International Journal of Computer Assisted Radiology and Surgery最新文献

{"title":"Improving the fairness and robustness of deep learning classification models through bias mitigation and post hoc calibration.","authors":"Majid Ramedani, Jaya C Terli, Ole H Martensen, Tobias Strauß, Katharina Buerger, Peter Dechent, Emrah Düzel, Klaus Fliessbach, Maria Gemenetzi, Wenzel Glanz, Julian Hellmann-Regen, Stefan Hetzer, Enise I Incesoy, Daniel Janowitz, Marie T Kronmüller, Christoph Laske, Falk Lüsebrink, Robert Perneczky, Oliver Peters, Lukas Preis, Josef Priller, Boris-Stephan Rauchmann, Ayda Rostamzadeh, Anja Schneider, Björn H Schott, Sebastian Sodenkamp, Annika Spottke, Eike J Spruth, Sophia Stöcklein, Jens Wiltfang, Frank Jessen, Stefan J Teipel, Martin Dyrba","doi":"10.1007/s11548-026-03805-x","DOIUrl":"https://doi.org/10.1007/s11548-026-03805-x","url":null,"abstract":"<p><strong>Purpose: </strong>The goal of this study was to enhance the fairness and clinical applicability of deep learning models for classifying Alzheimer's disease from structural MRI scans. To achieve this, we addressed two key challenges: age-related bias and unreliable probability estimates. Specifically, we investigated in-processing debiasing strategies to achieve age-invariant predictions while preserving diagnostic accuracy. We also systematically evaluated post hoc calibration methods to improve the reliability, interpretability, and clinical usefulness of predicted probabilities across diverse patient populations.</p><p><strong>Methods: </strong>Structural MRI data from three independent cohorts (ADNI, AIBL, and DELCODE) were preprocessed using a standardized pipeline. Age-related bias was addressed using two in-processing strategies, adversarial training (CF-Net) and GLM-based feature residualization (PMDN), implemented in a shared 3D convolutional neural network (CNN) framework. Additionally, post hoc calibration was performed using four regression methods: logistic, probit, isotonic, and Bayesian binning. We evaluated model performance, fairness, and calibration using classification as well as calibration-specific metrics.</p><p><strong>Results: </strong>PMDN achieved the best overall performance, with higher balanced accuracy (0.71) and improved calibration (Brier <math><mo>≈</mo></math> 0.22), showing substantially reduced variability compared to baseline and CF-Net. PMDN effectively removed age-related bias, achieving near-zero age prediction correlation; CF-Net, however, showed only partial and unstable debiasing. Post hoc calibration further improved probability reliability, with nonparametric methods (isotonic regression and Bayesian binning) yielding the lowest calibration errors (MCE <math><mo>≈</mo></math> 0.21-0.24).</p><p><strong>Conclusion: </strong>This study shows that, although deep learning models achieve similar classification performance, they differ significantly in terms of calibration, stability, and fairness. PMDN provides more consistent and effective age bias mitigation than the CF-Net approach, particularly in clinically relevant subgroups. Post hoc calibration improves probability reliability further, with nonparametric methods achieving the best calibration.</p>","PeriodicalId":51251,"journal":{"name":"International Journal of Computer Assisted Radiology and Surgery","volume":" ","pages":""},"PeriodicalIF":2.8,"publicationDate":"2026-10-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149217312","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Learning from the unseen: synthetic data for scoliosis AI segmentation.","authors":"Volodimir Buchakchiyskiy, Sidaty El Hadramy, Norbert Zentai, Michael Wehrli, Zdzislaw Krol, Giada Cescato, Maria Licci, Daniel Studer, Carol C Hasler, Philippe C Cattin","doi":"10.1007/s11548-026-03782-1","DOIUrl":"https://doi.org/10.1007/s11548-026-03782-1","url":null,"abstract":"<p><strong>Purpose: </strong>Severe scoliosis remains challenging for 3D vertebral segmentation because annotated pathological computed tomography (CT) data are scarce.</p><p><strong>Methods: </strong>We synthesize scoliosis-like CT volumes by jointly warping public healthy spine scans and labels and then train nnU-Net CNN and PRIMUS models under matched compute. We evaluate on in-house out-of-distribution (OOD) adolescent scoliosis CTs and the public Spine1K dataset.