Although low-dose CT can effectively reduce radiation risk, the projections are contaminated by quantum noise, resulting in severe noise and artifacts in the reconstructed image, which in turn compromises the accuracy of clinical diagnosis. To tackle this problem, we propose a reconstruction network, termed CSRCT, which incorporates convolutional sparse prior and a generalized sparse transform for low-dose CT imaging. Specifically, we adopt convolutional sparse representation (CSR) to model the reconstructed image through the convolution sum of dictionary and its corresponding coefficient maps, which suppresses random noise and preserves prior information. Furthermore, we propose a generalized sparse transform that leverages a gradient network to enhance sparsity, thereby enabling effective learning of sparse features, such as image details. By employing two optimization algorithms-the Alternating Direction Method of Multipliers (ADMM) and the Chambolle-Pock (CP) method, we solve the CSRCT model by deriving two deep unrolling networks: CSRCT-ADMM and CSRCT-CP. The experiments on simulated datasets demonstrate that, compared to existing methods, CSRCT exhibits superior performance in noise removal, artifact suppression, and texture detail preservation.
BackgroundPanoramic dental X-ray segmentation is critical clinically. However, inherent challenges-low contrast, blurred boundaries, complex root apices, and variable wisdom teeth-hinder pure CNN or Transformer models from effectively extracting global structures and local fine-grained features.ObjectiveLGMUNet is proposed, a dual-branch framework seamlessly integrating Mamba's long-range dependency modeling with CNN's local fine-grained extraction. This global-local synergy not only tackles the segmentation difficulties of complex dental topologies but also eliminates the semantic gaps inherent in conventional hybrid architectures.MethodsLGMUNet employs dynamic convolutional stems in shallow layers to extract local textures. In deep layers, the LGM module first aggregates channel information through an MLP, then feeds into a parallel dual-branch architecture: the Mamba branch models long-range dependencies via cross-scanning, and the DwConv branch captures local high-frequency details, with both branches interacting at same feature depth. The model was evaluated on the multi-source STDS and MICCAI 2023 datasets.ResultsIndependent repeat trials confirm that LGMUNet achieves state-of-the-art performance. On the STDS and MICCAI 2023 datasets, it achieves Dice coefficients of 92.64 ± 0.31% and 94.00 ± 0.19%, outperforming models like nnUNet and TMamba. It demonstrates superior accuracy and clinical robustness when addressing highly challenging cases, including impacted wisdom teeth and ambiguous boundaries.
To address the severe radial-edge artifacts and the loss of feature details in CT image reconstruction, this paper proposes a novel WF-Mamba algorithm integrating the Adaptive Wavelet Frequency Interaction Network (AW-net) with the Restructured Mamba Adjustment Module (RMAM). The AW-net leverages frequency-domain properties to initially enhance high-frequency components corresponding to edges and fine details in CT images. By integrating the Transformer architecture with the learnable 2D discrete wavelet transform (2D-DWT), we design the Optimized Encoding Block (OEB) and Wavelet Decoding Block (WDB) to improve the representation of image details and textures. Additionally, the Wavelet Aggregation Unit (WAU) is proposed to further enhance reconstruction quality by facilitating effective multi-scale information interaction. The RMAM incorporates a six-directional scanning mechanism to refine the initial reconstruction results generated by the AW-net, thereby enhancing the capture capability of fine-grained information and reducing computational complexity. To optimize model performance, a hybrid objective loss function is established by combining the L2 loss with Self-Supervised Wavelet Loss (SSWL). Experimental results demonstrate that the WF-Mamba achieves superior performance in eliminating radial edge artifacts, restoring textural details, and preserving boundary features compared with other discussed methods. Quantitatively, the WF-Mamba achieves a PSNR of 31.85 dB, an SSIM of 0.948, and a normalized MSE of 6.53 × 10-4 on the actual projection dataset, outperforming all compared methods. Furthermore, the reconstructed marked contour curves more closely approximate the reference image, demonstrating highly accurate restoration of the raw detailed structure.
