whole Slide Histopathological Imaging

Authors:

Julia Merta and Michał Marczyk

Abstract: Recently, computational pathology has undergone significant advancements. Numerous deep learning–based methods have been developed for analyzing whole slide images (WSI), transforming traditional histopatho-logical analysis. Multiple existing approaches share the most important data processing steps. However, most solutions are not universal, implement custom, insufficiently tested pipelines that are error-prone, limiting advanced WSI analysis. To address this limitation, we present HistoKit, an open-source Python package for WSI processing. The proposed tool supports the following operations: (i) tissue segmentation; (ii) detection of technical artifacts using GrandQC; (iii) parallelized patch extraction; (iv) staining normalization and data augmentation. Additionally, the GrandQC model prediction pipeline has been modified to reduce transition artifacts between patches, resulting in improved artifact detection. HistoKit was compared with three other libraries: TIA Toolbox, HistomicsTK, and Histolab. The tissue detection algorithm, combined with efficient postpro-cessing, produces visually superior tissue masks compared to those generated by other software. Only HistoKit allows the detection and removal of different classes of artifacts. Finally, the proposed solution enables significantly faster patch extraction than existing tools. Overall, HistoKit offers a reliable, user-friendly solution for processing histopathological images, enabling subsequent analysis and the development of computational pathology models. The source code is publicly available on GitHub: https://github.com/ZAEDPolSl/HistoKit.

Short Papers
Paper Nr: 28
Title:

Three-Dimensional Visualization of Inadequate Adhesive Placement under Dental Restorations: A Synchrotron Radiation Micro-CT Study

Authors:

Assem Hedayat

Abstract: Our objective is to delineate the inadequacies of adhesive placement under dental restorations using synchrotron radiation micro-CT (SRµCT) imaging, and three-dimensional (3D) rendering software. Human teeth were resourced randomly from the tooth bank at the College of Dentistry, University of Saskatchewan. All teeth were restored and divided into three groups: 1) random, previously restored teeth, 2) teeth restored earlier by our team for another study without knowledge of the goals of this research, and 3) teeth restored by our team with knowledge of our objectives. We scanned all teeth at the BioMedical Imaging & Therapy (BMIT) facility at the Canadian Light Source (CLS). The 05ID-2 beamline scanned the teeth at 50 keV and captured 4.3 µm pixel size images. Avizo® 9.0 software was used for 3D visualization of both sides of interfaces and plotting their gray value profile. All teeth representing the three groups revealed inadequate adhesive placement under dental fillings. The deficiency varied at the disconnected interfaces between the restorations and the adjacent dentin or enamel. Non-destructive SRµCT imaging proved that placement of dental adhesives is operator dependent. This technology is necessary to assess and ameliorate techniques for adequate adhesive placement in a way that reduces operators’ dependence.

Paper Nr: 34
Title:

Classifying Post-COVID-19 Motor and Vocal Tics Using Texture Analysis in Magnetic Resonance Images

Authors:

Murilo Costa de Barros, Kaue Tartarotti Nepomuceno Duarte, Yen-Ju Chu, Wang-Tso Lee and Marco Antonio Garcia de Carvalho

Abstract: Movement disorders are a heterogeneous group of neurological conditions marked by involuntary motor activity tics, tremors, dystonias and include manifestations reported after COVID-19. We propose a classification method to identify post-COVID-19 patients with motor and/or vocal tics by extracting radiomic texture features from structural MRI and applying a hierarchical Support Vector Machine (SVM). The pipeline comprised preprocessing and segmentation of brain volumes, region-wise texture feature extraction, and hierarchical SVM classification. The dataset included 76 subjects (38 controls, 38 post-COVID-19). Texture analysis reliably discriminated controls from COVID patients and separated symptomatic subgroups: ventral diencephalon (left) achieved 75% accuracy for Control vs COVID; medial orbitofrontal (left) reached 69% for COVID with motor tics vs without motor tics; superior temporal (left) reached 80% for COVID with vocal tics vs without vocal tics. We observed a predominance of left-hemisphere alterations and involvement of cortical and subcortical regions related to motor control and language. These findings support the utility of structural MRI texture features combined with SVMs to detect and characterize post-COVID-19 motor and vocal manifestations.

