Manual Segmentation Research Articles

A Practical Guide to Manual and Semi-Automated Neurosurgical Brain Lesion Segmentation.

The medical images used were sourced from the Medical Image Computing and Computer Assisted Interventions Society (MICCAI) Multimodal Brain Tumour Segmentation Challenge (BRATS) image database and from the local Picture Archival and Communication System (PACS) record with consent. Image pre-processing was carried out using MRIcron software (v1.0.20190902). ITK-SNAP (v3.8.0) was used in this guideline due to its availability and powerful built-in segmentation tools, although others (Seg3D, Freesurfer and 3D Slicer) are available. Quality control was achieved by employing expert segmenters to review. A pipeline was developed to demonstrate the pre-processing and manual and semi-automated segmentation of patient images for each cranial lesion, accompanied by image guidance and video recordings. Three sample segmentations were generated to illustrate potential challenges. Advice and solutions were provided within both text and video. Semi-automated segmentation methods enhance efficiency, increase reproducibility, and are suitable to be incorporated into future clinical practise. However, manual segmentation remains a highly effective technique in specific circumstances and provides initial training sets for the development of more advanced semi- and fully automated segmentation algorithms.

Automatic Segmentation of Heschl Gyrus and Planum Temporale by MRICloud.

This study used a cloud-based program, MRICloud, which parcellates T1 MRI brain scans using a probabilistic classification based on manually labeled multi-atlas, to create a tool to segment Heschl gyrus (HG) and the planum temporale (PT). MRICloud is an online platform that can automatically segment structural MRIs into 287 labeled brain regions. A 31-brain multi-atlas was manually resegmented to include tags for the HG and PT. This modified atlas set with additional manually labeled regions of interest acted as a new multi-atlas set and was uploaded to MRICloud. This new method of automated segmentation of HG and PT was then compared to manual segmentation of HG and PT in MRIs of 10 healthy adults using Dice similarity coefficient (DSC), Hausdorff distance (HD), and intraclass correlation coefficient (ICC). This multi-atlas set was uploaded to MRICloud for public use. When compared to reference manual segmentations of the HG and PT, there was an average DSC for HG and PT of 0.62 ± 0.07, HD of 8.10 ± 3.47 mm, and an ICC for these regions of 0.83 (0.68-0.91), consistent with an appropriate automatic segmentation accuracy. This multi-atlas can alleviate the manual segmentation effort and the difficulty in choosing an HG and PT anatomical definition. This protocol is limited by the morphology of the MRI scans needed to make the MRICloud atlas set. Future work will apply this multi-atlas to observe MRI changes in hearing-associated disorders.

Landmark registration in CT for lung model approximation in EIT

Abstract For the visualization of pulmonary ventilation with Electrical Impedance Tomography (EIT), most devices rely on generalized reconstruction models. Yet, fixed thorax dimensions, predetermined electrode locations, and standard lung shapes lead to multiple sources of EIT imaging distortion. The following work compares reconstructions of a library model, a practical model based on landmark fitting, and a detailed model based on manual segmentation. CT images of five pigs were segmented into torso surface, electrode positions, and lung borders. Practical models were created with reduced data, registration, and fitting. Detailed models were created by using the complete segmentation data. After EIT reconstruction and tidal image calculation, the overlap of CT lung segmentation and ventilated voxel was calculated. Compared with the results of a standardized model, the practical model reached an average ventilation/segmentation-overlap difference of an additional 12 %, and the detailed model achieved an average difference of 10 %. In conclusion, the shift of reconstructed impedance towards the segmented lung region is similar in both models when compared to the standard model, while the practical model requires a considerably less amount of information.

Automatic substantia nigra segmentation with Swin-Unet in susceptibility- and T2-weighted imaging: application to Parkinson disease diagnosis.