BACKGROUND:Automatic detection and segmentation of dental diseases from panoramic x-ray images have been a difficult task to achieve due to the problems of class imbalance and varying sizes of lesions, as well as the inconspicuous nature of early caries. This paper introduces SwinDent-Seg, which is a hybrid architecture of Swin Transformer-CNN that includes CSFF module, ECMA attention, and high-resolution P2 detection head. METHODS:The proposed model was trained and evaluated on a dataset of 1018 annotated panoramic radiographs comprising seven dental pathology classes. Model development and hyperparameter optimization were performed using a validation set (220 images). SwinDent-Seg was compared with multiple CNN- and Transformer-based baselines, including YOLOv8, YOLO11 and RT-DETR-L. A confidence-weighted ensemble combining SwinDent-Seg and YOLOv8m-seg was subsequently evaluated on an independent test set (81 images). RESULTS:On the validation set, SwinDent-Seg achieved the best single-model performance, with an mAP50 of 0.8947, outperforming all baseline models, including YOLOv8m-seg (mAP50 = 0.6770). The proposed architecture markedly improved caries detection, increasing AP50 from 0.134 to 0.8953. On the independent test set, the confidence-weighted ensemble (SwinDent-Seg + YOLOv8m-seg) achieved the best overall performance, with an mAP50 of 0.9289 and a macro-AUC of 0.9515, exceeding the performance of either individual model. CONCLUSION:SwinDent-Seg demonstrates superior single-model validation performance, while the confidence-weighted ensemble provides the highest independent test-set performance. These findings indicate that combining Transformer-based global contextual representation with CNN-based localization improves robust multi-class dental pathology segmentation and supports the development of reliable computer-aided diagnostic systems for panoramic radiographs.
Accurate X-ray computed tomography (CT) image segmentation of the abdominal organs is a key task in automated medical image analysis, with crucial applications in clinical decision-making, computer-aided diagnosis, and surgical planning. However, existing methods still face significant challenges: insufficient capability in modeling long-range contextual dependencies, hindering the adaptability to the complicated morphological variations and spatial relationships of abdominal organs; and inaccurate boundary segmentation due to blurred edges and irregular anatomical structures, particularly in regions with high tissue adhesiveness. To address these issues, we propose an efficient abdominal multi-organ segmentation model, EEA-UNet. Specifically, we design an efficient element-wise adaptive (EEA) attention mechanism integrated into the skip connections to enhance inter-organ feature interactions while maintaining computational efficiency. This module effectively expands the receptive field, improving long-range dependency modeling. An enhanced multi-scale feature fusion (EMF) module is introduced to strengthen decoding capability, coupled with an edge-awareness composite loss function to optimize segmentation accuracy for small organs and boundary regions.Experimental results on the Synapse dataset demonstrate the competitive performance of EEA-UNet, achieving a Dice score of 84.45% and an HD95 of 0.16. Our method demonstrates a favorable trade-off between segmentation accuracy and computational efficiency, showing improved results compared with several existing approaches in both visual comparison and quantitative metrics.
Cardiomegaly screening via manual Cardiothoracic Ratio (CTR) measurement remains a clinical bottleneck, while contemporary deep learning solutions often suffer from algorithmic bloating. To address the need for resource-efficient and interpretable triage, this study proposes a framework driven by implicit morphological inference, which bypasses the requirement for explicit heart segmentation. We developed UBNet-Seg, a lightweight U-Net variant (2.3 million parameters) trained on a heterogeneous dataset of 11,748 images to segment lung fields as a geometric proxy. Performance was rigorously evaluated across unseen domains using external NIH and OpenI datasets. The model achieved a lung Dice Coefficient of 95.85% and an inference time of 0.05 s. Fully automated accuracy reached 90.31% on NIH and 76.07% on OpenI. Crucially, the integration of a Human-in-the-Loop mechanism successfully neutralized domain shifts; the McNemar test confirmed that expert-guided refinement provided statistically significant accuracy improvements to 93.63% and 91.21% on NIH and OpenI, respectively (p<0.001). By validating that medial lung boundaries serve as reliable structural proxies, this framework offers a robust, explainable, and low-latency alternative suitable for resource-constrained clinical settings.
Accurate localization and counting of tiny electronic components in high-resolution X-ray images is a critical yet challenging task in nuclear science, radiation imaging, and industrial quality control. Traditional methods suffer from poor generalization in cluttered scenes, while deep learning approaches are limited by the lack of large-scale annotated datasets. This study aims to develop a semi-supervised detection framework that achieves high-precision component localization and counting in 3072×3072-pixel X-ray images, while significantly reducing manual annotation costs. We propose YOLO-DEP, a novel object detector that integrates the YOLOv11 architecture with a Deep Encoding Processor (DEP) and a Graph Attention Network (GAT). The DEP module enhances feature discrimination for dense and small targets via half-channel and spatial attention mechanisms. A semi-supervised label propagation strategy is designed to generate high-confidence pseudo-labels from only one labeled image per category, leveraging feature similarity graphs and GAT-based confidence filtering. We also introduce LEEC, a large-scale X-ray dataset for electronic component counting, containing 720 images across 49 component types. YOLO-DEP outperforms state-of-the-art detectors on both LEEC and DOTAv1 datasets. Specifically, YOLO-DEP-x achieves 79.2% mAP50 and 70.9% mAP50-95 on LEEC, with a counting error rate as low as 0.8, providing an immediately deployable solution for industrial automation, nuclear electronics inspection, fuel-assembly verification and broader radiation-based quality-control lines.
