Computer methods and programs in biomedicine最新文献

{"title":"The role of plaque morphology and composition in vulnerability assessment: Computational analysis using CT images and elastography","authors":"Nicoletta Curcio ,&nbsp;Giulia Matrone ,&nbsp;Michele Conti ,&nbsp;Giovanni Nano ,&nbsp;Paolo Righini ,&nbsp;Vlasta Bari ,&nbsp;Daniela Mazzaccaro","doi":"10.1016/j.cmpb.2025.109229","DOIUrl":"10.1016/j.cmpb.2025.109229","url":null,"abstract":"<div><h3>Objective</h3><div>This study seeks to assess the influence of using patient-specific data from different imaging methods on evaluating carotid plaque vulnerability via finite element analysis (FEA) instead of using data derived from the literature.</div></div><div><h3>Methods</h3><div>54 patients were considered in this analysis, who preoperatively underwent computed tomography angiography (CTA) and ultrasound (US) imaging evaluations. The composition (i.e. calcific, lipidic and mixed) and vulnerability (i.e. stable or vulnerable) of their plaques were evaluated by macroscopic and histologic assessment post-endarterectomy. In particular, the plaques of these 54 patients were classified as mixed. 3D reconstructions of the carotid artery were generated from CTA scans, and computational analyses were performed using two different simulation settings for material properties and loads: a) the material properties of the plaque components were set as an average of values available in the literature (LIT-based); b) the material property of the plaque fibrous content was modified using stiffness data derived from US shear-wave elastography imaging (SWE-based). Statistical analyses were conducted to compare stress parameters obtained from the different simulations within groups of vulnerable and stable plaques.</div></div><div><h3>Results</h3><div>Comparisons between LIT-based and SWE-based FEA revealed notable differences in stress parameters associated with plaque vulnerability. In particular, the stress values derived from SWE-based simulations provided distinct stratification of vulnerable versus stable plaques, whereas LIT-based models showed limited differentiation. Significant variations in von Mises (<em>p</em> = 0.015, <em>p</em> = 0.037) and maximum principal stress (<em>p</em> = 0.014) distributions were observed in SWE-based FEA.</div></div><div><h3>Conclusions</h3><div>Patient-specific modelling and computational analysis integrating CTA-derived morphological with US-derived biomechanical data could improve the assessment of plaque vulnerability in mixed-composition carotid plaques.</div></div>","PeriodicalId":10624,"journal":{"name":"Computer methods and programs in biomedicine","volume":"276 ","pages":"Article 109229"},"PeriodicalIF":4.8,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145881251","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Automatic construction of interconnected cable models of cardiac propagation on a surface","authors":"Elham Zakeri Zafarghandi,&nbsp;Vincent Jacquemet","doi":"10.1016/j.cmpb.2025.109228","DOIUrl":"10.1016/j.cmpb.2025.109228","url":null,"abstract":"<div><h3>Background and objective:</h3><div>Cardiac fibers may be represented by a network of interconnected cables for simulating electrical propagation. The lack of automatic cable mesh generation tool has hampered this modeling approach. We aim to provide and evaluate an algorithmic solution to this problem.</div></div><div><h3>Methods:</h3><div>We developed an open-source C++/Python package for the construction of a monolayer interconnected cable model from a triangulated surface with fiber orientation, targeting a given longitudinal and transverse space step. The workflow of the algorithm starts with the generation of evenly spaced streamlines aligned with fiber orientation. Another set of streamlines, orthogonal to the fibers, is used to specify lateral connections. The intersection between the two sets of streamlines gives the vertices of the cable mesh, determines its connectivity, and defines a polygonal tessellation of the surface that can be triangulated. Finite differences can then be applied to solve a reaction–diffusion equation on the cable mesh.</div></div><div><h3>Results:</h3><div>The approach was validated in increasingly complex configurations and up to near-cellular resolutions (20 to <span><math><mrow><mn>200</mn><mspace></mspace><mi>μ</mi><mi>m</mi></mrow></math></span>). Fiber orientation noise, singularities and abrupt changes in orientation reduced the local coupling by altering the microstructure of the tissue. The pipeline for mesh generation was tested using a publicly available cohort of 98 patient-specific geometries. The stability limit of the numerical scheme was assessed by spectral analysis of the diffusion matrix and was compared to triangular meshes and cartesian grids.