Journal of imaging informatics in medicine最新文献

{"title":"Operational Ontology for Radiology: A Framework for Profiling and Characterizing CT Imaging Data.","authors":"Njood Alsaihati, Yunrong Cai, Jered Wells, Justin Solomon, Donald P Frush, Ehsan Samei","doi":"10.1007/s10278-026-02370-8","DOIUrl":"https://doi.org/10.1007/s10278-026-02370-8","url":null,"abstract":"<p><p>Medical imaging data are increasingly used for benchmarking, quality improvement, research, and artificial intelligence applications. However, the lack of consistent metadata characterization limits the reliability and scalability of such analyses. Existing standards provide syntactic interoperability but often fail to capture the semantic clinical context and workflow details necessary for secondary uses such as AI development and radiation safety assessment, affecting the utility for reproducible profiling of imaging examinations across institutions. In this work, we describe key gaps in current approaches to imaging data standardization. These include syntactic interoperability without semantic equivalence, limited representation of clinically relevant context, inconsistent definitions of radiology procedures across institutions, and the absence of mechanisms to systematically profile imaging examinations for analysis. These limitations make it difficult to compare imaging practices, aggregate datasets, and interpret results across sites. To address these challenges, we have devised an ontology framework, the Operational Ontology for Radiology (OOR), that models the radiology study workflow from ordering through interpretation using explicitly defined entities, relationships, and annotated data properties. A CT-focused implementation of the framework was then formed, including 34 classes, 15 object properties, and 156 data properties in this present formation. The framework was deployed in a demonstrative use case scenario to illustrate the utility of OOR in drawing more clinically relevant radiation dose aggregates from clinical CT data. The work demonstrates that OOR provides both a structured representation of imaging data and a foundation for more transparent data aggregation, scalable secondary analysis, and reproducible characterization of imaging practices across institutions.</p>","PeriodicalId":516858,"journal":{"name":"Journal of imaging informatics in medicine","volume":" ","pages":""},"PeriodicalIF":3.1,"publicationDate":"2026-10-20","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149229681","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"M4DTrack: A DICOM-Based Imaging Informatics Workflow for User-Supervised 4DCT Lung Tumor Motion Reporting.","authors":"Tae Kyu Lee","doi":"10.1007/s10278-026-02435-8","DOIUrl":"https://doi.org/10.1007/s10278-026-02435-8","url":null,"abstract":"<p><p>Four-dimensional computed tomography (4DCT) is routinely used in radiotherapy simulation to evaluate respiration-induced tumor motion, but the motion information is often reviewed visually rather than captured as structured, phase-resolved data. We developed and evaluated M4DTrack, a DICOM-based imaging-informatics workflow for user-supervised 4DCT lung tumor motion analysis and structured, exportable three-dimensional center-of-mass (CoM) motion reporting. M4DTrack imports phase-resolved DICOM CT images, groups images by respiratory phase, displays linked axial/coronal/sagittal views, supports crosshair-based tumor localization and HU inspection, and applies seed or linked 3D-box initialization with sequential phase-to-phase HU-volume propagation. An 8-case lower-lobe lung 4DCT dataset (80 respiratory phases) was analyzed. Directional amplitudes, 3D vector amplitude, maximum phase-wise displacement, and agreement with independent ITV-GTV reference amplitudes were quantified. HU-weighted CoM amplitudes were 2.44 ± 1.44 mm, 2.64 ± 1.37 mm, and 6.04 ± 2.48 mm in X, Y, and Z patient-coordinate directions, respectively. Mean 3D vector amplitude was 7.26 ± 2.50 mm, and mean maximum phase-wise 3D displacement was 6.20 ± 2.10 mm. Biases versus ITV-GTV reference amplitudes were 0.01, - 0.84, and 0.23 mm for X, Y, and Z, respectively, with a 3D vector amplitude bias of - 0.29 mm. M4DTrack provides a user-supervised imaging-informatics bridge between qualitative 4DCT cine review and structured quantitative motion documentation. The pilot results support feasibility for generating exportable, phase-resolved tumor-motion metrics while preserving user-supervised review of 4DCT artifacts, HU-capture uncertainty, adjacent anatomy, and differences between CoM-based and extent-based measurements.