Enhancing brain image quality with 3D U-net for stripe removal in light sheet fluorescence microscopy.

Light Sheet Fluorescence Microscopy (LSFM) is increasingly popular in neuroimaging for its ability to capture high-resolution 3D neural data. However, the presence of stripe noise significantly degrades image quality, particularly in complex 3D stripes with varying widths and brightness, posing challenges in neuroscience research. Existing stripe removal algorithms excel in suppressing noise and preserving details in 2D images with simple stripes but struggle with the complexity of 3D stripes. To address this, we propose a novel 3D U-net model for Stripe Removal in Light sheet fluorescence microscopy (USRL). This approach directly learns and removes stripes in 3D space across different scales, employing a dual-resolution strategy to effectively handle stripes of varying complexities. Additionally, we integrate a nonlinear mapping technique to normalize high dynamic range and unevenly distributed data before applying the stripe removal algorithm. We validate our method on diverse datasets, demonstrating substantial improvements in peak signal-to-noise ratio (PSNR) compared to existing algorithms. Moreover, our algorithm exhibits robust performance when applied to real LSFM data. Through extensive validation experiments, both on test sets and real-world data, our approach outperforms traditional methods, affirming its effectiveness in enhancing image quality. Furthermore, the adaptability of our algorithm extends beyond LSFM applications to encompass other imaging modalities. This versatility underscores its potential to enhance image usability across various research disciplines.

An end-to-end deep learning pipeline to derive blood input with partial volume corrections for automated parametric brain PET mapping.

Dynamic 2-[18F] fluoro-2-deoxy-D-glucose positron emission tomography (dFDG-PET) for human brain imaging has considerable clinical potential, yet its utilization remains limited. A key challenge in the quantitative analysis of dFDG-PET is characterizing a patient-specific blood input function, traditionally reliant on invasive arterial blood sampling. This research introduces a novel approach employing non-invasive deep learning model-based computations from the internal carotid arteries (ICA) with partial volume (PV) corrections, thereby eliminating the need for invasive arterial sampling. We present an end-to-end pipeline incorporating a 3D U-Net based ICA-net for ICA segmentation, alongside a Recurrent Neural Network (RNN) based MCIF-net for the derivation of a model-corrected blood input function (MCIF) with PV corrections. The developed 3D U-Net and RNN was trained and validated using a 5-fold cross-validation approach on 50 human brain FDG PET scans. The ICA-net achieved an average Dice score of 82.18% and an Intersection over Union of 68.54% across all tested scans. Furthermore, the MCIF-net exhibited a minimal root mean squared error of 0.0052. The application of this pipeline to ground truth data for dFDG-PET brain scans resulted in the precise localization of seizure onset regions, which contributed to a successful clinical outcome, with the patient achieving a seizure-free state after treatment. These results underscore the efficacy of the ICA-net and MCIF-net deep learning pipeline in learning the ICA structure's distribution and automating MCIF computation with PV corrections. This advancement marks a significant leap in non-invasive neuroimaging.

Improved microvascular imaging with optical coherence tomography using 3D neural networks and a channel attention mechanism.

Skin microvasculature is vital for human cardiovascular health and thermoregulation, but its imaging and analysis presents significant challenges. Statistical methods such as speckle decorrelation in optical coherence tomography angiography (OCTA) often require multiple co-located B-scans, leading to lengthy acquisitions prone to motion artefacts. Deep learning has shown promise in enhancing accuracy and reducing measurement time by leveraging local information. However, both statistical and deep learning methods typically focus solely on processing individual 2D B-scans, neglecting contextual information from neighbouring B-scans. This limitation compromises spatial context and disregards the 3D features within tissue, potentially affecting OCTA image accuracy. In this study, we propose a novel approach utilising 3D convolutional neural networks (CNNs) to address this limitation. By considering the 3D spatial context, these 3D CNNs mitigate information loss, preserving fine details and boundaries in OCTA images. Our method reduces the required number of B-scans while enhancing accuracy, thereby increasing clinical applicability. This advancement holds promise for improving clinical practices and understanding skin microvascular dynamics crucial for cardiovascular health and thermoregulation.

3D U-Net Neural Network Architecture-Assisted LDCT to Acquire Vertebral Morphology Parameters: A Vertebral Morphology Comprehensive Analysis in a Chinese Population.