</div></div><div><h3>Conclusion:</h3><div>This physiologically based mesh generation tool may be used as a building block for the construction of multilayer three-dimensional models of the atria for the simulation of discrete propagation.</div></div>","PeriodicalId":10624,"journal":{"name":"Computer methods and programs in biomedicine","volume":"276 ","pages":"Article 109228"},"PeriodicalIF":4.8,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145838858","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Biomarker discovery study design consistent with the receiver-operator characteristic","authors":"Joakim Ekström,&nbsp;Ivaylo Stoimenov,&nbsp;Jim Åkerrén Ögren,&nbsp;Tobias Sjöblom","doi":"10.1016/j.cmpb.2025.109215","DOIUrl":"10.1016/j.cmpb.2025.109215","url":null,"abstract":"<div><h3>Background and Objective</h3><div>The field of early biomarker discovery is characterized by a lack of consensus on the choice of statistical methodology, which may impede later progress towards clinically useful biomarkers. The Receiver-Operator Characteristic (ROC) is a <em>de facto</em> standard for determining the performance of In Vitro Diagnostic (IVD) devices. In this study, we aimed to systematically identify and mitigate prevalent pitfalls in biomarker discovery efforts and propose a best-practice guideline based on a ROC analysis framework.</div></div><div><h3>Methods</h3><div>By maintaining a careful alignment to the study objectives through a sample procurement plan, study size determination and data analysis by the ROC framework, we formulated a biomarker discovery protocol. We performed Monte Carlo simulations to inform the investigator on the suitable number of study participants, the statistical power and sample bin allocation strategy. The main concept is illustrated using proteomic data of newly diagnosed cancer cases and concurrent external controls.</div></div><div><h3>Results</h3><div>The work demonstrates a regulatory-adherent pipeline to achieve an effect superior to the current best biomarker used as a predicate medical device. In our proof-of-concept ROC-based analysis in samples from a publicly available dataset, we detected statistically significant composite biomarkers, of which we validated a subset in an independent dataset acquired using the same proteomic analysis method. Intriguingly, commonly used feature selection methods do not identify the same composite biomarkers from the same data, and their selections show limited overlap with the ROC-based analysis.</div></div><div><h3>Conclusion</h3><div>The proposed approach can facilitate translation of scientific discoveries into regulatory approved biomarker tests fit for use in clinical medicine.</div></div>","PeriodicalId":10624,"journal":{"name":"Computer methods and programs in biomedicine","volume":"276 ","pages":"Article 109215"},"PeriodicalIF":4.8,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145827149","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"INTELLI-PVA: Informative sample annotation-based contrastive active learning for cross-domain patient-ventilator asynchrony detection","authors":"Lingwei Zhang ,&nbsp;Xue Feng ,&nbsp;Fei Lu ,&nbsp;Zepeng Ding ,&nbsp;Jiayi Yang ,&nbsp;Luping Fang ,&nbsp;Gangmin Ning ,&nbsp;Shuohui Yuan ,&nbsp;Huiqing Ge ,&nbsp;Qing Pan","doi":"10.1016/j.cmpb.2025.109203","DOIUrl":"10.1016/j.cmpb.2025.109203","url":null,"abstract":"<div><h3>Background and objective</h3><div>Patient-ventilator asynchrony (PVA) is prevalent in mechanically ventilated patients and adversely impacts clinical outcomes, but its real-time detection remains challenging. While artificial intelligence (AI) systems show promise for PVA detection, their cross-domain generalization faces two major limitations: variability in patient-ventilator interactions across different clinical settings, and morphological overlap between PVA types. These challenges necessitate specialized AI solutions rather than conventional re-annotation approaches.