Paper Nr: 54
Title:

DDPM-Based Histopathology Data Augmentation for Blood Vessel Segmentation

Authors:

Tomas Tanczos, Wanda Benesova, Jarmila Pavlovicova and Andrea Vajsová

Abstract: Deep learning approaches for histopathological image analysis require large, well-annotated datasets, which are difficult to obtain due to the need for domain expertise and privacy constraints. This work addresses dataset limitations through synthetic data augmentation using Denoising Diffusion Probabilistic Models (DDPMs). We propose a unified framework that explores two complementary strategies: fully synthetic image generation and partially synthetic generation via inpainting. Our approach systematically compares pixel-space and latent-space diffusion models on our in-house histopathology datasets consisting of post-transplant heart tissue biopsies from the Institute for Clinical and Experimental Medicine (IKEM). We evaluate the quality of synthetic data using standard metrics (KID, FID, LPIPS) as well as expert pathologist assessment. The synthetic samples are subsequently employed for augmentation in segmentation pipelines to assess improvements in detecting underrepresented structures, particularly blood vessels. Our experiments demonstrate that synthetic data augmentation enhances recall (a substantial gain in medical imaging where sensitivity to fine structures is critical) in vessel detection (from 0.34 to 0.42) while maintaining overall segmentation capability, though with some trade-offs in precision. These findings indicate that systematic diffusion-based augmentation represents a promising approach to addressing class imbalance in histopathological datasets, particularly for rare anatomical structures.

Paper Nr: 100
Title:

A Method for Privacy-Preserving and Explainable Pneumonia Detection and Localisation

Authors:

Miriam Di Renzo, Filomena Niro, Desirè Menditto, Patrizia Agnello, Marta Petyx, Mario Cesarelli, Fabio Martinelli, Antonella Santone and Francesco Mercaldo

Abstract: Pneumonia is a respiratory infection that affects the lungs, which are organs belonging to the lower respiratory tract of the human body. In general, this infection occurs because of the entry of pathogens and microparti-cles during the mechanisms of inhalation and exhalation. In medicine, it is possible to distinguish three main classes of pneumonia: bacterial, viral and fungal. Currently, Deep Learning (DL) methods have shown promising potential in diagnosing different categories of pneumonia using imaging data; however, their adoption in clinical practice remains limited for reasons related to patient privacy protection and the explainability of the models. For these reasons, in this paper we propose Federated Machine Learning (FML) approach for the classification of X-ray images connected to pneumonia while maintaining data confidentiality. The explain-ability of the model is guaranteed by the use of the Grad-Cam technique. The experimental analysis shows that the federated method is able to obtain significant performances, with a precision of 0.935, an accuracy of 0.933, a recall of 0.940 and, finally, an F-measure index of 0.939, while maintaining data confidentiality.

Paper Nr: 132
Title:

Reproducible PINN Framework for Patient-Specific Modeling and Biophysical Parameter Inference in Glioblastoma

Authors:

Juliette Vanderhaeghen, Cyril Corbet, François P. Duhoux, Christophe Pierreux and Sébastien Jodogne

Abstract: Cancer remains a major global health challenge, motivating personalized treatments. PDE-based models can capture tumor dynamics and enable patient-specific predictions, but traditional solvers like FDM or FEM can be computationally costly and require extensive calibration, while purely data-driven neural networks often lack interpretability. Physics-Informed Neural Networks (PINNs) address these limitations by embedding PDE constraints, supporting both forward simulations and inverse parameter estimation. We model glioblastoma growth using a reproducible, open-source PINN framework based on the Fisher–KPP equation. A systematic hyperparameter study evaluates architecture, activation functions, optimizers, learning rates, batching, and sampling strategies. Experiments on synthetic tumors show accurate dynamics and reliable recovery of biophysical parameters. We further provide a standalone Python implementation, transparent datasets, and practical guidelines for reproducible research in personalized oncology.