Accurately distinguishing between Parkinson disease (PD) and healthy controls (HCs) through reliable imaging method is crucial for appropriate therapeutic intervention. However, PD diagnosis is hindered by the subjective nature of the evaluation. We aimed to develop an automatic deep-learning method that can segment the substantia nigra areas on susceptibility-weighted imaging (SWI) and T2-weighted imaging (T2WI) and further differentiate patients with PD from HCs using a machine learning algorithm. Magnetic resonance imaging (MRI) data from 83 patients with PD and 83 age- and sex-matched HCs were obtained on the same 3.0-T MRI scanner. A deep learning method with Swin-Unet was developed to segment volumes of interest (VOIs) on SWI and then map the VOIs on SWI to the corresponding T2WI; features were then extracted from the VOIs on SWI and T2WI. Three machine learning models were developed and compared to differentiate those with PD from HCs. Swin-Unet achieved a better Dice coefficient than did U-Net in SWI segmentation (0.832 vs. 0.712). Machine learning models outperformed visual analysis (P>0.05), and logistic regression (LR) achieved the best performance [area under the curve (AUC) ≥0.819] and the most stable (relative standard deviations in AUC ≤0.05). The test results showed that the AUC of the LR model based on SWI segmentation was 0.894 while that of the LR model based on T2WI segmentation was 0.876. There was no significant difference in VOIs based on manual labeling or automatic segmentation across T2WI, SWI, or a combination of the two (P>0.05). The AUCs of the LR model based on automatic segmentation were close to those of the model based on manual labeling (P>0.05). Our approach could provide a powerful and useful method for automatically and rapidly diagnosing PD in the clinic with only T2WI.

Deep Location Soft-Embedding-Based Network With Regional Scoring for Mammogram Classification.

Early detection and treatment of breast cancer can significantly reduce patient mortality, and mammogram is an effective method for early screening. Computer-aided diagnosis (CAD) of mammography based on deep learning can assist radiologists in making more objective and accurate judgments. However, existing methods often depend on datasets with manual segmentation annotations. In addition, due to the large image sizes and small lesion proportions, many methods that do not use region of interest (ROI) mostly rely on multi-scale and multi-feature fusion models. These shortcomings increase the labor, money, and computational overhead of applying the model. Therefore, a deep location soft-embedding-based network with regional scoring (DLSEN-RS) is proposed. DLSEN-RS is an end-to-end mammography image classification method containing only one feature extractor and relies on positional embedding (PE) and aggregation pooling (AP) modules to locate lesion areas without bounding boxes, transfer learning, or multi-stage training. In particular, the introduced PE and AP modules exhibit versatility across various CNN models and improve the model's tumor localization and diagnostic accuracy for mammography images. Experiments are conducted on published INbreast and CBIS-DDSM datasets, and compared to previous state-of-the-art mammographic image classification methods, DLSEN-RS performed satisfactorily.

Presurgical Upgrade Prediction of DCIS to Invasive Ductal Carcinoma Using Time-dependent Deep Learning Models with DCE MRI.

Purpose To determine whether time-dependent deep learning models can outperform single time point models in predicting preoperative upgrade of ductal carcinoma in situ (DCIS) to invasive malignancy at dynamic contrast-enhanced (DCE) breast MRI without a lesion segmentation prerequisite. Materials and Methods In this exploratory study, 154 cases of biopsy-proven DCIS (25 upgraded at surgery and 129 not upgraded) were selected consecutively from a retrospective cohort of preoperative DCE MRI in women with a mean age of 59 years at time of diagnosis from 2012 to 2022. Binary classification was implemented with convolutional neural network (CNN)-long short-term memory (LSTM) architectures benchmarked against traditional CNNs without manual segmentation of the lesions. Combinatorial performance analysis of ResNet50 versus VGG16-based models was performed with each contrast phase. Binary classification area under the receiver operating characteristic curve (AUC) was reported. Results VGG16-based models consistently provided better holdout test AUCs than did ResNet50 in CNN and CNN-LSTM studies (multiphase test AUC, 0.67 vs 0.59, respectively, for CNN models [P = .04] and 0.73 vs 0.62 for CNN-LSTM models [P = .008]). The time-dependent model (CNN-LSTM) provided a better multiphase test AUC over single time point (CNN) models (0.73 vs 0.67; P = .04). Conclusion Compared with single time point architectures, sequential deep learning algorithms using preoperative DCE MRI improved prediction of DCIS lesions upgraded to invasive malignancy without the need for lesion segmentation. Keywords: MRI, Dynamic Contrast-enhanced, Breast, Convolutional Neural Network (CNN) Supplemental material is available for this article. © RSNA, 2024.