</div></div><div><h3>Methods</h3><div>We present the INTELLI-PVA framework for efficient cross-domain PVA detection on eight types. First, a hybrid two-stage PVA classifier was developed. A deep learning model, pre-trained on unannotated data using contrastive learning and fine-tuned using annotated data, identified four morphologically defined compound PVA types, each encompassing a reverse triggering (RT) and a non-RT type. A subsequent rule-based algorithm differentiated the subtypes within each compound type according to their triggering signatures. Then, the model was adapted to the target domain through an iterative active learning cycle, which selected the most informative samples for expert annotation and used them to fine-tune the model.</div></div><div><h3>Results</h3><div>Established and validated on data from two centers encompassing 1190 patients and 124.975 million respiratory cycles, INTELLI-PVA demonstrates superior detection performance (average F1-score: 0.849) in classifying the eight PVA classes using only 1000 annotated samples per target domain, and achieves respiratory therapist-level recognition ability (average Cohen's κ=0.850) across unseen ventilator configurations and patient demographics.</div></div><div><h3>Conclusions</h3><div>INTELLI-PVA achieves high-accuracy, cross-domain PVA detection with minimal annotation burden, establishing a practical and efficient pathway for deploying AI-assisted ventilation monitoring in diverse clinical settings.</div></div>","PeriodicalId":10624,"journal":{"name":"Computer methods and programs in biomedicine","volume":"275 ","pages":"Article 109203"},"PeriodicalIF":4.8,"publicationDate":"2026-02-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145734016","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"CoroSAM: adaptation of the Segment Anything Model for interactive segmentation in Coronary angiograms","authors":"Michela Ferrari ,&nbsp;Mario Urtis ,&nbsp;Edoardo Spairani ,&nbsp;Antonio Tescari ,&nbsp;Francesca Sessa ,&nbsp;Maurizia Grasso ,&nbsp;Francesco Prati ,&nbsp;Eloisa Arbustini ,&nbsp;Giovanni Magenes","doi":"10.1016/j.cmpb.2025.109172","DOIUrl":"10.1016/j.cmpb.2025.109172","url":null,"abstract":"<div><h3>Background and objectives</h3><div>X-ray Coronary Angiography (XCA) enables visualization of coronary arteries for disease and morphology assessment. Accurate segmentation of major coronary vessels is crucial for automated analysis of their geometric features but presents challenges due to anatomical complexity. This study introduces CoroSAM, an adaptation of the LiteMedSAM Foundation Model, employing parameter-efficient fine-tuning for interactive coronary artery segmentation in XCA images.</div></div><div><h3>Methods</h3><div>The proposed approach incorporates Convolutional Adapter layers within the image encoder's TinyViT blocks to enhance domain-specific feature extraction while maintaining computational efficiency. A point-based prompting strategy directly encodes vessel endpoints and branch points as additional input channels. Model evaluation employed 5-fold cross-validation on the ARCADE dataset and zero-shot testing on external datasets (XCAD, DCA1). Performance was compared with state-of-the-art models for user-guided segmentation.</div></div><div><h3>Results</h3><div>CoroSAM demonstrated superior performance on the ARCADE test set (Dice=0.87, Precision=0.86, Recall=0.89) while requiring fewer trainable parameters compared to competitive models. Statistical analysis confirmed significant improvements over alternative Adapter configurations. Zero-shot generalization yielded competitive performance on external datasets (XCAD: Dice=0.82; DCA1: Dice=0.73), demonstrating robust transferability across different image qualities.</div></div><div><h3>Conclusions</h3><div>Integrating specialized Convolutional Adapters and channel-encoded point prompts enables accurate delineation of major coronary vessels with minimal user intervention. CoroSAM's architecture facilitates efficient inference on standard computing hardware without dedicated GPUs, providing a practical tool for clinical applications. This approach establishes an adaptation framework that effectively balances segmentation accuracy with computational efficiency, making it suitable for routine analysis workflows.