Paper Nr: 160
Title:

Contrast Comparison and Enhancement of Images from Different Digital Mammography Equipment by Converting Imaging Systems Characteristic Curves

Authors:

Homero Schiabel, Matheus T. Pereira and Ana Cláudia Patrocínio

Abstract: This study presents a methodology for quantitative comparison and enhancement of contrast of digital mammography images acquired from different digital mammography systems. Using an anthropomorphic breast phantom, characteristic curves were derived from both raw and processed images obtained with three distinct mammography systems, each employing different detector technologies and target/filter combinations. Polynomial interpolation using cubic splines was applied to determine conversion functions between characteristic curves, allowing pixel-wise contrast transformations across systems. Quantitative contrast evaluation was performed using two complementary metrics: a new Contrast Index (ICC) and Contrast-to-Noise Ratio (CNR). Evaluation tests identified the images with lower and higher contrast among those selected from the exposures. The image with poorer contrast was enhanced by changing their pixel intensities using the appropriate conversion curve determined by the method. Results demonstrated that some converted images achieved higher ICC and CNR values compared to the corresponding original image, with one in particular yielding the most significant improvement. Histogram analysis confirmed broader gray-level distributions in converted images, suggesting enhanced structural definition. These findings indicate that the proposed approach can effectively change mammographic image contrast, potentially contributing to image standardization across acquisition systems and supporting more robust applications in computer-aided diagnosis and artificial intelligence-based lesion detection.

Paper Nr: 178
Title:

Assessing Modern Deep Vision Models for Chest X-Ray Diagnostics in Emergency Care

Authors:

Ferran Soler-Guiral, Manuel F. Dolz, José I. Aliaga and Katty Delgado-Barriga

Abstract: Accurate identification of pathologies in chest X-rays is essential for timely diagnosis, especially in emergency settings. This study presents a deep learning-based classification system structured around a hierarchical 32-category pathology taxonomy, tailored to the workflow of radiologists at a collaborating hospital. The system combines convolutional neural networks and vision transformers for deployment within medical imaging software to support triage. A two-phase hyperparameter optimization showed that the best-performing convo-lutional network achieved an AUC of 0.911, closely matching the 0.914 AUC of the top vision transformer, demonstrating the practicality of convolutional models in resource-constrained environments. The study also highlights inconsistencies in data augmentation and regularization, as performance varied minimally across augmentation levels, emphasizing the need for systematic methodologies in clinical model training.

Paper Nr: 197
Title:

Anatomy-Preserving Diffusion-Based Data Augmentation for Medical Image Segmentation

Authors:

Hiroki Tominari, Natsuki Nakayama, Naoko Arakawa and Hiroyasu Usami

Abstract: High-precision medical image processing tasks, such as tumor detection and segmentation, generally require large amounts of training data. However, constructing large-scale medical image datasets is challenging due to privacy and ethical constraints, as well as the rarity of certain diseases. To address this issue, data augmentation using generative models, including Generative Adversarial Networks (GANs), has been extensively investigated, and their effectiveness has been demonstrated in various medical imaging tasks (Frid-Adar et al., 2018). Nevertheless, GAN-based approaches often suffer from training instability, making it difficult to consistently ensure the quality and diversity of generated images. In recent years, advances in diffusion models have led to an increasing number of studies exploring diffusion-based data augmentation for medical image analysis, and more stable generation frameworks are gradually being established (Nazir et al., 2025). However, in the context of medical image segmentation, several practical aspects remain insufficiently organized, including (i) operation on small-scale datasets, (ii) reproducible procedures for prompt design, and (iii) implementations that explicitly preserve anatomical structures. In this study, we focus on these practical limitations and evaluate diffusion-based data augmentation, specifically using Stable Diffusion, for medical image segmentation. We present a reproducible augmentation workflow encompassing prompt generation, sample selection, anatomy-preserving image generation, and quantitative evaluation. Medical image datasets are augmented using Stable Diffusion, and multiple segmentation models, such as U-Net, are trained on the augmented datasets. The effectiveness of the proposed workflow is quantitatively evaluated using the Intersection over Union (IoU) and Dice coefficient across multiple datasets.