</p>","PeriodicalId":516858,"journal":{"name":"Journal of imaging informatics in medicine","volume":" ","pages":""},"PeriodicalIF":3.1,"publicationDate":"2026-10-20","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149229668","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"TurkRadLLM: A Compact MedGemma-Based Model for Structuring Turkish Lumbar and Cervical Spine MRI Reports-A Multi-reader Validation Study.","authors":"Ahmet Harun Turgan, Hasan Çolakoğlu, Akif Demir, Zakir Sakcı","doi":"10.1007/s10278-026-02427-8","DOIUrl":"https://doi.org/10.1007/s10278-026-02427-8","url":null,"abstract":"<p><p>To develop and validate a compact, locally deployable MedGemma-based model converting unstructured Turkish spine MRI reports into structured six-field output. In this retrospective single-institution study, 5210 lumbar and 5017 cervical Turkish spine MRI reports were used to fine-tune MedGemma 4B IT with QLoRA. Silver-standard labels from a proprietary large language model served as training targets and as the reference for automated fidelity analyses, but not as the clinical reference standard. Evaluation comprised multi-reader validation of 180 authentic reports interleaved with 20 adversarial cases, a second round in which the readers scored the silver labels themselves, and comparison on the full held-out set against a rule-based baseline, the non-fine-tuned base model, and a much larger model. Every output was valid, schema-conformant JSON, and the grouped acceptance rate was 100% (95% CI 97.9-100), though one reader judged negative-finding lists incomplete in 57.8% of reports. On the full held-out set (1025 reports), fine-tuning increased agreement with the silver-label impression from 17.7% to 67.1% and urgent-flag sensitivity against the silver reference from 18.8% to 62.5% (both p < .001); the larger non-fine-tuned model reached 27.5% and 50.0%, and a rule-based baseline 99.0% on body region. When the readers scored the silver labels themselves, the model was rated higher on negative findings and lower on the impression. Deployment required 14.75 GB of GPU memory on a consumer card. Fine-tuning a compact medical language model produced structured output rated acceptable by clinicians, though negative-finding completeness remained reader-dependent. The gain lay in content-level fidelity rather than output validity.</p>","PeriodicalId":516858,"journal":{"name":"Journal of imaging informatics in medicine","volume":" ","pages":""},"PeriodicalIF":3.1,"publicationDate":"2026-10-20","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149229740","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"3D-DDN Faster R-CNN: Pulmonary Nodule Detection Through Cross-Slice Feature Fusion and Dynamic Displacement Modeling.","authors":"Chuantao Guan, Dongliang Yang, Zihao Leng, Shuo Liu, Ang Zu, Dongmei Luo, Hai Chen, Jiamei Wang","doi":"10.1007/s10278-026-02432-x","DOIUrl":"https://doi.org/10.1007/s10278-026-02432-x","url":null,"abstract":"<p><p>Accurate pulmonary nodule detection in computed tomography (CT) remains challenging because small- and low-contrast nodules may resemble vessels and other anatomical structures on individual axial slices. We propose 3D-DDN Faster R-CNN, a multi-slice detection framework that incorporates structured local axial context into a two-stage detector. The five-slice interactive blending (5-SIB) strategy organizes five consecutive CT slices as an ordered contextual group and preserves slice-specific feature streams extracted by a shared two-dimensional backbone. The cross-slice feature fusion (CSFF) module applies slice-wise spatial self-attention at shallow feature levels and learnable cross-slice aggregation at deep feature and region-of-interest levels. The dynamic displacement constraint (DDC) models proposal displacement across adjacent slices using a positive-sample regression loss and a margin-based negative-sample contrastive loss, while a class-weighted focal loss alleviates foreground-background imbalance. The model was evaluated on a private hospital dataset, LIDC-IDRI, and LNDb using patient-level data partitions. On the private hospital dataset, 3D-DDN achieved an average precision at an IoU threshold of 0.5 (AP_0.5) of 75.80% ± 1.05%, a sensitivity of 89.33% ± 1.91%, and a competition performance metric (CPM) of 91.20% ± 1.50%. The corresponding AP_0.5values on LIDC-IDRI and LNDb were 66.27% ± 0.45% and 48.20% ± 0.30%, respectively. Ablation experiments showed that direct five-slice stacking degraded performance, whereas structured feature fusion and displacement supervision recovered the performance loss and improved the overall detection profile. These results indicate that ordered cross-slice representation, hierarchical feature fusion, and proposal-displacement supervision provide complementary mechanisms for incorporating local three-dimensional context.