</div></div>","PeriodicalId":10624,"journal":{"name":"Computer methods and programs in biomedicine","volume":"274 ","pages":"Article 109172"},"PeriodicalIF":4.8,"publicationDate":"2026-02-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145602812","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Development of a predictive model for Ki-67 index of meningiomas by integrating deep-learning, radiomics and clinical features utilizing fully automated segmentation results","authors":"Xin Ma ,&nbsp;Yajing Zhao ,&nbsp;Kaiyue Zhang ,&nbsp;Nan Mei ,&nbsp;Xuanxuan Li ,&nbsp;Jin Cui ,&nbsp;Jie Chen ,&nbsp;Yuxi Xie ,&nbsp;Yiping Lu ,&nbsp;Bo Yin","doi":"10.1016/j.cmpb.2025.109205","DOIUrl":"10.1016/j.cmpb.2025.109205","url":null,"abstract":"<div><h3>Purpose</h3><div>To investigate the efficacy of clinical information, traditional radiological, radiomics and deep-learning features combinations for constructing a predictive model for the Ki-67 index of meningiomas.</div></div><div><h3>Material and methods</h3><div>This study acquired retrospective (198 cases) and prospective (22 cases) meningioma data between 2015 and 2020. Within the retrospective data, 160 cases were utilized for training, while 38 were allocated to an independent test. Ki-67 expression levels were dichotomized into low and high groups using a 4% threshold based on previous research. The study developed and evaluated five classifier models combining clinical information, radiomics and deep-learning features to predict Ki-67 expression levels. Model performance was evaluated via the receiver operating characteristic (ROC) curves and the area under the curve (AUC), obtaining a 95% confidence interval (CI) using DeLong testing. Subsequently, the most effective model was validated using prospective data from 22 cases.</div></div><div><h3>Results</h3><div>The eXtreme Gradient Boosting (XGBoost) classifier model showed optimal performance among the five classifier models. The AUC for the independent test dataset was 0.717 (CI: 0.575-0.858). After optimization, the AUC of the test dataset is 0.767 (CI: 0.631-0.903). The AUC for the prospective test data set was 0.773 (CI: 0.590-0.955). Decision curve analysis (DCA) showed that combining clinical information, radiomics, and deep-learning features resulted in the best predictive performance of the XGBoost classifier.</div></div><div><h3>Conclusion</h3><div>An integrated radiomics model enables Ki-67 prediction and has great potential to estimate the risk of tumor regrowth and recurrence non-invasively.</div></div>","PeriodicalId":10624,"journal":{"name":"Computer methods and programs in biomedicine","volume":"275 ","pages":"Article 109205"},"PeriodicalIF":4.8,"publicationDate":"2026-02-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145787096","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"The added value of radiomic analysis for predicting spontaneous preterm birth in the first trimester","authors":"William Cancino ,&nbsp;Carlos Hernan Becerra-Mojica ,&nbsp;Said Pertuz","doi":"10.1016/j.cmpb.2025.109181","DOIUrl":"10.1016/j.cmpb.2025.109181","url":null,"abstract":"<div><h3>Background and Objective:</h3><div>Preterm birth (PTB) is a public health problem. Researchers have worked to identify ways to detect women at risk for PTB early in pregnancy. Although existing biomarkers allow some women to be detected in the second trimester, detection in the first trimester is warranted to initiate earlier interventions. This study aims to explore the added value of radiomic analysis of transvaginal ultrasound (TVUS) images in the construction of first-trimester risk assessment models for predicting spontaneous preterm birth (sPTB).</div></div><div><h3>Methods:</h3><div>A retrospective cohort study was conducted including pregnant women who attended their screening ultrasound examination between 11+0 and 13+6 weeks. Data on medical history (MH), cervical length (CL), and cervical consistency index (CCI) were collected, along with TVUS images. These images were subjected to a computerized radiomic analysis to extract features, which were used to train different machine learning models incorporating a feature selection process. The area under the receiver operating characteristic curve (AUC) with 95% confidence intervals (CI) was used to evaluate the models’ performance in predicting sPTB.</div></div><div><h3>Results:</h3><div>A total of 253 pregnant women were included in the study, where 225 had a term birth and 28 had a sPTB. In sPTB prediction, MH, CL, and CCI obtained AUCs of 0.68 (95% CI, 0.56–0.79), 0.61 (95% CI, 0.48–0.73) and 0.67 (95% CI, 0.55–0.79), respectively. Among the evaluated machine learning models, logistic regression achieved the highest AUC and was therefore used to build the radiomic feature–based model (RF), which reached an AUC of 0.67 (95% CI, 0.56–0.76). When combining RF with MH and CCI, AUCs of 0.73 (95% CI, 0.63–0.82) and 0.74 (95% CI, 0.64–0.83) were achieved, respectively. The best performance was obtained by combining RF with both MH and CCI, reaching an AUC of 0.76 (95% CI, 0.68–0.84).