Paper Nr: 264
Title:

Towards an Efficient Morphological Characterization of Naturally-Derived Hydroxyapatite Scaffolds: Preliminary Outcomes

Authors:

Elisabetta Salerno, Alice Betti, Simone Borghi, Claudio Ongaro, Barbara Zardin, Diego Angeli, Melania Maglio, Gregorio Marchiori, Margherita Peruzzini and Jessika Bertacchini

Abstract: The morphological characterization of porous scaffolds for bone regeneration is crucial to quantify fluid dynamic parameters that influence cell response and scaffold integration in bio-engineering devices. In this study, hydroxyapatite scaffolds of natural origin are considered, presenting an irregular internal architecture. Three distinct imaging techniques are explored to extract pore features, including: (i) microscope images combined to semi-automatic image processing in ImageJ, (ii) microscope images combined to a Computer Vision approach based on Python tools and (iii) micro-Computed Tomography. Pore distributions are obtained from two samples and used to compute scaffold flow resistance and flow shear stress. The ImageJ-based method produces the most consistent results but requires heavy manual refinement, whereas the other two methods offer faster data processing but their detection accuracy must be improved. Hence, an integrated strategy to be implemented in future developments is proposed.

Paper Nr: 307
Title:

Optimized U-Net Convolutional Autoencoder for Low-Dose CT Denoising: Technical Improvements and Clinical Validation

Authors:

Simone Damiani, Patrizio Barca, Marco Giannelli, Francesca Lizzi, Emanuele Neri, Alessandra Retico, Chiara Romei, Camilla Scapicchio, Maria Irene Tenerani, Antonio Traino, Arman Zafaranchi and Maria Evelina Fantacci

Abstract: Low-Dose Computed Tomography (LDCT) plays a crucial role in the early detection of lung cancer, but reducing radiation exposure inevitably increases image noise, which can hinder diagnostic accuracy. While numerous denoising strategies have been proposed, Deep Learning (DL) approaches are often constrained by the limited availability of large, high-quality clinical datasets. We argue that transfer learning can bridge this gap by enabling robust LDCT denoising from scarce clinical data. To support this position, we present a U-Net based Convolutional Autoencoder (UNbCAE) initially trained on phantom images and subsequently adapted to clinical chest-CTs through transfer learning. Our analysis on the LUNA16 dataset indicates that this approach reduces noise magnitude by a factor of (3.4±0.6), while qualitative assessment by two experienced radiologists confirmed improved perceived image quality and enhanced pulmonary nodule detectability. We position this framework as an efficient and generalizable solution for LDCT denoising under data-limited conditions, highlighting its potential for application in lung cancer screening settings.

Paper Nr: 321
Title:

Comparison of Colour Normalization Techniques for Skin Optical Imaging in Digital Tissue Phantoms

Authors:

Leah DeVos, Guennadi Saiko and Alexandre Douplik

Abstract: Background: Subtle changes in skin colour due to cardiovascular activity may provide important clinical information. However, clinically relevant information extracted from it is susceptible to variations in lighting conditions, which may impact the interpretation of the captured data. Aim: We aimed to evaluate colour normalization techniques for producing illumination-independent colour representations of the skin. Material and Methods: Tissue reflectance was computed using Monte Carlo modelling of light propagation in multilayer skin phantoms, then convolved with four different illuminant spectra and corresponding sensor responses to obtain colour representation in CIELUV colour space. Three colour normalization techniques (spectral normalization on white reference, Bradford CAT, and Von Kries CAT) were applied and compared to evaluate their effectiveness in reducing illumination-dependent variability. Results: Among the evaluated methods, the Von Kries chromatic adaptation transform (CAT) produced the most consistent colour representation across varying illuminants. Conclusion: These results can guide the future development of image processing techniques for physiological and anatomical optical imaging.