</p><p><strong>Results: </strong>On 18 in-house adolescent scoliosis CTs, the best CNN setting increased mean Dice from 0.765 to 0.842 and vertebra success rate from 73.2 to 81.6%. In 55 surgeon-defined apical labels, mean Dice increased from 0.720 to 0.854. Spine1K performance remained high when synthetic data supplemented real scans; synthetic-only training was less stable.</p><p><strong>Conclusion: </strong>Pathology-informed synthetic data improve robustness to severe scoliosis without new annotations or architecture changes. Code and interactive visualizations are available at: https://chepchikvaleri.github.io/synth-data-s/.</p>","PeriodicalId":51251,"journal":{"name":"International Journal of Computer Assisted Radiology and Surgery","volume":" ","pages":""},"PeriodicalIF":2.8,"publicationDate":"2026-10-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149171963","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Resection zone prediction for parenchyma-sparing hepatectomy planning: a comparative study of three modeling paradigms.","authors":"Janine Rothert, Joy Rakshit, Judith L Salz, Florentine Huettl, Viola Ehses, Tobias Huber, Hauke Lang, Sylvia Saalfeld, Georg Hille","doi":"10.1007/s11548-026-03798-7","DOIUrl":"https://doi.org/10.1007/s11548-026-03798-7","url":null,"abstract":"<p><strong>Purpose: </strong>Parenchyma-sparing hepatectomy planning depends on accurate resection zones that preserve functional liver tissue without compromising oncological margins. This work investigates how different levels of anatomical and functional information complexity influence resection zone prediction for parenchyma-sparing surgical planning for primary liver cancer.</p><p><strong>Methods: </strong>We compare three modeling paradigms: a geometric distance-based approach, an explicit perfusion-based method using vascular anatomy, and a deep learning-based model built on the U-Net architecture. All methods operate on segmentation-derived representations of the liver, tumor, and vessels. Performance is evaluated using overlap- and distance-based metrics.</p><p><strong>Results: </strong>The distance-based model produces predictions with limited surface deviation (HD <math><mmultiscripts><mrow></mrow> <mn>95</mn> <mrow></mrow></mmultiscripts> </math> 33.89  mm) but lower overlap due to undersegmentation (DSC 58.18 %). The perfusion-based method achieves a favorable balance between overlap (DSC 67.41 %) and boundary accuracy (HD <math><mmultiscripts><mrow></mrow> <mn>95</mn> <mrow></mrow></mmultiscripts> </math> 37.92  mm) but tends to overestimate the predicted region due to strict binary perfusion assumptions. The deep learning model attains the highest overlap accuracy (DSC 76.31 %) while exhibiting larger distance errors (HD <math><mmultiscripts><mrow></mrow> <mn>95</mn> <mrow></mrow></mmultiscripts> </math> 65.21  mm), reflecting localized boundary inaccuracies.</p><p><strong>Conclusion: </strong>None of the models is universally outperforming the other two for parenchyma-sparing resection planning. Deep learning shows strong predictive performance, particularly for larger resection volumes, while geometric and perfusion-based models offer greater interpretability and clinical controllability. The distance-based approach is well suited for maximal parenchyma-sparing resections, whereas perfusion-based modeling is advantageous for tumors near vessels with potential perfusion loss. Our results highlight variations due to different surgical strategies and can thus provide valuable guidance for individual patient's surgery planning.</p>","PeriodicalId":51251,"journal":{"name":"International Journal of Computer Assisted Radiology and Surgery","volume":" ","pages":""},"PeriodicalIF":2.8,"publicationDate":"2026-10-07","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149129200","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"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 support system for total hip arthroplasty: evaluation of lower limb loads using virtual and dummy human body models.","authors":"Risa Shoji, Motoki Takagi, Tomonori Baba, Akihiko Hanafusa","doi":"10.1007/s11548-026-03791-0","DOIUrl":"https://doi.org/10.1007/s11548-026-03791-0","url":null,"abstract":"<p><strong>Purpose: </strong>The direct anterior approach (DAA) is one of the procedures used for total hip arthroplasty (THA) and offers advantages such as faster postoperative recovery. A lower limb traction operating table has been developed to assist in intraoperative movement and retention of the lower limb. However, these movements depend on the experience of individual surgeons, and no quantitative safety index is currently available. The authors are developing a surgical support system for THA that evaluates lower limb loads to achieve safe and reliable surgery. This study aimed to establish a system using dummy and virtual human body models that can quantitatively evaluate lower limb loads during procedures.