Radiomics machine learning algorithm facilitates detection of small pancreatic neuroendocrine tumors on CT

PurposeThe purpose of this study was to develop a radiomics-based algorithm to identify small pancreatic neuroendocrine tumors (PanNETs) on CT and evaluate its robustness across manual and automated segmentations, exploring the feasibility of automated screening. Materials and methodsPatients with pathologically confirmed T1 stage PanNETs and healthy controls undergoing dual-phase CT imaging were retrospectively identified. Manual segmentation of pancreas and tumors was performed, then automated pancreatic segmentations were generated using a pretrained neural network. A total of 1223 radiomics features were independently extracted from both segmentation volumes, in the arterial and venous phases separately. Ten final features were selected to train classifiers to identify PanNETs and controls. The cohort was divided into training and testing sets, and performance of classifiers was assessed using area under the receiver operator characteristic curve (AUC), specificity and sensitivity, and compared against two radiologists blinded to the diagnoses. ResultsA total of 135 patients with 142 PanNETs, and 135 healthy controls were included. There were 168 women and 102 men, with a mean age of 55.4 ± 11.6 (standard deviation) years (range: 20–85 years). Median PanNET size was 1.3 cm (Q1, 1.0; Q3, 1.5; range: 0.5–1.9). The arterial phase LightGBM model achieved the best performance in the test set, with 90 % sensitivity (95 % confidence interval [CI]: 80–98), 76 % specificity (95 % CI: 62–88) and an AUC of 0.87 (95 % CI: 0.79–0.94). Using features from the automated segmentations, this model achieved an AUC of 0.86 (95 % CI: 0.79–0.93). In comparison, the two radiologists achieved a mean 50 % sensitivity and 100 % specificity using arterial phase CT images. ConclusionRadiomics features identify small PanNETs, with stable performance when extracted using automated segmentations. These models demonstrate high sensitivity, complementing the high specificity of radiologists, and could serve as opportunistic screeners.

Enhanced TumorNet: Leveraging YOLOv8s and U-net for superior brain tumor detection and segmentation utilizing MRI scans

Brain tumors, characterized by abnormal cell growth, pose a significant challenge in clinical imaging due to their complex and diverse structures. Early and accurate identification, classification, localization, and segmentation of these tumors are critical to reducing mortality. However, the extensive data generated by MRI scans makes manual segmentation time-consuming and impractical for clinical use. To address these challenges, we propose, a hybrid deep learning model that precisely segmented tumor regions with U-Net to enable YOLOv8s to efficiently detect, classify and localize tumors. The model was trained and validated using The Cancer Imaging Archive (TCIA) dataset, which includes MRI images of brain tumors, and the Cancer Genome Atlas (CGA) low-grade glioma dataset, which includes data from 110 patients with FLAIR aberrant segmentation masks. The proposed hybrid model was evaluated using several performance metrics, including F1 score, specificity, recall, precision, accuracy, and ROC-AUC score. Hybrid proposed performed highly, achieving a precision of 97.8 %, accuracy of 98.6 %, recall of 95.2 %, F-1 score of 96.3 %, specificity of 89.1 %, and ROC-AUC score of 98.5 %. The integration of YOLOv8s and U-Net in Enhanced TumorNet offers a powerful solution for the automated analysis of brain tumors in MRI scans, significantly improving detection and segmentation accuracy. This hybrid approach holds great potential for clinical applications, enhancing the efficiency and effectiveness of brain tumor diagnosis.