Paper Nr: 352
Title:

Attention-Guided U-Net for Cell Nucleus Segmentation in Microscopy Images

Authors:

Saqib Nazir and Ardhendu Behera

Abstract: Cell nuclei segmentation is a fundamental step in computational pathology and biomedical image analysis, enabling downstream tasks such as disease diagnosis and drug discovery. However, existing deep learning based approaches often rely on heavy encoders or complex multi-branch designs, leading to large parameter counts and limited practicality in clinical settings. We propose a lightweight encoder–decoder architecture that achieves improved segmentation performance with significantly reduced model complexity. Our custom residual encoder leverages dilated convolutions and Squeeze-and-Excitation (SE) modules to capture rich contextual features, while a Spatial Pyramid Pooling bottleneck enhances multi-scale representation. Extensive experiments on three diverse benchmarks demonstrate that the proposed model consistently outperforms or matches State-of-the-art (SOTA) models while using fewer parameters. Specifically, it achieves Dice scores of 0.9817 on BCS, 0.9262 on DSB, and 86.45 on NuInsSeg, surpassing SOTA models in most cases despite much smaller computational footprint.

Paper Nr: 432
Title:

Erosion–Dilation Trajectory Radiomics for Classifying 3D High-Resolution Breast Microcalcifications

Authors:

Redona Braihmetaj, Jan Cornelis and Bart Jansen

Abstract: Background: Breast microcalcifications (MCs) are key early indicators of breast cancer. In mammography, radiomics features are commonly computed once on 2D MC clusters and at a single segmentation scale. This prevents testing how the discriminative signal for benign/malignant MC classification changes from the densely calcified core to progressively larger surrounding regions, or whether it is best captured by how features evolve as the analyzed MCs are systematically expanded. Using high-resolution 3D ex-vivo micro-CT, we quantify benign/malignant discrimination along a per-MC erosion–dilation trajectory. Materials and Methods: Biopsies from 94 patients were scanned with ex-vivo micro-CT and 3504 MCs were segmented. For each MC, we generated 11 segmentation masks by applying erosions Ek and dilations Dk (k ∈ 2,4,6,8,10) around the baseline mask M0; Ek/Dk denote erosion/dilation stage k. Analyses were restricted to MCs with complete trajectories (i.e., not vanishing under erosion), yielding 84 patients and 608 MCs. Radiomic features were extracted at each stage and evaluated with patient-level classification comparing: (i) separate models for each stage of the erosion-dilation trajectory t, (ii) a single and same 80-feature subset reused across all stages, and (iii) a trajectory model concatenating these 80 features across stages. Results: Dense cores (E10) retained diagnostic signal (AUC of 0.71) compared to M0, while performance increased towards intermediate dilations, peaking at D8 (AUC of 0.851). Reusing the same 80 features improved AUC at every stage, with the largest gain at E10 (AUC improvement 0.4). Using the 11 segmentations yields a trajectory model which further increased AUC to 0.865. Conclusions: Radiomics of individual MCs reveals extra information when computed on an erosion–dilation trajectory, rather than on a single segmentation mask.

Paper Nr: 437
Title:

Blood Typing with Computer Vision and Machine Learning: A Preliminary Approach

Authors:

Bruno Silva, Enmanuel Abilheira, Ljiljana Dukanovic, Afonso Pinheiro and Vítor Carvalho

Abstract: Blood typing is a critical process in medical transfusions requiring accurate and efficient classification methods. Traditional serological techniques, while reliable, face challenges including infrastructure requirements, time constraints, and human error susceptibility and existing automated blood-typing approaches are often limited by high computational requirements and lack validation for deployment in low-resource point-of-care (POC) settings. This study developed an intelligent blood typing system using artificial intelligence to analyse hemagglutination reactions, aiming to improve efficiency and accessibility while reducing dependency on specialized personnel. The research employed multiple machine learning approaches on a dataset of 3,090 pre-labelled blood typing reaction images from the CRIAM device. Methods included fine-tuned Convolutional Neural Networks (CNNs), traditional machine learning classifiers, namely Logistic Regression, Extreme Gradient Boosting (XGBoost) and Support Vector Machine (SVM), and hybrid approaches combining Vision Transformer (ViT) or CNN embeddings with traditional classifiers. Data preprocessing, augmentation, and stratified K-fold cross-validation were implemented to ensure robust model comparison and performance. The best results were achieved by combining DINOv2 embeddings with Logistic Regression, with 99.87±0.12% F1-Score. This approach required minimal computational resources to run the model (1GB RAM) to enable deployment on low-power, battery-operated devices. The complete dataset embeddings occupied only 4.6MB storage, allowing continuous model improvement through incremental learning. Limitations of this study include the use of data collected from a single device, incorporating multiple devices could improve model generalization, also, the research is lacking real world, in field, validation. The model successfully addresses blood type classification, avoiding the need for specialized technicians and providing rapid results without extensive laboratory infrastructure.