</p><p><strong>Methods: </strong>The dummy human body model consisted of a skeletal specimen with six silicone tubes mimicking four types of muscles. Tension sensors were installed to measure tension forces. A virtual human body model with 40 muscles, along with the operating table model, was constructed using the musculoskeletal simulation software OpenSim. The movements of the operating table and the dummy human body model were measured using a 3D motion analysis system, and identical movements were applied for simulation.</p><p><strong>Results: </strong>Traction and extension movements showed that the trend of reaction forces in the traction direction and the relative magnitudes of muscle tension forces in the dummy model were consistent with the simulation results. In addition, the simulation enabled evaluation of muscle forces not physically represented in the dummy model.</p><p><strong>Conclusion: </strong>The validity of the simulation was verified by comparing measured and simulated forces applied to both the virtual and dummy human body models. The combined use of these models enabled evaluation of loads on the sole and muscles surrounding the hip joint. This approach enables quantitative and comprehensive biomechanical evaluation of intraoperative procedures and contributes to the development of a safe surgical support system.</p>","PeriodicalId":51251,"journal":{"name":"International Journal of Computer Assisted Radiology and Surgery","volume":" ","pages":""},"PeriodicalIF":2.8,"publicationDate":"2026-10-05","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149118255","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Evaluation of an automated workflow for TSN-based bounded-latency communication in ISO IEEE 11073 SDC-based medical networks.","authors":"Maja Dohms, Noah Wickel, Klaus Radermacher, Armin Janß","doi":"10.1007/s11548-026-03801-1","DOIUrl":"https://doi.org/10.1007/s11548-026-03801-1","url":null,"abstract":"<p><strong>Purpose: </strong>Interoperable, vendor-independent communication between medical devices in operating rooms has been enabled by the ISO IEEE 11073 Service-oriented Device Connectivity (SDC) standard since 2019. However, mechanisms for predictable and deterministic transmission behavior have not yet been standardized. One approach to enable bounded-latency communication is the use of Time-Sensitive Networking (TSN), an extension of standard Ethernet that enables time-synchronized and controlled data transmission to achieve reliable and time-deterministic data delivery.</p><p><strong>Methods: </strong>We implemented and evaluated the integration of TSN into SDC-based medical networks to provide bounded-latency communication through time-division mechanisms. An automated process was developed to extract transmission requirements from SDC devices and generate TSN configuration files for network deployment. The evaluation was conducted in a prototype SDC network under realistic traffic conditions. Furthermore, an SDC library designed with the aim of enabling faster data transmission via QUIC was used in the evaluation.</p><p><strong>Results: </strong>Automated TSN configuration is technically feasible and can improve bounded-latency communication in busy networks. With appropriately dimensioned schedules, the maximum SDC latency remains within the same range under high background traffic as under unloaded conditions. However, improperly dimensioned transmission windows can cause instability, connection loss and latency peaks of up to four times the planned transmission time.</p><p><strong>Conclusion: </strong>SDC with TSN offers a promising approach for achieving bounded-latency communication in interoperable medical networks. Further optimization is required to eliminate latency peaks and meet the timing requirements of safety-critical medical applications.