BackgroundGlaucoma is a leading cause of irreversible vision loss and is characterized by subtle structural changes in the optic disc and optic cup. However, existing automated detection systems often suffer from weak boundary delineation, dataset variability, and unstable feature learning, which limit their generalizability and clinical reliability.ObjectiveThis study aims to develop a unified and anatomically guided framework for accurate and reliable automated glaucoma detection from fundus images.MethodsThe proposed pipeline begins with contrast-enhanced preprocessing to improve image quality, followed by an Attention-guided Multi-scale Edge-aware Segmentation Network (AME-SegNet) for precise segmentation of the optic disc and optic cup. Both deep convolutional features and clinically relevant geometric features are extracted and optimized using Bitterling Colony Optimization (BCO) to select the most discriminative attributes. A Convolutional Transformer (CT) is then employed to integrate local convolutional representations with global attention mechanisms for robust classification. Additionally, the Honey Badger Algorithm (HBA) is used for automatic parameter tuning to ensure stable convergence.ResultsExperimental evaluation demonstrates high segmentation performance with Dice scores of 97.36% for the optic disc and 96.72% for the optic cup on the Drishti-GS1 dataset. The classification model achieves accuracies of 98.63% on RIM-ONE and 98.96% on ORIGA-Light datasets, indicating strong generalization capability.ConclusionsThe proposed framework exhibits robust performance, high accuracy, and strong generalization across multiple datasets. These results highlight its effectiveness and clinical potential for reliable automated glaucoma screening and early diagnosis.
Background Radiomic feature extraction from cone-beam computed tomography (CBCT) images in radiotherapy has potential for predicting tumor control and treatment-related toxicity. However, the reliability of CBCT-based radiomics is limited by variations in scatter intensity associated with differences in patient size.Objective To evaluate the impact of patient size on CBCT-derived radiomic features and investigate whether a novel quantitative CBCT technique can reduce patient size-induced radiomic feature variability.Methods Phantoms representing small and large body habitus were scanned using a linac-mounted CBCT incorporating a two-dimensional antiscatter grid (CBCT-2DASG) and standard clinical protocols. Ninety-three intensity and texture features were extracted from CBCT and multidetector CT (MDCT) images, and robust features were identified by analyzing the concordance correlation between small and large phantom images.Results Compared with MDCT, 53% of CBCT-2DASG features were robust to patient size variation, versus 8 features from standard CBCT. The CBCT-2DASG reduced size-dependent feature variation, approximately fourfold on average for intensity-based features, with improvement varying across feature types.Conclusions CBCT-2DASG with effective scatter suppression reduces patient size-dependent radiomic feature variability in phantom experiments, improving the robustness of CBCT-derived features to patient size variation. This represents a promising step toward more reliable CBCT-based radiomics in radiotherapy.
Real-world skin lesion classification faces three major challenges: severe class imbalance, high intra-class variability, and the need to reject out-of-distribution (OOD) samples. Conventional monolithic models often struggle to address these issues simultaneously. To mitigate this limitation, we propose a multi-stage decoupled hybrid framework that combines Supervised Contrastive Learning (SupCon) with structural reconstruction. First, representation learning is decoupled using SupCon. Compared with standard cross-entropy training, SupCon alleviates feature degradation under long-tailed distributions by encouraging a more balanced feature space. Second, to address open-set recognition, we integrate contrastive semantic features with structural anomaly scores derived from an independent Convolutional Autoencoder (CAE). These complementary signals-semantic confidence and reconstruction error-are fused through a linear boundary formulation to support both known-class classification and unknown-sample rejection. On the ISIC 2019 dataset, the proposed framework achieves a Balanced Accuracy of 78.5% on known classes and improves the unknown-class F1-score to 51.3%. These results indicate that semantic-structural fusion enhances robustness under long-tailed and open-set conditions.