Paper Nr: 438
Title:

AI-Based Classification of Rapid Plasma Reagin Test Reactions for Syphilis Detection: A Preliminary Approach

Authors:

Enmanuel Abilheira, Bruno Silva, Ljiljana Dukanovic, Afonso Pinheiro and Vítor Carvalho

Abstract: The Rapid Plasma Reagin (RPR) test used for syphilis screening remains error-prone due to the subjective interpretation of its flocculation reactions and manual handling of reagents. We present a high-accuracy, low-latency Artificial Intelligence (AI)-based system for automated RPR classification on edge devices. We tested Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs), trained on a curated dataset of 243 RPR reactions images. After data augmentation, the dataset size was increased to 174,960 images. The best model achieved a 99.75% accuracy on the test set. The model runs in under 50 ms per images on a virtual resource-constrained hardware. This work demonstrates robust performance and real-world utility.

Paper Nr: 105
Title:

Efficient Diabetic Foot Ulcer Classification: A Comparative Study of Derivative-Free Methods for Resource-Constrained Clinical Environments

Authors:

Mohammed A. B. Mahmoud

Abstract: Diabetes affects over 820 million people globally, as reported by the World Health Organization. A significant complication of diabetes is diabetic foot ulcers (DFUs), which necessitate early detection to prevent severe outcomes, including lower limb amputation. This study proposes a method comparing the performance of gradient-based and non-gradient-based approaches for DFU analysis. Specifically, we evaluate a gradient-based vision transformer (ViT) against two non-gradient methods: the Cascaded Forward (CaFo) algorithm and the Pseudoinverse Learning autoencoder (PILAE). ViT, a backbone of many recent large language models (LLMs), relies on the backpropagation (BP) algorithm during training. However, BP is associated with several limitations, such as vanishing/exploding gradients, overfitting, local minima, architectural constraints, and high computational demands. With an emphasis on clinical deployability, this work offers the first thorough comparison of gradient-free techniques (PILAE and CaFo) versus a cutting-edge gradient-based Vision Transformer for DFU classification. Our findings show that PILAE offers a workable, affordable alternative for early DFU identification in resource-constrained contexts by achieving competitive diagnostic performance with noticeably decreased computing overhead.

Paper Nr: 139
Title:

Using Multi-Task Learning and CAM-Based Ensembles for Detection of Abnormalities in X-Rays

Authors:

Ahana Roy Choudhury, Rebanto Nath and Akshat Sharma

Abstract: Musculoskeletal disorders are one of the leading causes of disability worldwide. Accurate and timely diagnosis can play a significant role in reducing pain, mobility issues and in preventing disability. We aim to develop an abnormality detection framework that utilizes X-rays and ensures interpretability by visualizing the Class Activation Maps (CAMs). We use three classic CNN architectures, DenseNet169, ResNet50, and VGG19 to create an ensemble of 3 Convolutional Neural Networks (CNNs) and perform our experiments on the MURA Dataset. We introduce multi-task learning to enable the backbones shared between the tasks to learn additional, useful information. Besides, we explore CAM-based ensemble techniques for combining the outputs of the models. CAM-based ensemble strategies allow us to make decisions based on the CAMs produced by a model for a specific image and do not rely on computations or decisions based on the validation set, which may not generalize adequately to the test set. On the validation set, our Cohen’s Kappa score improves by 2.7% to 6% for the three models by introducing multi-task learning. Our CAM-based technique for combining the outputs of the models results in an improvement of 4.6% over the best performing individual model.