</div></div><div><h3>Conclusion:</h3><div>The performance of the models improved by adding radiomic features, highlighting the potential of radiomic analysis for sPTB prediction in the first trimester.</div></div>","PeriodicalId":10624,"journal":{"name":"Computer methods and programs in biomedicine","volume":"275 ","pages":"Article 109181"},"PeriodicalIF":4.8,"publicationDate":"2026-02-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145682518","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"MDM-DTA: Message Passing Neural Network with molecular descriptors and Mixture of Experts for drug–target affinity prediction","authors":"Yang Dai ,&nbsp;Xiaoyu Tan ,&nbsp;Haoyu Wang ,&nbsp;Gengchen Ma ,&nbsp;Yujie Xiong ,&nbsp;Xihe Qiu","doi":"10.1016/j.cmpb.2025.109163","DOIUrl":"10.1016/j.cmpb.2025.109163","url":null,"abstract":"<div><h3>Background and Objective:</h3><div>Drug–target affinity (DTA) prediction is a pivotal task in computational drug discovery, enabling the estimation of binding affinities between small molecules and their target proteins. This process is essential for reducing the costs, development time, and risks inherent in traditional drug development pipelines. Current DTA prediction models primarily rely on separate extraction and concatenation of drug and protein features. However, these models often fail to account for the complex semantic relationships within protein sequences, which limits their ability to accurately predict affinity.</div></div><div><h3>Methods:</h3><div>In response to these challenges, we propose MDM-DTA, a novel framework leveraging a Mixture of Experts (MoE) strategy to integrate diverse molecular and protein representations. For drug representation, MDM-DTA utilizes molecular graphs, which are processed via Message Passing Neural Networks (MPNNs), alongside molecular descriptors that are passed through a three-layer convolutional neural network (CNN). Protein features are extracted using a deep convolutional network enhanced with Squeeze-and-Excitation (SE) mechanisms to capture inter-channel dependencies. Furthermore, protein sequence semantics are encoded through pre-trained embeddings from a knowledge-guided Bidirectional Encoder Representations from Transformers (BERT) model and the Evolutionary Scale Modeling 2 (ESM2) model, enabling the model to capture contextual relationships within protein sequences.</div></div><div><h3>Results:</h3><div>Extensive experiments on three benchmark datasets demonstrate that MDM-DTA consistently outperforms state-of-the-art models of similar complexity in terms of predictive accuracy. The incorporation of both structural and semantic features significantly enhances the model’s ability to predict drug–target binding affinities, highlighting the importance of a multi-modal representation approach.</div></div><div><h3>Conclusions:</h3><div>The proposed MDM-DTA framework effectively integrates both molecular and semantic protein representations, providing superior performance in DTA prediction tasks. The results underscore the potential of MDM-DTA to improve the accuracy of computational drug discovery models, facilitating the identification of novel drug candidates and advancing the field of in silico drug development.</div></div>","PeriodicalId":10624,"journal":{"name":"Computer methods and programs in biomedicine","volume":"274 ","pages":"Article 109163"},"PeriodicalIF":4.8,"publicationDate":"2026-02-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145596058","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Benchmarking foundation models and parameter-efficient fine-tuning for prognosis prediction in medical imaging","authors":"Filippo Ruffini ,&nbsp;Elena Mulero Ayllón ,&nbsp;Linlin Shen ,&nbsp;Paolo Soda ,&nbsp;Valerio Guarrasi","doi":"10.1016/j.cmpb.2025.109196","DOIUrl":"10.1016/j.cmpb.2025.109196","url":null,"abstract":"<div><h3>Background and Objectives:</h3><div>Despite the significant potential of Foundation Models (FMs) in medical imaging, their application to prognosis prediction remains challenging due to data scarcity, class imbalance, and task complexity, limiting their clinical adoption. This study introduces the first structured benchmark to assess the robustness and efficiency of transfer learning strategies for FMs compared with convolutional neural networks (CNNs) in predicting COVID-19 patient outcomes from chest X-rays. The goal is to systematically compare fine-tuning strategies, classical and parameter-efficient, under realistic clinical constraints related to data scarcity and class imbalance, offering empirical guidance for AI deployment in clinical workflows.