Accurate segmentation of stenosis in X-ray angiography (XRA) images is crucial for the objective assessment of stenosis severity and subsequent treatment planning in coronary artery disease. Current clinical practice primarily relies on subjective visual evaluation, which suffers from significant inter-observer variability. In this work, we propose a deep learning model enhanced with a novel Hybrid Context-Aware Attention (HCA) module. HCA employs a parallel dual-pathway design that integrates global inter-channel attention and grouped multi-scale spatial aggregation. This integration enhances feature discriminability and spatial-context modeling, leading to more accurate and anatomically consistent stenosis segmentation in XRA. Evaluated on three independent datasets, our method achieves competitive performance against existing approaches across multiple metrics, demonstrating consistent leading performance. Ablation and attention visualization studies further confirm the contribution of the designed module to reducing segmentation errors and enhancing focus on stenotic regions. These findings demonstrate that the proposed model is an effective and generalizable approach for stenosis segmentation in XRA, with the potential to support standardized assessment in clinical practice.
BackgroundPanoramic radiography is widely used in dental diagnostics; however, conventional systems require a dedicated panoramic scan, increasing radiation exposure and limiting adaptation to patient-specific anatomical variability. Cone-beam computed tomography (CBCT) provides volumetric information that may potentially replace separate panoramic acquisitions.PurposeThis study proposes CT2Pano, a projection-domain reconstruction method that generates patient-specific panoramic images directly from CBCT projection data without requiring an additional panoramic scan.MethodsThe CT2Pano algorithm automatically detects teeth from CBCT images and estimates a patient-specific dental arch curve using Bézier curve optimization. Based on the extracted arch, a virtual panoramic scan trajectory is defined and panoramic projections are extracted from the corresponding CBCT projections. Geometric correction combined with subpixel rebinning compensates for magnification variations caused by the fixed rotation center of the CBCT systems. The final panoramic image is reconstructed using a shift-and-add (SAA) method that enhances structures within the focal trough while suppressing out-of-plane structures.ResultsSimulation and phantom experiments demonstrate that CT2Pano generates panoramic images with anatomical fidelity comparable to conventional panoramic imaging systems. The method effectively compensates for geometric distortion and maintains image quality even in the presence of metal artifacts or patient misalignment.ConclusionsCT2Pano enables patient-specific panoramic reconstruction directly from CBCT projection data, eliminating the need for an additional panoramic scan. By operating in the projection domain, the proposed framework preserves measurement information while reducing radiation exposure and improving imaging workflow efficiency, highlighting its potential for CBCT-based panoramic imaging in dental diagnostics.
Background AI-enhanced CT reconstruction enables strong noise suppression and dose reduction, but aggressive denoising can distort diagnostically relevant texture, lowering reader confidence-the "texture preservation paradox." Clinical evidence linking texture assessment with validation across AI approaches remains fragmented. Objective To systematically evaluate AI-enhanced CT reconstruction techniques with emphasis on texture preservation and to summarize clinical validation evidence supporting dose optimization. Methods This systematic review followed PRISMA guidelines and searched PubMed, IEEE Xplore, Scopus, and Web of Science (January 2020 to November 2025) for peer-reviewed studies evaluating AI-enhanced CT reconstruction with texture-related image quality assessment and clinical validation. Results Fifteen clinical studies involving 1847 patients were included in the qualitative synthesis. AI-enhanced CT reconstruction showed consistently improved image quality compared with conventional reconstruction, with representative reports of higher PSNR and SSIM, reduced noise, and improved lesion conspicuity. Multiple studies emphasized preservation of diagnostically relevant texture, addressing the limitation that conventional metrics do not fully capture perceptual fidelity. Reader-based validation indicated improved diagnostic confidence, better inter-reader consistency, and maintained diagnostic acceptability at reduced radiation dose. Conclusions AI-enhanced CT reconstruction shows promising clinical utility for improving image quality and supporting dose optimization while preserving texture characteristics. However, heterogeneity in study design and evaluation metrics warrants cautious interpretation and highlights the need for standardized assessment.