</p>","PeriodicalId":516858,"journal":{"name":"Journal of imaging informatics in medicine","volume":" ","pages":""},"PeriodicalIF":3.1,"publicationDate":"2026-10-20","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149229742","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"The Performance Ceiling of LIDC-IDRI Is Label-Defined Rather than Image-Defined: A Computable Bound and a Purposive Audit of Published Studies.","authors":"Yen-Hsiang Wang","doi":"10.1007/s10278-026-02420-1","DOIUrl":"https://doi.org/10.1007/s10278-026-02420-1","url":null,"abstract":"<p><p>Benchmarks labelled by panels of human readers underpin most medical image analysis. We ask whether such a dataset has a single attainable ceiling. All 1319 annotation files of the Lung Image Database Consortium image collection (LIDC-IDRI) were parsed into 7034 reader-level ratings. Modelling each nodule by a latent malignancy propensity of which the four reader verdicts are conditionally independent draws, we prove that no image-based model can exceed, against a reader-derived label, the area under the receiver operating characteristic curve (AUC) of the propensity itself; we derive how that bound depends on the labelling policy, test the independence assumption against a prediction the fit does not use, simulate its failures, and audit a purposive sample of 22 published studies against the rules they declare. Among 905 nodules read by all four radiologists, all four agreed exactly for 11.7% and 58.0% carried ratings spanning two or more points. Across three latent families, retaining every nodule the ceiling is 0.898-0.913; discarding the 19.6% the readers split two-two lifts it to 0.946-0.954; retaining only unanimous nodules discards 57.7% and lifts it to 0.991-0.994. On the binary scale, no family predicted more reader agreement than observed; simulated reader heterogeneity would make these ceilings slightly low. Of 27 reported figures, 11 lie above the bound implied by the study's own declared rule, four beyond sampling error and two if modest heterogeneity is also allowed for. The ceiling is set by the label, not by the images.</p>","PeriodicalId":516858,"journal":{"name":"Journal of imaging informatics in medicine","volume":" ","pages":""},"PeriodicalIF":3.1,"publicationDate":"2026-10-19","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149225568","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"SVF-net: A Dual-encoder Network with Adaptive Gated Fusion for Colorectal Polyp Segmentation.","authors":"Zepeng Zhang, Long Yu, Shengwei Tian","doi":"10.1007/s10278-026-02415-y","DOIUrl":"https://doi.org/10.1007/s10278-026-02415-y","url":null,"abstract":"<p><p>Automatic colorectal polyp segmentation in retrospective still colonoscopy images remains challenging for small, low-contrast, partially occluded, or irregular lesions. We propose SVF-Net, a heterogeneous dual-encoder network combining a frozen SAM2 Hiera-L image trunk with a trainable MaxViT-RMLP-Small specialist branch. At four feature scales, the two streams are projected to a common width, spatially aligned, and reconciled by complementary channel-spatial gating before progressive decoding. Across five public benchmarks, three-run mean mDice is 0.922 on CVC-ClinicDB, 0.898 on Kvasir-SEG, 0.805 on CVC-ColonDB, 0.770 on ETIS-LaribPolypDB, and 0.896 on CVC-300. Matched projected-concatenation and fixed-weight ( <math><mrow><mi>G</mi> <mo>=</mo> <mn>0.5</mn></mrow> </math> ) controls retain the surrounding fusion path while removing input-dependent weighting; across ten matched training runs, adaptive gating gives higher mean mDice than both controls on all five datasets. Dataset-wide gate diagnostics on all 380 CVC-ColonDB and 196 ETIS cases indicate non-collapsed, scale- and region-dependent activations, while targeted one-factor sensitivity checks cover the MaxViT learning-rate multiplier, gate initialization, and auxiliary-loss weight. Matched per-case analyses further show that cross-domain gains are heterogeneous across lesion size, brightness, and an image-derived specular-reflection proxy. SVF-Net requires 92.8 ms per image (10.8 FPS) and 8.5 GB peak GPU memory on an NVIDIA RTX 3090 under the reported profiling protocol. The study evaluates retrospective static images only and does not establish real-time colonoscopy performance, prospective clinical effectiveness, or clinician benefit.