</p>","PeriodicalId":51251,"journal":{"name":"International Journal of Computer Assisted Radiology and Surgery","volume":" ","pages":""},"PeriodicalIF":2.8,"publicationDate":"2026-09-29","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149082724","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Evaluation of 5G cellular and Wi-Fi 6 for cable-free surgical navigation using ISO/IEEE 11073 SDC.","authors":"Noah Wickel, Philipp Schollmaier, Klaus Radermacher, Armin Janß","doi":"10.1007/s11548-026-03774-1","DOIUrl":"https://doi.org/10.1007/s11548-026-03774-1","url":null,"abstract":"<p><strong>Purpose: </strong>As clinics strive for optimal patient safety and surgical outcomes as well as efficient and cost-effective operation, operating rooms have been investigated for workflow disruptions, risks, and inefficiencies. Notably, medical device cabling and tubing has been identified to cause surgical staff to trip and fall or to take detours to avoid them. Wireless devices could alleviate cable-associated risks, but few of them exist in practice as it remains a technical challenge to provide wireless connectivity with the required responsiveness and reliability for demanding applications, such as surgical navigation.</p><p><strong>Methods: </strong>Due to their promise of high reliability and low latency, private 5G networks have been proposed as enablers for wireless medical applications. In this work, the prototype of an interoperable tracking provider employing 5G-SA (stand-alone) and sixth-generation WLAN communications is evaluated regarding latency and reliability.</p><p><strong>Results: </strong>The performance using Wi-Fi is improved compared to the first-generation prototype, consistently reporting tracking events below 30/100 ms delay in the average/worst case, respectively. Using 5G-SA, the system can only operate at lower update rates, which causes outliers above 250 ms and experiences congestion as well as connection losses.</p><p><strong>Conclusion: </strong>The latency results using Wi-Fi are improved compared to the previous generation and generally enable navigation as well as other hand-eye coordination-based tasks. However, the tested 5G system does not yet provide the necessary performance required for the surgical navigation task and requires further adjustments of the hardware and software.</p>","PeriodicalId":51251,"journal":{"name":"International Journal of Computer Assisted Radiology and Surgery","volume":" ","pages":""},"PeriodicalIF":2.8,"publicationDate":"2026-09-28","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149076489","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Assessing selective interaction for reliable and trustworthy robotic surgical assistance.","authors":"Jacinto Colan, Ana Davila, Yutaro Yamada, Kazunari Misawa, Yasuhisa Hasegawa","doi":"10.1007/s11548-026-03799-6","DOIUrl":"https://doi.org/10.1007/s11548-026-03799-6","url":null,"abstract":"<p><strong>Purpose: </strong>Achieving reliable robotic assistance in collaborative surgery requires balancing the precision of autonomy with human oversight. This study assesses a Selective interaction strategy, in which the robot initiates communication only when task confidence is low, to explore whether uncertainty-triggered interaction may support reliability and operator trust relative to conventional Autonomous and Supervised modalities.</p><p><strong>Methods: </strong>We implemented a collaborative framework for soft tissue resection capable of three interaction levels: (1) Autonomous, executing tasks without feedback; (2) Supervised, requiring confirmation for every critical action; and (3) Selective, which triggers queries based on internal uncertainty thresholds. We conducted a preliminary user study on resection of phantom tissue, assessing quantitative reliability (cutting path error, task completion time) and qualitative metrics (NASA-TLX, Trust/Distrust questionnaires).</p><p><strong>Results: </strong>Results suggest that the Selective modality optimizes the performance-workload trade-off. It showed a trend toward lower mean path error and shorter mean completion time compared to the Supervised modality. It also demonstrated lower mean distrust and mental demand scores alongside similar positive trust scores.</p><p><strong>Conclusion: </strong>The preliminary data suggest that uncertainty-triggered communication can provide a favorable balance between task performance, workload, and operator perception. A larger study is required to determine whether the observed trends represent statistically reliable effects.