Globally, breast cancer is recognized as the leading form of malignant tumor in women, significantly impacting public health on an international scale. However, current breast cancer classification methods often overlook the imbalanced data distribution within breast cancer datasets. This leads to models being biased toward majority-class samples during classification, while underperforming on minority-class samples, thereby reducing overall classification accuracy. Moreover, multi-classification approaches can provide physicians with richer diagnostic information and better support the development of treatment strategies. To address these challenges, we introduce a breast cancer multi-classification algorithm, CLID (Contrastive Learning for Imbalanced Distribution), which utilizes contrastive learning to address the problem of imbalanced data distribution. Specifically, a preprocessed convolutional neural network is employed to extract features from histopathological images. Learnable archetype vectors are employed to model inter-category similarities, and a contrastive learning loss function is used to train the model. Experimental results on the BreakHis and BRACS datasets demonstrate that the proposed method achieves superior performance compared with existing approaches and representative imbalance-aware learning strategies, showing strong robustness under different magnifications.
BackgroundIt is common for X-ray computed tomography (CT) images to be reconstructed differently for various clinical examination purposes. This is primarily because we aim to meet two clinical requirements: improving spatial resolution and reducing noise by changing the reconstruction parameters.ObjectiveTwo deep learning-based methods, super resolution (SR) and denoising, respectively, have been proposed to address these requests. We present a single neural network that can perform SR and denoising simultaneously.MethodsWe propose using existing ultra-high-resolution CT (UHR-CT) data to achieve high spatial resolution and reduces noise. We generated specific input data, which is normal-resolution and high-noise data, simulated from UHR-CT data. Afterwards, we apply the network to NR-CT data, the resulting method, called SR-Denoise deep-learning reconstruction (DLR). In experiments, we measured modulation transfer function as the quantitative study. We also evaluated the performance using both simulated and real clinical data in NR-CT data, with UHR-CT data serving as the ground truth.ResultsSR-Denoise DLR achieved performance on both SR and denoising tasks that was equivalent to training them individually and outperformed the methods currently used in clinical settings.ConclusionsSR-Denoise DLR utilizes the spatial resolution of UHR-CT and takes advantage of NR-CT to significantly reduce noise.
Intraoperative cone beam computed tomography (CBCT) is critical for pedicle screw planning; however, image quality is frequently compromised by artifacts and low contrast, potentially leading to adverse clinical outcomes. To address these limitations, we propose the Spatiotemporal Adaptive Warm-Start Diffusion Model (STADW-M), a novel framework aimed to generate high-quality synthetic CT (sCT) images from CBCT data, thereby enhancing surgical precision. The STADW-M integrates an Artifact-Aware Adaptive Diffusion Module to mitigate localized artifact distributions and a Dually-Guided Structural Consistency Module to preserve anatomical integrity. Furthermore, we employ a CBCT Warm-Start strategy alongside composite loss functions to optimize textural fidelity and accelerate model convergence. Quantitative experiments demonstrated significant improvements over original CBCT images: with RMSE decreased from 890.1 to 152.9 HU, MAE decreasing from 859.7 to 102.6 HU, and PSNR increased from 13.6 to 27.9 dB. Crucially, the generated sCTs maintained high anatomical consistency with reference CTs. In clinical validation, automated screw planning based on sCTs achieved a 100% Grade A standard, with 94.7% of screws placed without cortical breach and 5.3% exhibiting only minor (<2 mm) erosion. The proposed method effectively synthesizes high-quality CT images, preserving vertebral anatomy and significantly improving the accuracy and safety of intraoperative pedicle screw planning.
OBJECTIVE:To compare treatment plans created using the Eclipse and Prowess treatment planning systems for patients diagnosed with high-grade brain tumors in the left temporal lobe and to evaluate target volume coverage and doses delivered to critical organs using intensity-modulated radiotherapy (IMRT) technique. MATERIALS AND METHODS:This retrospective comparative planning study included 15 patients. Plans were created on CT images using step-and-shoot IMRT with 6 MV photons. Eclipse used the Pencil Beam Convolution (PBC) algorithm, while Prowess used the Collapsed Cone Convolution (CCCS) algorithm. A total dose of 60 Gy in 30 fractions was prescribed. Dosimetric parameters were analyzed statistically. RESULTS:Both TPSs achieved comparable target coverage, with no significant differences in conformity and homogeneity indices (p = 0.075 and p = 0.590, respectively) or dose-volume parameters (D95, D98, D2; p > 0.05). Eclipse provided significantly lower doses to the ipsilateral lens, brainstem, left optic nerve, cochlea, and whole brain V40 (p < 0.05). CONCLUSION:Both TPSs demonstrated adequate target coverage; however, Eclipse achieved statistically significant dose reductions in several critical structures, which may have clinical relevance in reducing toxicity risk.