Accurate automated segmentation of autophagic bodies in yeast vacuoles using cellpose 2.0

ABSTRACT Segmenting autophagic bodies in yeast TEM images is a key technique for measuring changes in autophagosome size and number in order to better understand macroautophagy/autophagy. Manual segmentation of these images can be very time consuming, particularly because hundreds of images are needed for accurate measurements. Here we describe a validated Cellpose 2.0 model that can segment these images with accuracy comparable to that of human experts. This model can be used for fully automated segmentation, eliminating the need for manual body outlining, or for model-assisted segmentation, which allows human oversight but is still five times as fast as the current manual method. The model is specific to segmentation of autophagic bodies in yeast TEM images, but researchers working in other systems can use a similar process to generate their own Cellpose 2.0 models to attempt automated segmentations. Our model and instructions for its use are presented here for the autophagy community. Abbreviations: AB, autophagic body; AvP, average precision; GUI, graphical user interface; IoU, intersection over union; MVB, multivesicular body; ROI, region of interest; TEM, transmission electron microscopy; WT,wild type.

LungPath: artificial intelligence-driven histologic pattern recognition for improved diagnosis of early-stage invasive lung adenocarcinoma.

Early-stage invasive lung adenocarcinoma (ADC) characterized by a predominant micropapillary or solid pattern exhibit an elevated risk of recurrence following sub-lobar resection, thus determining histological subtype of early-stage invasive ADC prior surgery is important for formulating lobectomy or sub-lobar resection. This study aims to develop a deep learning algorithm and assess its clinical capability in distinguishing high-risk or low-risk histologic patterns in early-stage invasive ADC based on preoperative computed tomography (CT) scans. Two retrospective cohorts were included: development cohort 1 and external test cohort 2, comprising patients diagnosed with T1 stage invasive ADC. Electronic medical records and CT scans of all patients were documented. Patients were stratified into two risk groups. High-risk group: comprising cases with a micropapillary component ≥5% or a predominant solid pattern. Low-risk group: encompassing cases with a micropapillary component <5% and an absence of a predominant solid pattern. The overall segmentation model was modified based on Mask Region-based Convolutional Neural Network (Mask-RCNN), and Residual Network 50 (ResNet50)_3D was employed for image classification. A total of 432 patients participated in this study, with 385 cases in cohort 1 and 47 cases in cohort 2. The fine-outline results produced by the auto-segmentation model exhibited a high level of agreement with manual segmentation by human experts, yielding a mean dice coefficient of 0.86 [95% confidence interval (CI): 0.85-0.87] in cohort 1 and 0.84 (95% CI: 0.82-0.85) in cohort 2. Furthermore, the deep learning model effectively differentiated the high-risk group from the low-risk group, achieving an area under the curve (AUC) of 0.89 (95% CI: 0.88-0.90) in cohort 1. In the external validation conducted in cohort 2, the deep learning model displayed an AUC of 0.87 (95% CI: 0.84-0.88) in distinguishing the high-risk group from the low-risk group. The average diagnostic time was 16.00±3.2 seconds, with an accuracy of 0.82 (95% CI: 0.81-0.83). We have developed a deep learning algorithm, LungPath, for the automated segmentation of pulmonary nodules and prediction of high-risk histological patterns in early-stage lung ADC based on CT scans.

Intelligent void identification of particle packing system of caved ore and rock

Analyzing the void structure of the packing system of caved ore and rock and establishing the causal relationship between void structure and macroscopic properties are critical research areas in the field of gravity flow. However, the current method of analyzing void structure primarily involves manual threshold segmentation, leading to low analysis efficiency and the need for improved accuracy. This study introduces a novel approach using a dataset of two-dimensional Computed Tomography (CT) slices of an irregular limestone particle packing system. The primary goal is to enhance existing algorithms by focusing on data and network aspects to address the challenges in identifying this type of image dataset. As a result, a model named Data Attention gate with Recurrent Residual convolutional neural network based on U-net (DA2RU-net) is developed and evaluated for its convergence, accuracy, and robustness. The findings indicate the following: (1) To identify voids in the packing system of caved ore and rock using CT slices, it is advisable to employ 1000 images with dimensions between 300 and 400 pixels. It is essential to ensure that both large and small voids are represented in the image dataset. (2) The DA2RU-net model demonstrates superior performance with an average Dice value of 0.9859 for identifying large voids and 0.9654 for identifying small voids, surpassing other iterations of the U-net model and traditional algorithms. (3) The DA2RU-net model shows robustness to variations in brightness levels.