Paper Nr: 244
Title:

Multi-Backend Deep Learning Framework for Multimodal Microscopy

Authors:

Inés Varona-Peña, Sara Cruz-Adrados, Rosa-María Menchón-Lara and Biagio Mandracchia

Abstract: We present bioMONAI, an open-source deep learning framework designed to support multimodal microscopy workflows by extending the MONAI ecosystem with microscopy-oriented tools. bioMONAI provides a modular and reproducible infrastructure for integrating heterogeneous microscopy data, including electron and light microscopy, within unified deep learning pipelines. The framework offers native support for common microscopy data formats, domain-aware preprocessing transforms, specialized loss functions, and visualization utilities, while maintaining full interoperability with established deep learning backends through Keras 3 integration. Rather than proposing novel learning algorithms, bioMONAI focuses on reducing implementation complexity, improving workflow transparency, and facilitating reproducible experimentation in advanced microscopy settings. A practical workflow example using multispectral fluorescence microscopy illustrates how bioMONAI enables end-to-end pipeline construction with minimal code. The framework is publicly available to support further methodological validation and community-driven development. Availability: available at https://github.com/deepCLEM/biomonai.

Paper Nr: 286
Title:

Robust Segmentation of the Lungs Using Active Shape Models Combined with Neural Networks on PA Chest X-ray Images

Authors:

Adam Tumay, Daniel Hadhazi and Gabor Hullam

Abstract: The segmentation of the lung shadow on Posterior Anterior (PA) chest X-ray (CXR) images is often an important preprocessing step for CADe/CADx methods as it is required for numerous tasks (e.g. nodule segmentation, pulmonary lesion detection etc.) and PA CXR is one of the cheapest and most commonly used modalities worldwide. Neural network models trained for lung segmentation are able to achieve state-of-the-art accuracy, however they lack robustness often introducing non-local errors, furthermore are sensitive to even slight changes in the input distributions - unable to generalize well in inference time, since limited training data is available. In this paper based on our previous work on heart segmentation, we propose a novel method for segmenting the lung shadow by combining Convolutional Neural Networks (CNNs) with Active Shape Models (ASMs), where the contour is iteratively refined by neural network’s predictions regularized by the shape model. Results show that our approach is able to robustly segment the lung shadow even in cases where pure neural solutions failed to generalize well.

Paper Nr: 301
Title:

A Hybrid Dimensionality Reduction and Clustering Framework for Early Parkinson’s Disease Characterization Using Clinical and DaTSCAN Features

Authors:

Malak Elghali and Wafa Moualhi

Abstract: Parkinson’s disease (PD) is a complex neurodegenerative disorder characterized by the progressive loss of dopaminergic neurons. Early diagnosis remains a major challenge, particularly in SWEDD (Scans Without Evidence of Dopaminergic Deficit) cases, where patients exhibit Parkinson-like symptoms despite normal DaTSCAN imaging. This paper proposes a hybrid unsupervised framework that combines dimensionality reduction and clustering for early PD characterization. We evaluate Linear Discriminant Analysis (LDA), Uniform Manifold Approximation and Projection (UMAP), and their sequential integration (UMAP+LDA) using multimodal clinical and imaging features from the PPMI dataset (548 subjects: 341 PD, 156 HC, and 51 SWEDD). Three clustering algorithms K-Means, Hierarchical Clustering (HCA), and Gaussian Mixture Models (GMM)—are assessed using Accuracy, Sensitivity, Specificity, and F1-score. The results demonstrate that UMAP and LDA+UMAP provide superior separability compared to LDA alone. Among the clustering methods, the UMAP+GMM configuration achieves the most balanced performance, with an Accuracy of 0.68, F1-score of 0.60, Sensitivity of 0.78, and Specificity of 0.43, enabling an interpretable discrimination between PD and HC subjects. Overall, the proposed framework offers a robust and clinically meaningful approach for exploratory analysis and early Parkinson’s disease characterization.