</div></div><div><h3>Methods:</h3><div>Four publicly available COVID-19 chest X-ray datasets were used, covering mortality, severity, and ICU admission, with varying sample sizes and class imbalances. CNNs pretrained on ImageNet and FMs pretrained on general or biomedical datasets were adapted using full fine-tuning, linear probing, and parameter-efficient methods. Models were evaluated under full-data and few-shot regimes using Matthews Correlation Coefficient (MCC) and Precision–Recall AUC (PR-AUC) with cross-validation and class-weighted losses.</div></div><div><h3>Results:</h3><div>CNNs with full fine-tuning performed robustly on small, imbalanced datasets, while FMs with Parameter-Efficient Fine-Tuning (PEFT), particularly LoRA and BitFit, achieved competitive results on larger datasets. Severe class imbalance degraded PEFT performance, whereas balanced data mitigated this effect. In few-shot settings, FMs showed limited generalization, with linear probing yielding the most stable results.</div></div><div><h3>Conclusions:</h3><div>No single fine-tuning strategy proved universally optimal. CNNs remain dependable for low-resource scenarios, whereas FMs benefit from parameter-efficient methods when data are sufficient.</div></div>","PeriodicalId":10624,"journal":{"name":"Computer methods and programs in biomedicine","volume":"275 ","pages":"Article 109196"},"PeriodicalIF":4.8,"publicationDate":"2026-02-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145676217","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Explainable multimodal fusion for breast carcinoma diagnosis: A systematic review, open problems, and future directions","authors":"Mohammad Mehedi Hassan ,&nbsp;Anika Tahsin ,&nbsp;Md Golam Rabiul Alam ,&nbsp;Deema Alzamil ,&nbsp;Sahil Garg ,&nbsp;Md. Zia Uddin ,&nbsp;Nurul Choudhury ,&nbsp;Giancarlo Fortino","doi":"10.1016/j.cmpb.2025.109152","DOIUrl":"10.1016/j.cmpb.2025.109152","url":null,"abstract":"<div><div>Breast carcinoma (BC) remains one of the most common and lethal malignancies in women worldwide, making an early and accurate diagnosis a public health priority. Recent artificial intelligence (AI) research has increasingly explored multimodal fusion, with imaging, clinical records, histopathology, and genomic data to produce richer, more reliable predictions. In parallel, explainable AI (XAI) techniques aim to address a key barrier to clinical use: transparency about how deep learning models make decisions. This systematic review examines 49 peer-reviewed studies published between 2015 and 2025, following the PRISMA guidelines to analyze the landscape of multimodal learning and XAI in the diagnosis and prognosis of BC. We categorize reported fusion strategies from simple feature concatenation to advanced attention-based, gated, and hybrid architectures, designed to manage data heterogeneity and missing modalities. We also document model designs, including transfer learning, transformers, graph neural networks (GNNs), autoencoders, and ensembles, supported by preprocessing methods like stain normalization and GAN-based augmentation, and the use of XAI techniques: Grad-CAM, SHAP, and attention weights, which help bridge the gap between complex AI systems and clinical workflows. Across included studies, multimodal models often outperformed unimodal baselines; however, effect sizes varied by dataset, validation design (cross-validation versus external), and the handling of missing modalities. Our review highlights persistent open problems, including the limited availability of multimodal datasets, inconsistent benchmarks, and the scarcity of interpretable models in real world settings. Future research on BC care should focus on developing ensemble-based fusion approaches, validating them in various clinics, and embedding clinician expertise. All of which are crucial for developing AI systems that are accurate, transparent, generalizable, and trustworthy.</div></div>","PeriodicalId":10624,"journal":{"name":"Computer methods and programs in biomedicine","volume":"274 ","pages":"Article 109152"},"PeriodicalIF":4.8,"publicationDate":"2026-02-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145511891","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}