Development and validation of a prediction model for malignant sinonasal tumors based on MR radiomics and machine learning.

This study aimed to utilize MR radiomics-based machine learning classifiers on a large-sample, multicenter dataset to develop an optimal model for predicting malignant sinonasal tumors and tumor-like lesions. This study included 1711 adult patients (875 benign and 836 malignant) with sinonasal tumors or tumor-like lesions from three institutions. Patients from institution 1 (n = 1367) constituted both the training and validation cohorts, while those from institution 2 and 3 (n = 158/186) made up the test cohorts. Manual segmentation of the region of interest of the tumor was performed on T1WI, T2WI, and contrast-enhanced T1WI (CE-T1WI). Data normalization, dimensional reductions, feature selection, and classifications were performed using ten machine-learning classifiers. Four fusion models, namely T1WI + T2WI, T1WI + CE-T1WI, T2WI + CE-T1WI, and T1WI + T2WI + CE-T1WI, were constructed using the top ten features with the highest contribution in feature selection in the optimal models of T1WI, T2WI, and CE-T1WI. The Delong test compared areas under the curve (AUC) between models. The AUCs of training/validation/test1/test2 datasets for T1WI, T2WI, and CE-T1WI were 0.900/0.842/0.872/0.839, 0.876/0.789/0.842/0.863, and 0.899/0.824/0.831/0.707, respectively. The fusion model from T1WI + T2WI + CE-T1WI had the highest AUC. The AUCs of training/validation/test1/test2 datasets were 0.947/0.849/0.871/0.887. The T1WI + T2WI + CE-T1WI model demonstrated a significantly higher AUC than the T2WI + CE-T1WI model in both cohorts (p < 0.05) and outperformed the T2WI model in test 1 (p = 0.008) and the T1WI model in test 2 (p = 0.006). This fusion model based on radiomics from T1WI + T2WI + CE-T1WI images and machine learning can improve the power in predicting malignant sinonasal tumors with high accuracy, resilience, and robustness. Our study proposes a radiomics-based machine learning fusion model from T1- and T2-weighted images and contrast-enhanced T1-weighted images, which can non-invasively identify the nature of sinonasal tumors and improve the performance in predicting malignant sinonasal tumors. Differentiating benign and malignant sinonasal tumors is difficult due to similar clinical presentations. A radiomics model from T1 + T2 + contrast-enhanced T1 images can identify the nature of sinonasal tumors. This model can help distinguish benign and malignant sinonasal tumors.

Deep learning to predict risk of lateral skull base cerebrospinal fluid leak or encephalocele.

Skull base features, including increased foramen ovale (FO) cross-sectional area, are associated with lateral skull base spontaneous cerebrospinal fluid (sCSF) leak and encephalocele. Manual measurement requires skill in interpreting imaging studies and is time consuming. The goal of this study was to develop a fully automated deep learning method for FO segmentation and to determine the predictive value in identifying patients with sCSF leak or encephalocele. A retrospective cohort study at a tertiary care academic hospital of 34 adults with lateral skull base sCSF leak or encephalocele were compared with 815 control patients from 2013-2021. A convolutional neural network (CNN) was constructed for image segmentation of axial computed tomography (CT) studies. Predicted FO segmentations were compared to manual segmentations, and receiver operating characteristic (ROC) curves were constructed. 295 CTs were used for training and validation of the CNN. A separate dataset of 554 control CTs was matched 5:1 on age and sex with the sCSF leak/encephalocele group. The mean Dice score was 0.81. The sCSF leak/encephalocele group had greater mean (SD) FO cross-sectional area compared to the control group, 29.0 (7.7) mm2 versus 24.3 (7.6) mm2 (P = .002, 95% confidence interval 0.02-0.08). The area under the ROC curve was 0.69. CNNs can be used to segment the cross-sectional area of the FO accurately and efficiently. Used together with other predictors, this method could be used as part of a clinical tool to predict the risk of sCSF leak or encephalocele.