</p>","PeriodicalId":51251,"journal":{"name":"International Journal of Computer Assisted Radiology and Surgery","volume":" ","pages":""},"PeriodicalIF":2.8,"publicationDate":"2026-09-28","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149076498","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Introducing FRAM as a tool for surgical process modeling: a case study for understanding surgical scheduling complexity.","authors":"Sidra Rashid, Rebecca Weber, Alexander Geiger, Lars Wagner, Lukas Bernhard, Emily Spicker, Alissa Jell, Johannes Fottner, Dirk Wilhelm, Niklas Grabbe","doi":"10.1007/s11548-026-03800-2","DOIUrl":"https://doi.org/10.1007/s11548-026-03800-2","url":null,"abstract":"<p><strong>Introduction: </strong>Surgical processes are inherently variable leading to unexpected delays and cancellations. Due to inefficient and outdated processes, there is a need to understand and improve potential bottlenecks. However, traditional surgical process models (SPMs) focus on intraoperative procedures and sequential workflows, but most cannot adequately capture parallel workflows, variability, and sociotechnical interactions. We propose the use of Functional Resonance Analysis Method (FRAM) in conjunction with FRAMalyse to enable systematic measurement of variability propagation and potential process bottleneck identification. Additionally, we introduce a structured FRAM Model Interpreter (FMI) cycle-mapping interpretation protocol that maps function activations across cycles to temporal process phases, an approach for interpreting FRAM simulation outputs that has not been used in prior qualitative applications of the method.</p><p><strong>Material and methods: </strong>Based on stakeholder interviews and workflow observations, we developed a FRAM model of the daily surgical scheduling process at University Hospital Technical University of Munich (TUM). FMI profiles for each function were determined based on expert interviews and function activations were recorded during model simulation. Additionally, the model was analyzed using FRAMalyse to identify functions with high variability potential arising from their coupling structure, thereby providing a basis for future bottleneck validation and surgical scheduling optimization.</p><p><strong>Results: </strong>We identified 98 interconnected FRAM functions across four agents and phases. Scheduling and rescheduling functions spanned multiple cycles, indicating continuous and reciprocal coordination. Surgical team allocation function showed the highest variability, while attending surgeons performed the most functions and central patient management (CPM) had the highest coupling density.</p><p><strong>Conclusion: </strong>This study introduces the first quantitative application of FRAM and FRAMalyse in the healthcare domain, systematically identifying candidate bottlenecks across the full surgical scheduling pathway as a foundation for process improvement.</p>","PeriodicalId":51251,"journal":{"name":"International Journal of Computer Assisted Radiology and Surgery","volume":" ","pages":""},"PeriodicalIF":2.8,"publicationDate":"2026-09-28","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149070523","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Deep learning-based segmentation of peritoneal cancer index regions from CT imaging.","authors":"Pieter C Gort, Lotte J S Fleurkens-Ewals, Lenah D Kampmeijer, Anna F van Herwijnen, Marion W Tops-Welten, Cris H B Claessens, Joost Nederend, Ignace H J T De Hingh, Max J Lahaye, Misha D P Luyer, Fons van der Sommen","doi":"10.1007/s11548-026-03796-9","DOIUrl":"https://doi.org/10.1007/s11548-026-03796-9","url":null,"abstract":"<p><strong>Purpose: </strong>Peritoneal metastases (PM) are staged using the surgically determined peritoneal cancer index (PCI), which requires invasive laparoscopic assessment. Although CT is routinely used for preoperative evaluation, imaging-based assessment of PM extent remains challenging and is often less structured than surgical PCI scoring. A recent consensus study defined radiological PCI (rPCI) regions for cross-sectional imaging. We present the first deep learning approach to automatically segment 13 rPCI regions on CT.