Efficient Generative-Adversarial U-Net for Multi-Organ Medical Image Segmentation

Journal of Imaging, Год журнала: 2025, Номер 11(1), С. 19 - 19

Опубликована: Янв. 12, 2025

Manual labeling of lesions in medical image analysis presents a significant challenge due to its labor-intensive and inefficient nature, which ultimately strains essential resources impedes the advancement computer-aided diagnosis. This paper introduces novel image-segmentation framework named Efficient Generative-Adversarial U-Net (EGAUNet), designed facilitate rapid accurate multi-organ labeling. To enhance model’s capability comprehend spatial information, we propose Global Spatial-Channel Attention Mechanism (GSCA). mechanism enables model concentrate more effectively on regions interest. Additionally, have integrated Mapping Convolutional Blocks (EMCB) into feature-learning process, allowing for extraction multi-scale information adjustment feature map channels through optimized weight values. Moreover, proposed progressively enhances performance by utilizing generative-adversarial learning strategy, contributes improvements segmentation accuracy. Consequently, EGAUNet demonstrates exemplary public datasets while maintaining high efficiency. For instance, evaluations CHAOS T2SPIR dataset, achieves approximately 2% higher Jaccard metric, 1% Dice nearly 3% precision metric comparison advanced networks such as Swin-Unet TransUnet.

Язык: Английский

---

The predictive value of radiomics and deep learning for synchronous distant metastasis in clear cell renal cell carcinoma

Discover Oncology, Год журнала: 2025, Номер 16(1)

Опубликована: Янв. 25, 2025

The objective of this research was to devise and authenticate a predictive model that employs CT radiomics deep learning methodologies for the accurate prediction synchronous distant metastasis (SDM) in clear cell renal carcinoma (ccRCC). A total 143 ccRCC patients were included training cohort, 62 validation cohort. images from all normalized, tumor regions manually segmented via ITK-SNAP software. Radiomic features extracted FAE toolkit. least absolute shrinkage selection operator (LASSO) algorithm employed select build various machine models. Additionally, largest cross-section cropped train model. Multiple models trained predict SDM patients. results best then fused with those create combined Of 944 radiomic identified, 15 closely associated SDM. With these features, support vector (SVM) emerged as most effective, demonstrating areas under curve (AUC) 0.860 0.813 respectively. Among models, ResNet101 performed optimally, achieving AUC 0.815 0.743 yielded an 0.863. Decision analysis suggested offers superior clinical applicability. integrates learning, showing significant potential predicting It holds promise supporting decision-making, reducing missed diagnoses SDM, guiding further enhancing their systemic examinations.

Язык: Английский

---

Highlighting the Advanced Capabilities and the Computational Efficiency of DeepLabV3+ in Medical Image Segmentation: An Ablation Study

BioMedInformatics, Год журнала: 2025, Номер 5(1), С. 10 - 10

Опубликована: Фев. 14, 2025

Background: In clinical practice, identifying the location and extent of tumors lesions is crucial for disease diagnosis treatment. Artificial intelligence, particularly deep neural networks, offers precise automated segmentation, yet limited data high computational demands often hinder its application. Transfer learning helps mitigate these challenges by significantly reducing costs, although applying models can still be resource intensive. This study aims to present flexible computationally efficient architecture that leverages transfer delivers highly accurate results across various medical imaging problems. Methods: We evaluated three datasets with varying similarities ImageNet: ISIC 2018 (skin lesions), CBIS-DDSM (breast masses), Shenzhen Montgomery CXR Set (lung segmentation). An ablation on tested pre-trained backbones, architectures, loss functions. Results: The optimal configuration—DeepLabV3+ a ResNet50 backbone Log-Cosh Dice loss—was validated remaining datasets, achieving state-of-the-art results. Conclusion: Computationally simpler architectures deliver robust performance without extensive resources, establishing DeepLabV3+ as baseline future studies. domain, enhancing quality more critical improving segmentation accuracy than increasing model complexity.

Язык: Английский

---

Artificial intelligence-enabled automated analysis of transmission electron micrographs to evaluate chemotherapy impact on mitochondrial morphology in triple negative breast cancer

bioRxiv (Cold Spring Harbor Laboratory), Год журнала: 2025, Номер unknown

Опубликована: Фев. 23, 2025

ABSTRACT Advancements in transmission electron microscopy (TEM) have enabled in-depth studies of biological specimens, offering new avenues to large-scale imaging experiments with subcellular resolution. Mitochondrial structure is growing interest cancer biology due its crucial role regulating the multi-faceted functions mitochondria. We and others established mitochondria triple-negative breast (TNBC), an aggressive subtype limited therapeutic options. Building upon our previous work demonstrating regulatory mitochondrial dynamics metabolic adaptation survival chemotherapy-refractory TNBC cells, we sought extend those findings a analysis micrographs. Here present UNet artificial intelligence (AI) model for automatic annotation assessment morphology feature quantification. Our trained on 11,039 manually annotated across 125 micrographs derived from variety orthotopic patient-derived xenograft (PDX) mouse tumors adherent cell cultures. The achieves F1 score 0.85 test at pixel level. To validate ability detect expected structural features, utilized primary skeletal muscle cells genetically modified lack Dynamin-related protein 1 (Drp1). algorithm successfully detected significant increase elongation, alignment well-established Drp1 as driver fission. Further, subjected vitro vivo models conventional chemotherapy treatments commonly used clinical management TNBC, including doxorubicin, carboplatin, paclitaxel, docetaxel (DTX). found substantial within-sample heterogeneity both observed consistent reduction elongation DTX-treated specimens. went compare mammary matched lung metastases highly metastatic PDX uncovering length lesions. large, curated dataset provides high statistical power frequent chemotherapy-induced shifts shapes sizes residual left behind after treatment. successful application AI capture marks step forward high-throughput structures, enhancing understanding how morphological changes may relate efficacy mechanism action. Finally, micrograph - now publicly available serves unique gold-standard resource developing, benchmarking, applying computational models, while further advancing investigations into impact biology.