High-resolution segmentations of the hypothalamus and its subregions for training of segmentation models

Segmentation of brain structures on magnetic resonance imaging (MRI) is a highly relevant neuroimaging topic, as it is a prerequisite for different analyses such as volumetry or shape analysis. Automated segmentation facilitates the study of brain structures in larger cohorts when compared with manual segmentation, which is time-consuming. However, the development of most automated methods relies on large and manually annotated datasets, which limits the generalizability of these methods. Recently, new techniques using synthetic images have emerged, reducing the need for manual annotation. Here we provide a dataset composed of label maps built from publicly available ultra-high resolution ex vivo MRI from 10 whole hemispheres, which can be used to develop segmentation methods using synthetic data. The label maps are obtained with a combination of manual labels for the hypothalamic regions and automated segmentations for the rest of the brain, and mirrored to simulate entire brains. We also provide the pre-processed ex vivo scans, as this dataset can support future projects to include other structures after these are manually segmented.

Automated analysis of scattering-based light sheet microscopy images of anal squamous intraepithelial lesions.

We developed an algorithm for automatically analyzing scattering-based light sheet microscopy (sLSM) images of anal squamous intraepithelial lesions. We developed a method for automatically segmenting sLSM images for nuclei and calculating seven features: nuclear intensity, intensity slope as a function of depth, nuclear-to-nuclear distance, nuclear-to-cytoplasm ratio, cell density, nuclear area, and proportion of pixels corresponding to nuclei. 187 images from 80 anal biopsies were used for feature analysis and classifier development. The automated nuclear segmentation method provided reliable performance with the precision of 0.97 and recall of 0.91 when compared with the manual segmentation. Among the seven features, six showed statistically significant differences between high-grade squamous intraepithelial lesion (HSIL) and non-HSIL (non-dysplastic or low-grade squamous intraepithelial lesion, LSIL). A classifier using linear support vector machine (SVM) achieved promising performance in diagnosing HSIL versus non-HSIL: sensitivity of 90%, specificity of 70%, and area under the curve (AUC) of 0.89 for per-image diagnosis, and sensitivity of 90%, specificity of 80%, and AUC of 0.92 for per-biopsy diagnosis.

Vasculature segmentation in 3D hierarchical phase-contrast tomography images of human kidneys.

Efficient algorithms are needed to segment vasculature in new three-dimensional (3D) medical imaging datasets at scale for a wide range of research and clinical applications. Manual segmentation of vessels in images is time-consuming and expensive. Computational approaches are more scalable but have limitations in accuracy. We organized a global machine learning competition, engaging 1,401 participants, to help develop new deep learning methods for 3D blood vessel segmentation. This paper presents a detailed analysis of the top-performing solutions using manually curated 3D Hierarchical Phase-Contrast Tomography datasets of the human kidney, focusing on the segmentation accuracy and morphological analysis, thereby establishing a benchmark for future studies in blood vessel segmentation within phase-contrast tomography imaging.

Innovative Detection and Segmentation of Mobility Activities in Patients Living with Parkinson's Disease Using a Single Ankle-Positioned Smartwatch.