Cardiovascular disease remains the leading cause of mortality globally, with escalating risk factors and increasing pressure on the healthcare system. Despite the critical importance of early diagnosis, it is impeded by challenges, such as data imbalance, feature complexity, and variability in diagnostic processes. These challenges necessitate the development of sophisticated intelligent systems to ensure the accurate and timely prediction of heart disease. This work presents an intelligent system that integrates advanced machine learning techniques for heart disease prediction, employing the capuchin search algorithm graylag goal optimization (CSAGGO) and Quantum-SpinalNet (Q-SpinalNet) for enhanced classification. The methodology begins with data preprocessing using Principal Component Analysis (PCA) and synthetic minority oversampling technique (SMOTE) to address issues of dimensionality and class imbalance. An improved Fuzzy C-Means Gaussian Mixture Model (FCM-GMM) was utilized for clustering, while the Least Absolute Shrinkage and Selection Operator (LASSO) identified the most informative features. To enhance interpretability, SHapley Additive exPlanations (SHAP) values were employed to elucidate the influence of individual features on predictions, providing actionable insights for healthcare professionals. The hybrid CSAGGO-Q-SpinalNet framework surpasses the existing methods, offering a robust, efficient, and explainable solution for heart disease prediction. The proposed system achieved exceptional performance metrics, including 98.44% accuracy, 96.89% sensitivity, 96.83% specificity, and 96.22% precision on the Cleveland dataset. Additionally, the model demonstrated low error rates with a 4.24% false positive rate (FPR), 4.38% false negative rate (FNR), and 4.05% false discovery rate (FDR). This system holds significant promise for real-world clinical applications by facilitating early diagnosis and personalized treatment strategies.
This paper presents an iterative algorithm for reconstruction of dynamic non-monotonic process in a three-dimensional object (4D tomography) using only one projection per each time step. In this case, a priori knowledge about the unchanging (initial) structure of the object is used, and the processing of intermediate data at each iteration is performed in the projection data space. The proposed algorithm is compared with the previously proposed algorithm for monotonic processes and its extensions for non-monotonic processes, where the processing of intermediate data in the space of reconstructed objects is used. Virtual experiments have been conducted to reconstruct the dynamics of voxels values changes in time. The advantage of the new approach in both speed and quality of reconstruction is shown. The influence of the internal structure of dynamic regions on the reconstruction quality is investigated. The new 4D tomography algorithm presented in this paper shows the possibility of reconstructing a dynamic process in a three-dimensional volume using only one projection per time step, which can be useful in different fields of science such as materials science, geology, medicine, etc. The proposed algorithm is demonstrated on an unambiguous mathematical phantom, serving as a proof of concept. Its application to more complex or real experimental data will require further study and adaptation.
BackgroundExisting methods for segmenting liver cancer from single-modal medical images fail to effectively leverage potential correlations across modalities. These correlations between the anatomical structures of liver cancer and the liver are also crucial for accurate liver cancer segmentation. These challenges not only limit the performance and scalability of liver cancer segmentation models but also pose significant challenges for researchers seeking to develop multimodal, low-annotation-dependent solutions. Therefore, it is necessary to propose a universal liver cancer segmentation network for abdominal computed tomography (CT) and magnetic resonance (MR) medical images.MethodBased on the above, we propose a two-stage universal liver cancer segmentation network for 3D dual-modality abdominal nuclear medical images using a mixed-label, multi-type training strategy. In stage 1, two CT and MR liver segmentation models are trained to generate liver mask images for CT and MR multimodal abdominal images without liver mask label images, thereby solving the laborious technical problem of liver labeling. In stage 2, a mixed-label strategy is proposed, where a mixed-label pool is constructed from CT and MR liver mask images generated by the aforementioned liver segmentation models, along with liver label images and their corresponding liver cancer label images. Subsequently, a universal liver cancer segmentation model is trained using a mixed-label, multi-type training strategy that fully considers potential correlations among different medical imaging modalities, liver cancer, and the liver.ResultsThe proposed liver cancer segmentation model, based on Medformer and the proposed mixed-label, multi-type abdominal-image training strategy, performs best, validating the effectiveness of the proposed strategy.DiscussionThe proposed two-stage universal liver cancer segmentation network, based on a mixed-label and multi-type training strategy, can effectively segment liver cancer in different 3D dual-modality abdominal CT and MR images, which may become an indispensable quantitative analysis tool for liver cancer in clinical practice.