</p>","PeriodicalId":516858,"journal":{"name":"Journal of imaging informatics in medicine","volume":" ","pages":""},"PeriodicalIF":3.1,"publicationDate":"2026-10-19","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149225611","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Automated MRI Series Classification Across Heterogeneous Datasets Using Limited Manual Annotation.","authors":"Ryan Fishman, Evan Calabrese","doi":"10.1007/s10278-026-02425-w","DOIUrl":"https://doi.org/10.1007/s10278-026-02425-w","url":null,"abstract":"<p><p>The purpose of this study is to determine the smallest tested fraction of manually labeled training data required for a multimodal artificial intelligence (AI) classifier to achieve satisfactory MRI series classification performance (macro-F1 ≥ 0.80) across five heterogeneous public MRI datasets using digital imaging and communications in medicine (DICOM) metadata, image data, or both. This retrospective study included 3804 examinations and 14,054 MRI series from five public DICOM datasets spanning brain, knee, lumbar spine, liver, and breast MRI, accessed February-April 2026. AI classification models included a random forest classifier using DICOM header metadata, a deep convolutional neural network classifier using the geometric center slice of each MRI series, and decision-fusion models. Models were trained with 5%-40% of studies manually labeled in 5-percentage-point increments, with additional 1%-4% fractions for the knee and lumbar spine. Model performance was evaluated on the remaining studies using study-level splits across five random seeds. Results were summarized as mean ± standard deviation. Zero-label unsupervised clustering results were evaluated separately on the full dataset. The macro-F1 target was reached at the smallest tested training fraction of 1% in the lumbar spine (0.945 ± 0.013), at 2% in the knee (0.818 ± 0.006), at 10% in the brain (0.872 ± 0.027), and at 20% in the liver (0.827 ± 0.016) and the breast (0.847 ± 0.012). The best-performing model differed by dataset, reflecting different contributions from metadata versus image data. Across five separately evaluated public MRI datasets, at least one header-based, image-based, or combined model achieved macro-F1 ≥ 0.80 using no more than 20% manually labeled studies.</p>","PeriodicalId":516858,"journal":{"name":"Journal of imaging informatics in medicine","volume":" ","pages":""},"PeriodicalIF":3.1,"publicationDate":"2026-10-19","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149225778","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Automated Multi-class Segmentation of Mandibular Odontogenic Cysts and the Mandibular Canal on CBCT Images Using a Hybrid Deep Learning Framework.","authors":"Adalet Çelebi, Dilan Aykanat, Özgür Tonkal, Deniz Bora Küçük, Ayşe Özcan Küçük, Muammer Türkoğlu, Fethi Atıl","doi":"10.1007/s10278-026-02393-1","DOIUrl":"https://doi.org/10.1007/s10278-026-02393-1","url":null,"abstract":"<p><p>Mandibular odontogenic cysts can complicate surgical planning if located near the mandibular canal. Accurate delineation of both structures is therefore clinically important. This study presents a deep learning-based method for simultaneously segmenting mandibular odontogenic cysts and the mandibular canal in cone-beam computed tomography (CBCT) images. The developed method relies on a hybrid deep learning architecture integrating a ConvNeXt-Base encoder, a Dilated Context Block, a U-Net-based decoder, and a Deep Supervision mechanism. This framework simultaneously segments the mandibular canal and odontogenic cyst regions within the same architecture, promoting multi-scale contextual feature learning. We evaluated the performance of the proposed method using fivefold cross-validation experiments on a real-world clinical CBCT dataset comprising 103 patient images. We assessed segmentation performance using overlap-based metrics together with boundary-aware HD95 and ASSD metrics. The experimental results demonstrated that the model achieved a mean intersection over union (mIoU) of 76.28% ± 1.83%. Excluding the background class yielded macro IoU and macro F1-score values of 64.46% ± 2.74% and 77.00% ± 1.89%, respectively. Meanwhile, the weighted IoU and weighted F1-score values were 78.71% ± 4.25% and 87.63% ± 2.63%, respectively. In boundary-based evaluations, the mandibular canal had an HD95 of 2.44 ± 2.98 mm and an ASSD of 0.45 ± 0.30 mm. The corresponding values for odontogenic cysts were 5.59 ± 3.65 mm and 1.43 ± 0.88 mm, respectively. These results suggest that the mandibular canal is more consistently delineated, while cyst segmentation remains more challenging, particularly at irregular and low-contrast boundaries. The proposed framework allows for the simultaneous and objective segmentation of mandibular odontogenic cysts and the mandibular canal within a unified, multi-class CBCT analysis framework. By providing reproducible delineation of both structures, this method could serve as a technical foundation for future quantitative analyses of lesion-canal relationships. However, further validation using larger, multicenter, volumetric CBCT datasets is required before clinical translation.