The automatic detection of activities of daily living (ADL) is necessary to improve long-term home-based monitoring of Parkinson's disease (PD) symptoms. While most body-worn sensor algorithms for ADL detection were developed using laboratory research systems covering full-body kinematics, it is now crucial to achieve ADL detection using a single body-worn sensor that remains commercially available and affordable for ecological use. to detect and segment Walking, Turning, Sitting-down, and Standing-up activities of patients with PD using a Smartwatch positioned at the ankle. Twenty-two patients living with PD performed a Timed Up and Go (TUG) task three times before engaging in cleaning ADL in a simulated free-living environment during a 3 min trial. Accelerations and angular velocities of the right or left ankle were recorded in three dimensions using a Smartwatch. The TUG task was used to develop detection algorithms for Walking, Turning, Sitting-down, and Standing-up, while the 3 min trial in the free-living environment was used to test and validate these algorithms. Sensitivity, specificity, and F-scores were calculated based on a manual segmentation of ADL. Sensitivity, specificity, and F-scores were 96.5%, 94.7%, and 96.0% for Walking; 90.0%, 93.6%, and 91.7% for Turning; 57.5%, 70.5%, and 52.3% for Sitting-down; and 57.5%, 72.9%, and 54.1% for Standing-up. The median of time difference between the manual and automatic segmentation was 1.31 s for Walking, 0.71 s for Turning, 2.75 s for Sitting-down, and 2.35 s for Standing-up. The results of this study demonstrate that segmenting ADL to characterize the mobility of people with PD based on a single Smartwatch can be comparable to manual segmentation while requiring significantly less time. While Walking and Turning were well detected, Sitting-down and Standing-up will require further investigation to develop better algorithms. Nonetheless, these achievements increase the odds of success in implementing wearable technologies for PD monitoring in ecological environments.

Renal tumor segmentation, visualization, and segmentation confidence using ensembles of neural networks in patients undergoing surgical resection.

To develop an automatic segmentation model for solid renal tumors on contrast-enhanced CTs and to visualize segmentation with associated confidence to promote clinical applicability. The training dataset included solid renal tumor patients from two tertiary centers undergoing surgical resection and receiving CT in the corticomedullary or nephrogenic contrast media (CM) phase. Manual tumor segmentation was performed on all axial CT slices serving as reference standard for automatic segmentations. Independent testing was performed on the publicly available KiTS 2019 dataset. Ensembles of neural networks (ENN, DeepLabV3) were used for automatic renal tumor segmentation, and their performance was quantified with DICE score. ENN average foreground entropy measured segmentation confidence (binary: successful segmentation with DICE score > 0.8 versus inadequate segmentation ≤ 0.8). N = 639/n = 210 patients were included in the training and independent test dataset. Datasets were comparable regarding age and sex (p > 0.05), while renal tumors in the training dataset were larger and more frequently benign (p < 0.01). In the internal test dataset, the ENN model yielded a median DICE score = 0.84 (IQR: 0.62-0.97, corticomedullary) and 0.86 (IQR: 0.77-0.96, nephrogenic CM phase), and the segmentation confidence an AUC = 0.89 (sensitivity = 0.86; specificity = 0.77). In the independent test dataset, the ENN model achieved a median DICE score = 0.84 (IQR: 0.71-0.97, corticomedullary CM phase); and segmentation confidence an accuracy = 0.84 (sensitivity = 0.86 and specificity = 0.81). ENN segmentations were visualized with color-coded voxelwise tumor probabilities and thresholds superimposed on clinical CT images. ENN-based renal tumor segmentation robustly performs in external test data and might aid in renal tumor classification and treatment planning. Ensembles of neural networks (ENN) models could automatically segment renal tumors on routine CTs, enabling and standardizing downstream image analyses and treatment planning. Providing confidence measures and segmentation overlays on images can lower the threshold for clinical ENN implementation. Ensembles of neural networks (ENN) segmentation is visualized by color-coded voxelwise tumor probabilities and thresholds. ENN provided a high segmentation accuracy in internal testing and in an independent external test dataset. ENN models provide measures of segmentation confidence which can robustly discriminate between successful and inadequate segmentations.

Neural network auto-segmentation of serial-block-face scanning electron microscopy images exhibit collagen fibril structural differences with tendon type and health.

