The choice of segmentation method for HR-pQCT scans influences accuracy of bone microarchitecture measurements. Smaller or under-mineralized bone can present challenges in accurate extraction of bone structure using global thresholding methods, as local variations in intensity are not considered and finer structural details are not detected. This is especially important in small hand bones, where bone structure is finer, or younger populations, where bone tissue may be under mineralized. This study compared accuracy of global thresholding methods using Gaussian and Laplace-Hamming filters, and an adaptive local thresholding (AT) method in HR-pQCT scans of carpal bones. Eight ex vivo human cadaveric forearms (n = 64 carpal bones) were analyzed. Three specimens (n = 24 carpal bones, 2 female, mean age: 82.7 ± 4.6yr) were used for AT parameter optimization, and 5 specimens (n = 40 carpal bones, 3 female, mean age: 82.0 ± 6.4yr) were used to compare trabecular microarchitecture accuracy and spatial agreement relative to micro-CT (μCT, 20μm isotropic resolution). Micro-CT images were segmented using a Gaussian filter and Otsu's method, and HR-pQCT images were segmented using Gaussian filtering and global thresholding, Laplace-Hamming filtering and global thresholding, and the AT method. Trabecular thickness (Tb.Th), separation (Tb.Sp), and bone volume fraction (Tb.BV/TV) accuracy were evaluated, and spatial agreement was assessed using Dice similarity coefficients (DSC), 95th percentile Hausdorff distances (HD95), and average symmetric surface distances (ASSD). The AT method yielded the smallest absolute and relative errors, and lowest bias across all trabecular parameters. Compared to the Gaussian and fixed threshold method, AT reduced mean absolute error by 36% for Tb.Th, 14% for Tb.Sp, and 15% for Tb.BV/TV, and achieved the highest spatial agreement with μCT (DSC = 0.84, HD95 = 0.061mm, and ASSD = 0.018mm). These findings extend prior HR-pQCT segmentation validation studies to carpal bones and demonstrate that AT outperforms the standard and Laplace-Hamming methods.

Evaluation of dual-energy CT virtual monoenergetic imaging for target delineation in rectal cancer radiotherapy.

Data-centric approach to uncover and mitigate representation bias in AI segmentation of esophageal tumors on CT

Deep learning synthesis of virtual T2-weighted fat-suppressed MR images: a multi-center study.

Development and validation of a fully automated transformer-based 3D framework for pancreatic fat quantification in pancreatic steatosis.

In chronic diseases, accelerated muscle mass loss is associated with poor clinical outcomes. Computed tomography (CT) is considered a reference standard for assessing muscle mass, but it is limited for longitudinal assessment. Ultrasound (US) is more suitable for longitudinal measurements, but limited reliability data or reference values exist to inform clinical adoption. This pilot study evaluated the reliability of rectus femoris (RF) muscle US measurements [cross-sectional area (CSA) and shear-wave elastography (SWE) stiffness] and investigated their relationship with CT-derived truncal muscle mass. Forty healthy living liver donors undergoing abdominal CT were included. CT-derived skeletal muscle area and skeletal muscle index at T12 and L3 were quantified using deep learning. US B-mode and SWE RF images obtained with manual and automated measurements. Reliability was assessed using intraclass correlation (ICC). Agreement between manual and automated methods was evaluated using the Dice coefficient. US and CT measurements associations were evaluated using Pearson correlation and multiple linear regression. Inter-reader agreement for manual US CSA was excellent (ICC=0.95, 95% CI: 0.88-0.97). Test-retest reliability of SWE was good (ICC=0.78, 95% CI: 0.67-0.87). Automated and manual methods showed strong agreement (Dice coefficient 0.94) and good reliability (ICC=0.85, 95% CI: 0.75-0.91). RF CSA demonstrated weak but significant correlations with CT-derived skeletal muscle area at both T12 and L3 levels (r=0.37 to 0.40, P<0.05). US parameters showed moderate predictive value for CT-derived skeletal muscle index at L3 (adjusted R²=0.70). In conclusion, RF US measurements are reliable, and automated measurements are feasible but show a modest correlation with CT-derived muscle mass measurements.

To evaluate the use of the Model Predictive Filtering (MPF) method to improve temporal resolution of magnetic resonance temperature imaging (MRTI) for monitoring laser interstitial thermal therapy (LITT) ablations. Using a Green's function method for solving differential equations, a treatment-specific power matrix Q was derived from a LITT heating and used in the Pennes bioheat equation (PBHE) to model a subsequent higher power heating and supplement subsampled k-space data. This MPF method was evaluated using both 3D segmented EPI data and a tissue mimicking phantom and clinical LITT treatment data after retrospective subsampling. Reconstruction accuracy was assessed via thermal dose and analysis of the hottest voxel and region-of-voxels over time. In the phantom data, temporal resolution equivalent to a 12-slice acquisition was produced with larger fields-of-view (24 and 36 slices, R = 2 and 3) with good hottest voxel-over-time accuracy and 240 CEM43 volume agreement (Dice similarity coefficient, DSC 0.7). In the in vivo data, MPF reconstruction showed excellent 240 CEM43 volume agreement for both orthogonal slices (DSC 0.9 for R = 2 and 3). The sagittal and coronal slices showed excellent hottest voxel accuracy for subsampling of R 3, with an RMSE ≤ 1°C. Hottest voxel RMSE remained within 1°C-3°C up to a subsampling factor of 5. The MPF algorithm allowed for large field-of-view (FOV) volumetric temperature imaging without decreasing temporal resolution in phantom heatings. Bi-planar clinical treatment data reconstruction showed good accuracy for the application of MPF to in vivo data.