</p><p><strong>Methods: </strong>62 contrast-enhanced CT scans were retrospectively collected across the full PCI range. Each scan was annotated into non-overlapping rPCI regions by one researcher, reviewed by a second, with disagreements resolved by a radiologist. Using fivefold cross-validation, we compared nnU-Net and Swin UNETR with Dice, 95th-percentile Hausdorff distance (HD95) and Average Surface Distance (ASD). We introduce an anatomically constrained pipeline that trains on merged super-regions and splits them during post-processing using TotalSegmentator landmarks at the hips and the ligament of Treitz.</p><p><strong>Results: </strong>On the identical 62-scan cohort, the baseline nnU-Net reached an overall Dice of 0.81 and outperformed Swin UNETR (0.76). The proposed anatomically constrained pipeline improved the overall Dice to 0.84 and reduced boundary error (HD95 <math><mrow><mn>13.7</mn> <mo>→</mo> <mn>11.8</mn></mrow> </math>  mm; ASD <math><mrow><mn>4.1</mn> <mo>→</mo> <mn>3.4</mn></mrow> </math>  mm), with the largest gains in the small-bowel regions, approaching the interobserver Dice of 0.87.</p><p><strong>Conclusions: </strong>Automated rPCI region segmentation on CT is feasible and approaches interobserver agreement. Encoding anatomical boundary constraints substantially improves segmentation quality in the most challenging regions. This provides a reproducible foundation for noninvasive, imaging-based PCI assessment. The main limitations are the single-center cohort and the small interobserver subset. Code is available at: https://github.com/PieterGort/rpci-region-segmentation.</p>","PeriodicalId":51251,"journal":{"name":"International Journal of Computer Assisted Radiology and Surgery","volume":" ","pages":""},"PeriodicalIF":2.8,"publicationDate":"2026-09-27","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149070531","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Performance changes in automated lesion detection under federated learning with sequential institution addition.","authors":"Yukihiro Nomura, Shouhei Hanaoka, Aiki Yamada, Tomomi Takenaga, Takahiro Nakao, Toshiya Nakaguchi, Takeharu Yoshikawa, Osamu Abe","doi":"10.1007/s11548-026-03795-w","DOIUrl":"https://doi.org/10.1007/s11548-026-03795-w","url":null,"abstract":"<p><strong>Purpose: </strong>Federated learning (FL) enables multiple institutions to collaboratively train machine learning models while keeping data local and has attracted attention in medical image processing, including computer-aided detection (CAD). In FL, performance is expected to improve through retraining as additional institutions participate. The purpose of this study was to investigate how CAD software performance changes as the number of participating institutions is sequentially increased within an FL framework.</p><p><strong>Methods: </strong>We used two types of CAD software for cerebral aneurysm detection in magnetic resonance (MR) angiography images and brain metastasis detection in contrast-enhanced T1-weighted MR images. Datasets from different institutions or scanner vendors were treated as independent FL clients and incorporated sequentially. Training strategies included from-scratch training, fine-tuning (FT) of selected layers, and full fine-tuning (FFT) of all parameters. Performance was assessed using the competition performance metric on test sets from the initial institutions as well as from all participating institutions.</p><p><strong>Results: </strong>For both CAD software types, sequential institution addition combined with FT generally showed higher median performance changes than from-scratch training. FT showed performance comparable to that of FFT while requiring substantially fewer trainable parameters. Performance improvements generally accumulated with sequential institution addition, whereas simultaneous addition resulted in less consistent improvements.</p><p><strong>Conclusions: </strong>Sequential institution addition under FL may improve CAD software performance when combined with appropriate FT strategies. FT that updates only a subset of layers may achieve performance changes comparable to those of FFT while requiring substantially fewer trainable parameters across different lesion detection tasks.</p>","PeriodicalId":51251,"journal":{"name":"International Journal of Computer Assisted Radiology and Surgery","volume":" ","pages":""},"PeriodicalIF":2.8,"publicationDate":"2026-09-27","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149070584","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}