A U-Net machine learning algorithm was adapted to automatically segment tendon collagen fibril cross-sections from serial block face scanning electron microscopy (SBF-SEM) and create three-dimensional (3D) renderings. We compared the performance of routine Otsu thresholding and U-Net for a positional tendon that has low fibril density (rat tail tendon), an energy-storing tendon that has high fibril density (rat plantaris tendon), and a high fibril density tendon hypothesized to have disorganized 3D ultrastructure (degenerated rat plantaris tendon). The area segmentation of the tail and healthy plantaris tendon had excellent accuracy for both the Otsu and U-Net, with an Intersection over Union (IoU) of 0.8. With degeneration, only the U-Net could accurately segment the area, whereas Otsu IoU was only 0.45. For boundary validation, the U-Net outperformed Otsu segmentation for all tendons. The fibril diameter from U-Net was within 10% of the manual segmentation, however, the Otsu underestimated the fibril diameter by 39% in healthy plantaris and by 84% in the degenerated plantaris. Fibril geometry was averaged across the entire image stack and compared across tendon types. The tail had a lower fibril area fraction (58%) and larger fibril diameter (0.31 µm) than the healthy plantaris (67% and 0.21 µm) and degenerated plantaris tendon (66% and 0.19 µm). This method can be applied to a large variety of tissues to quantify 3D collagen fibril structure.

Inter-visit and inter-reader reproducibility of multi-parametric diffusion-weighted MR imaging in longitudinally imaged patients with metabolic dysfunction-associated fatty liver disease and healthy volunteers.

BackgroundDespite the widespread use of diffusion-weighted imaging (DWI) in metabolic dysfunction-associated fatty liver disease (MAFLD), MRI acquisition and quantification techniques vary in the literature suggesting the need for established and reproducible protocols. The goal of this study was to assess inter-visit and inter-reader reproducibility of DWI- and IVIM-derived parameters in patients with MAFLD and healthy volunteers using extensive sampling of the “fast” compartment, non-rigid registration, and exclusion voxels with poor fit quality. MethodsFrom June 2019 to April 2023, 31 subjects (20 patients with biopsy-proven MAFLD and 11 healthy volunteers) were included in this IRB-approved study. Subjects underwent MRI examinations twice within 40 days. 3.0 T DWI was acquired using a respiratory-triggered spin-echo diffusion-weighted echo-planar imaging sequence (b-values of 0, 10, 20, 30, 40, 50, 100, 200, 400, 800 s/mm2). DWI series were co-registered prior to voxel-wise non-linear regression of the IVIM model and voxels with poor fit quality were excluded (normalized root mean squared error ≥ 0.05). IVIM parameters (perfusion fraction, f; diffusion coefficient, D; and pseudo-diffusion coefficient, D*), and apparent diffusion coefficients (ADC) were computed from manual segmentation of the right liver lobe performed by two analysts on two MRI examinations. ResultsAll results are reported for f, D, D*, and ADC respectively. For inter-reader agreement on the first visit, ICC were of 0.985, 0.994, 0.986, and 0.993 respectively. For intra-reader agreement of analyst 1 assessed on both imaging examinations, ICC between visits were of 0.805, 0.759, 0.511, and 0.850 respectively. For inter-reader agreement on the first visit, mean bias and 95 % limits of agreement were (0.00 ± 0.03), (−0.01 ± 0.03) × 10−3 mm2/s, (0.70 ± 10.40) × 10−3 mm2/s, and (−0.02 ± 0.04) × 10−3 mm2/s respectively. For intra-reader agreement of analyst 1, mean bias and 95 % limits of agreement were (0.01 ± 0.09) × 10−3 mm2/s, (−0.01 ± 0.21) × 10−3 mm2/s, (−13.37 ± 56.19) × 10−3 mm2/s, and (−0.01 ± 0.16) × 10−3 mm2/s respectively. Except for parameter D* that was associated with between-subjects parameter variability (P = 0.009), there was no significant variability between subjects, examinations, or readers. ConclusionWith our approach, IVIM parameters f, D, D*, and ADC provided excellent inter-reader agreement and good to very good inter-visit or intra-reader agreement, thus showing the reproducibility of IVIM-DWI of the liver in MAFLD patients and volunteers.