Towards a shuttle-based workflow for daily online magnetic resonance imaging-guided particle therapy: Anatomical robustness and organ variability in the female pelvis.

This work established and validated a radiomics-driven framework to quantify therapeutic response in multiple myeloma (MM). By leveraging CT-based Delta features, we aimed to delineate sub-visual temporal changes and assess their incremental predictive value beyond conventional clinical biomarkers. We retrospectively evaluated 131 newly diagnosed MM patients (April 2019-November 2024), partitioned into training and validation cohorts (7:3 ratio). Baseline and post-induction CT images (mean interval: 132days), reconstructed via high-resolution sharp kernels, served as the basis for analysis. Following a rigorous inter-observer stability filter (ICC > 0.90), Delta-radiomic features were extracted from the ribs, manubrium, and thoracic spine. A stability-first selection pipeline, utilizing mRMR and LASSO regression, identified the most robust signatures for model construction. ISS staging (I/II/III: 19.9%/29.0%/51.1%) was balanced across cohorts. The multi-regional radiomics signature (RadDeltaAll) demonstrated superior stability compared to single-site models. In the training set, the integrated framework yielded an AUC of 0.855. In the independent validation cohort, while single-site performance varied, the comprehensive RadDeltaAll model maintained high diagnostic stability . AUC=0.842. Notably, the clinical biomarker 24hU-Pro was retained as a "biological anchor"; although it did not statistically enhance predictive accuracy, it provided essential context for systemic renal burden without compromising model integrity. CT-based Delta radiomics constitutes a robust, non-invasive biomarker for MM response assessment. While 24hU-Pro serves as an indispensable biological reference, the multi-regional radiomic signature remains the primary engine of predictive power. This underscores the clinical necessity of quantitative, skeleton-aware analysis in steering individualized induction therapy. MM: Multiple myeloma; CT: Computed tomography; ROI: Region of interest; AUC: Area under the curve; ESM: Electronic supplementary material; WBLDCT: Whole-body low-dose computed tomography; IMWG: International Myeloma Working Group; LASSO: Least Absolute Shrinkage and Selection Operator; mRMR: minimum Redundancy Maximum Relevance; SVM: Support Vector Machine; RF: Random Forest; XGBoost: Extreme Gradient Boosting; GLCM: Gray-Level Co-occurrence Matrix; GLRLM: Gray-Level Run-Length Matrix; GLSZM: Gray-Level Size Zone Matrix; NGTDM: Neighboring Gray-Tone Difference Matrix; DSC: Dice Similarity Coefficient; DCA: Decision Curve Analysis.

To develop a novel transparent and lightweight machine learning model, the Green Learning (GL), for automated prostate segmentation (PS) and clinically significant prostate cancer (csPCa) detection on magnetic resonance imaging (MRI). Men who underwent 3-T MRI and prostate biopsy (PBx) were identified. MRI was acquired and interpreted according to the Prostate Imaging-Reporting and Data System (PI-RADS), version 2 or 2.1. The GL was created to automate PS and csPCa detection on biparametric MRI. The performance was compared to the standard-of-care radiologists using PI-RADS, and a conventional deep learning (DL) U-Net model as benchmarking. The PS performance was evaluated by the Dice similarity coefficient (DSC). The area under the curve (AUC) for patient-level csPCa detection was assessed. Model size and computational workload, measured by floating point operations (FLOPs), were reported. A total of 602 MRIs were randomly divided for training (N = 483) and testing (N = 119). Overall, 224 patients had csPCa on PBx. The median DSC for PS was higher for GL than U-Net (0.91 vs 0.88, P < 0.001). The AUC for csPCa detection of GL was similar to PI-RADS (0.75 vs 0.76, P = 0.8) and U-Net (vs 0.74, P = 0.3). A combination of GL and PI-RADS showed a higher AUC of 0.81 than PI-RADS alone (P = 0.02). Compared with U-Net, the GL had smaller magnitude parameters (1.21× 106 vs 177× 106) and less computational workload (9.8× 109 vs 1027× 109 FLOPs). A novel GL model fully automatically detects csPCa on prostate biparametric MRI with comparable performance to PI-RADS and DL. Combined with PI-RADS, GL significantly improves csPCa detection.