Язык: Английский

---

AOA-guided hyperparameter refinement for precise medical image segmentation

Alexandria Engineering Journal, Год журнала: 2025, Номер 120, С. 547 - 560

Опубликована: Фев. 24, 2025

Язык: Английский

---

Deep Convolutional Neural Networks in Medical Image Analysis: A Review

Information, Год журнала: 2025, Номер 16(3), С. 195 - 195

Опубликована: Март 3, 2025

Deep convolutional neural networks (CNNs) have revolutionized medical image analysis by enabling the automated learning of hierarchical features from complex imaging datasets. This review provides a focused CNN evolution and architectures as applied to analysis, highlighting their application performance in different fields, including oncology, neurology, cardiology, pulmonology, ophthalmology, dermatology, orthopedics. The paper also explores challenges specific outlines trends future research directions. aims serve valuable resource for researchers practitioners healthcare artificial intelligence.

Язык: Английский

---

A Two stage deep learning network for automated femoral segmentation in bilateral lower limb CT scans

Scientific Reports, Год журнала: 2025, Номер 15(1)

Опубликована: Март 17, 2025

This study presents the development of a deep learning-based two-stage network designed for efficient and precise segmentation femur in full lower limb CT images. The proposed incorporates dual-phase approach: rapid delineation regions interest followed by semantic femur. experimental dataset comprises 100 samples obtained from hospital, partitioned into 85 training, 8 validation, 7 testing. In first stage, model achieves an average Intersection over Union 0.9671 mean Average Precision 0.9656, effectively delineating femoral region with high accuracy. During second attains Dice coefficient 0.953, sensitivity 0.965, specificity 0.998, pixel accuracy 0.996, ensuring When compared to single-stage SegResNet architecture, demonstrates faster convergence during reduced inference times, higher accuracy, overall superior performance. Comparative evaluations against TransUnet further highlight network's notable advantages robustness. summary, offers efficient, accurate, autonomous solution large-scale complex medical imaging datasets. Requiring relatively modest training computational resources, exhibits significant potential scalability clinical applicability, making it valuable tool advancing image supporting diagnostic workflows.

Язык: Английский

---

Bone Appetit: Skellytour Sets the Table for Robust Skeletal Segmentation

Radiology Artificial Intelligence, Год журнала: 2025, Номер 7(2)

Опубликована: Март 1, 2025

Язык: Английский

---

Deep Learning Approach for ASD Detection with GAN-Driven MRI Data Augmentation

International Journal of Management Technology and Social Sciences, Год журнала: 2025, Номер unknown, С. 78 - 93

Опубликована: Март 21, 2025

Autism spectrum disorder (ASD) is a highly heterogeneous neurodevelopment that primarily characterized by impaired communication, behaviours and social interaction. These will remain strong commonsense can not only diagnose ASD accurately but provide as swiftly, especially, when trying to work with small-labelled datasets in the presence of neuro-imaging data such MRI photos, etc. Here, we present hybrid deep learning framework combining CNNs Swin attention maps discriminate from different groups children diagnosis autism based on complementarity features they extract. Furthermore, introduce GAN-based synthetic augmentation method tackle scarcity issue. we’re operating slices already prepared for them each scan, again crop middle slice each, then wrap-in needed resize so it could be input model. This results bigger/ broader training set conditioning GAN generator produce images closely resemble typical neuroimaging data. Similarly, combined CNN model provides short large feature extraction depends depth-wise convolution comprehensive "lossless" classification approach layered parts. Trained validated non-ASD subjects, generated superlative Hence, have metrics e.g. accuracy, precision, recall, AUC-ROC check performance our validation accuracy score 75%+. Experiments show improvement using mechanisms. does, however, appears most unfortunate prevalent problem, annotation has too high resource requirements make this often pointless like learning, lot machines just end up wasted cleaned up, becoming clear all papers review undergo problem real-life mindset. In future, explore 3D research similar methods examine stronger mechanisms, which may yield an incremental identification ASD. The finding help adaptation machine learn apply disease’s purpose potentially raise feasibility access diagnostic tools.

Язык: Английский

---

Synergistic Multi-Granularity Rough Attention UNet for Polyp Segmentation

Journal of Imaging, Год журнала: 2025, Номер 11(4), С. 92 - 92

Опубликована: Март 21, 2025

Automatic polyp segmentation in colonoscopic images is crucial for the early detection and treatment of colorectal cancer. However, complex backgrounds, diverse morphologies, ambiguous boundaries make this task difficult. To address these issues, we propose Synergistic Multi-Granularity Rough Attention U-Net (S-MGRAUNet), which integrates three key modules: Hybrid Filtering (MGHF) module extracting multi-scale contextual information, Dynamic Granularity Partition Synergy (DGPS) enhancing polyp-background differentiation through adaptive feature interaction, (MGRA) mechanism further optimizing boundary recognition. Extensive experiments on ColonDB CVC-300 datasets demonstrate that S-MGRAUNet significantly outperforms existing methods while achieving competitive results Kvasir-SEG ClinicDB datasets, validating its accuracy, robustness, generalization capability, all effectively reducing computational complexity. This study highlights value multi-granularity extraction attention mechanisms, providing new insights practical guidance advancing theories medical image segmentation.

Язык: Английский

---