</p>","PeriodicalId":516858,"journal":{"name":"Journal of imaging informatics in medicine","volume":" ","pages":""},"PeriodicalIF":3.1,"publicationDate":"2026-10-19","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149225761","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"BLIX-Ray: Chest X-Ray Report Generation with Vision-Language Models and Reinforcement Learning.","authors":"Taymaz Akan, Mohanna Orujpour, Pedram Salehpour, Sergio Alvarado Montalvo, Mohammad A N Bhuiyan","doi":"10.1007/s10278-026-02317-z","DOIUrl":"https://doi.org/10.1007/s10278-026-02317-z","url":null,"abstract":"<p><p>The radiology report is essential for directing the course of treatment for a patient. However, writing radiology reports is a huge burden for radiologists. Manual reporting faces challenges due to intra-labeler and inter-labeler variability, causing inconsistent patient treatment and diagnostic results, and complicating matters due to alertness and changing judgment. To address these challenges, we propose an automated radiology report generation model tailored for chest X-ray imaging. We introduce BLIX-Ray, a vision-language model (VLM) designed to generate radiology reports using a transfer learning approach. BLIX-Ray fine-tunes the pre-trained BLIP model to extract and align rich visual and textual representations suited to the medical imaging domain. We use reinforcement learning (RL) to directly optimize report generation based on language evaluation metrics in order to further enhance performance. We evaluate BLIX-Ray on two publicly available datasets: IU-Xray and CheXpert Plus. Experimental results demonstrate that our model achieves the highest CIDEr and BERTScore on both datasets, while n-gram-oriented baselines retain an advantage on BLEU. BLIX-Ray integrates reinforcement learning and transfer learning to improve medical report generation under limited and heterogeneous data conditions. The results highlight the potential of vision-language models to support automated and standardized radiology reporting. Notably, BLIX-Ray achieves the best reported CIDEr (0.6534) and BERTScore (0.8765) among the compared methods on IU-Xray and leads on CIDEr, METEOR, and BERTScore on CheXpert Plus (BERTScore 0.8407), demonstrating strong report-level consensus and semantic similarity with the reference reports.</p>","PeriodicalId":516858,"journal":{"name":"Journal of imaging informatics in medicine","volume":" ","pages":""},"PeriodicalIF":3.1,"publicationDate":"2026-10-19","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149225054","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"PW-GAN: Pseudo-Warping Field Guided GAN for Unsupervised Denoising of Fetal Brain MRI Images.","authors":"Linxin Chen, Guixing Fu, Pengjie Yang, Jiaying Lin, Kaida Ning, Zongzheng Chen, Yue Liu, Tao Tan, Hongsheng Liu, Bingguang Liu, Mengting Liu","doi":"10.1007/s10278-026-02416-x","DOIUrl":"https://doi.org/10.1007/s10278-026-02416-x","url":null,"abstract":"<p><p>Fetal brain MRI commonly uses thick-slice acquisitions to mitigate motion artifacts and preserve in-plane image quality. Multiple motion-corrupted stacks of 2D slices can be retrospectively reconstructed into isotropic 3D volumes; however, volumes reconstructed from thicker slices generally exhibit lower image quality than those reconstructed from thinner slices because limited through-plane information results in residual noise and inter-slice structural discontinuities. Existing post-reconstruction denoising methods are limited by their dependence on large-scale datasets comprising noisy-clean image pairs or insufficient modeling of inter-slice structural consistency. In this work, we propose a Pseudo-Warping Field Guided Unsupervised Generative Adversarial Network (PW-GAN), which formulates post-reconstruction denoising as unpaired translation between low- and high-quality MRI domains. By incorporating a pseudo-warping field module, our method significantly enhances inter-slice continuity in the reconstructed volumes while effectively suppressing residual noise and inter-slice structural discontinuities. On simulated data, PW-GAN achieved a PSNR of 39.65 dB and an SSIM of 0.9932. On real clinical data, it reduced the inter-slice warping error by up to 40% compared with unsupervised baselines. These results demonstrate that PW-GAN outperforms existing unsupervised models and achieves performance comparable to several state-of-the-art supervised methods. Qualitatively, PW-GAN produced cleaner results with smoother inter-slice structural transitions while preserving anatomical structures, demonstrating its effectiveness in post-reconstruction denoising of fetal brain MRI acquired from thick-slice scans.</p>","PeriodicalId":516858,"journal":{"name":"Journal of imaging informatics in medicine","volume":" ","pages":""},"PeriodicalIF":3.1,"publicationDate":"2026-10-19","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149225554","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}