Beyond Human Variability: Deep Learning for Intravascular Ultrasound Segmentation With Noisy Labels.

Delineating the CTVboost in postoperative breast radiation therapy (RT) may be challenging after preoperative systemic therapy (PST), due to treatment-induced anatomical changes and the potential discordance between the postoperative surgical cavity, as identified by surgical clips, and the spatial extent of the pretreatment tumor. The objective of this study was to assess whether [18F]-FDG PET/CT (PET/CT)-guided delineation using deformable image registration (DIR) improves inter-observer reproducibility compared to conventional strategies relying primarily on surgical clips and anatomical landmarks. Fifty-eight patients treated with PST, surgery, and postoperative RT were retrospectively included. Three radiation oncologists performed four delineation strategies per patient: (1) CTVCLI, based on clips and anatomical landmarks; (2) CTVPET, a 10mm isotropic expansion around the PET-defined biological tumor volume (BTV); (3) CTVCOM, generated by a 15-20mm isotropic expansion around the BTV centroid and adapted to pathological features; and (4) CTVINT, manually refined volumes integrating clips, PET signal, and registration accuracy. Inter-observer variability (IOV) was assessed using the Dice similarity coefficient (DSC) and Hausdorff distance (HD). All patients underwent lumpectomy after PST, with a mean age of 53.9 years, and 44.2% had triple-negative tumors. The mean volume of the CTVCLI was 34.4cm3, compared to 36.5cm3 for CTVPET, 21.2cm3 for CTVCOM, and 43.6cm3 for CTVINT. Mean DSCs were 0.86 for both CTVPET and CTVCOM, 0.77 for CTVINT, and 0.62 for CTVCLI (p<10-8). PET/CT-guided delineation after DIR significantly improves inter-observer reproducibility of CTVboost definition after PST compared with conventional methods.

Evaluation of the Effectiveness of an AI-assisted, Dual-Template Workflow in Improving the Accuracy of Tooth Autotransplantation.

This study aimed to evaluate the performance of deep learning models for segmenting the augmented bone following transalveolar sinus floor elevation (TSFE). Cone-beam computed tomography (CBCT) data from 103 patients undergoing TSFE, acquired at preoperative (T0) and immediate postoperative (T1) were retrospectively analysed. Four deep learning models (UNETR++, Swin Transformer, U-Net, 3D-VNet) were trained and validated for segmenting the augmented bone. Performance was assessed using the Dice similarity coefficient (DSC), intersection over union (IoU), sensitivity, precision, 95% Hausdorff Distance (HD95), and accuracy. UNETR++ demonstrated the best performance, with an average DSC of 0.8477, IoU of 0.7356, sensitivity of 0.8337, precision of 0.8622, HD95 of 0.9234 mm, and accuracy of 0.8730. UNETR++ segmentations exhibited excellent reproducibility compared with manual segmentation. The automated segmentation process significantly reduced measurement time to 14.96 ± 2.57 seconds. Deep learning models, particularly UNETR++, provide an accurate and efficient method for segmenting augmented bone after TSFE.

BIBSNet: A deep learning baby image brain segmentation network for MRI scans.

Semantic token-guided hierarchical adversarial knowledge distillation for 3D abdominal organ segmentation.

Ultrasound image segmentation remains a significant challenge due to inherent low contrast and blurred anatomical boundaries. Fully supervised deep learning approaches require extensive annotated datasets, which are costly and labor-intensive to acquire. This study aims to develop an effective semi-supervised segmentation framework for ultrasound images with limited annotations. We propose a novel semi-supervised segmentation framework tailored for ultrasound images, leveraging frequency component augmentation and edge mask enhancement to promote structural consistency between weakly and strongly augmented inputs. Specifically, discrete wavelet transform (DWT) is used to decompose ultrasound images into low-frequency and high-frequency sub-bands. A high-frequency component replacement strategy is introduced for strongly augmented images, and an edge mask enhancement module is designed to further emphasize anatomical boundaries. Experiments conducted on 3 public fetal ultrasound imaging segmentation datasets-PSFHS, HC18, and CCAUI-demonstrate that our method achieves average Dice similarity coefficients (DSC) of 0.81 and 0.91, respectively, using only 10 annotated images. This represents a 2-3% DSC improvement over existing semi-supervised methods such as FixMatch. Ablation studies confirm the effectiveness of both the high-frequency augmentation and edge enhancement components. The proposed framework offers a promising direction for ultrasound image segmentation in settings with limited annotations, effectively improving segmentation accuracy by combining frequency-domain augmentation and edge-aware enhancement. Code will be available at https://github.com/apple1986/WTEM-SemiSeg.

PDAFormer 3+: A full-scale connected modified transformer with parallel dual attention for 3D medical image segmentation.

cervical nuclei segmentation through synergic conditional generative adversarial network in cervical smear images.

Fully Automated Deep Learning-Based Pipeline for Evans Index Measurement from Raw 3D MRI.