To evaluate the feasibility of acquiring vertebral height from chest low-dose computed tomography (LDCT) images using an artificial intelligence (AI) system based on 3D U-Net vertebral segmentation technology and the correlation and features of vertebral morphology with sex and age of the Chinese population. Patients who underwent chest LDCT between September 2020 and April 2023 were enrolled. The Altman and Pearson's correlation analyses were used to compare the correlation and consistency between the AI software and manual measurement of vertebral height. The anterior height (Ha), middle height (Hm), posterior height (Hp), and vertebral height ratios (VHRs) (Ha/Hp and Hm/Hp) were measured from T1 to L2 using an AI system. The VHR is the ratio of Ha to Hp or the ratio of Hm to Hp of the vertebrae, which can reflect the shape of the anterior wedge and biconcave vertebrae. Changes in these parameters, particularly the VHR, were analysed at different vertebral levels in different age and sex groups. The results of the AI methods were highly consistent and correlated with manual measurements. The Pearson's correlation coefficients were 0.855, 0.919, and 0.846, respectively. The trend of VHRs showed troughs at T7 and T11 and a peak at T9; however, Hm/Hp showed slight fluctuations. Regarding the VHR, significant sex differences were found at L1 and L2 in all age bands. This innovative study focuses on vertebral morphology for opportunistic analysis in the mainland Chinese population and the distribution tendency of vertebral morphology with ageing using a chest LDCT aided by an AI system based on 3D U-Net vertebral segmentation technology. The AI system demonstrates the potential to automatically perform opportunistic vertebral morphology analyses using LDCT scans obtained during lung cancer screening. We advocate the use of age-, sex-, and vertebral level-specific criteria for the morphometric evaluation of vertebral osteoporotic fractures for a more accurate diagnosis of vertebral fractures and spinal pathologies.

Dual Attention-Based 3D U-Net Liver Segmentation Algorithm on CT Images.

The liver is a vital organ in the human body, and CT images can intuitively display its morphology. Physicians rely on liver CT images to observe its anatomical structure and areas of pathology, providing evidence for clinical diagnosis and treatment planning. To assist physicians in making accurate judgments, artificial intelligence techniques are adopted. Addressing the limitations of existing methods in liver CT image segmentation, such as weak contextual analysis and semantic information loss, we propose a novel Dual Attention-Based 3D U-Net liver segmentation algorithm on CT images. The innovations of our approach are summarized as follows: (1) We improve the 3D U-Net network by introducing residual connections to better capture multi-scale information and alleviate semantic information loss. (2) We propose the DA-Block encoder structure to enhance feature extraction capability. (3) We introduce the CBAM module into skip connections to optimize feature transmission in the encoder, reducing semantic gaps and achieving accurate liver segmentation. To validate the effectiveness of the algorithm, experiments were conducted on the LiTS dataset. The results showed that the Dice coefficient and HD95 index for liver images were 92.56% and 28.09 mm, respectively, representing an improvement of 0.84% and a reduction of 2.45 mm compared to 3D Res-UNet.

An improved 3D-UNet-based brain hippocampus segmentation model based on MR images.

OBJECTIVE: Accurate delineation of the hippocampal region via magnetic resonance imaging (MRI) is crucial for the prevention and early diagnosis of neurosystemic diseases. Determining how to accurately and quickly delineate the hippocampus from MRI results has become a serious issue. In this study, a pixel-level semantic segmentation method using 3D-UNet is proposed to realize the automatic segmentation of the brain hippocampus from MRI results. METHODS: Two hundred three-dimensional T1-weighted (3D-T1) nongadolinium contrast-enhanced magnetic resonance (MR) images were acquired at Hangzhou Cancer Hospital from June 2020 to December 2022. These samples were divided into two groups, containing 175 and 25 samples. In the first group, 145 cases were used to train the hippocampus segmentation model, and the remaining 30 cases were used to fine-tune the hyperparameters of the model. Images for twenty-five patients in the second group were used as the test set to evaluate the performance of the model. The training set of images was processed via rotation, scaling, grey value augmentation and transformation with a smooth dense deformation field for both image data and ground truth labels. A filling technique was introduced into the segmentation network to establish the hippocampus segmentation model. In addition, the performance of models established with the original network, such as VNet, SegResNet, UNetR and 3D-UNet, was compared with that of models constructed by combining the filling technique with the original segmentation network. RESULTS: The results showed that the performance of the segmentation model improved after the filling technique was introduced. Specifically, when the filling technique was introduced into VNet, SegResNet, 3D-UNet and UNetR, the segmentation performance of the models trained with an input image size of 48 × 48 × 48 improved. Among them, the 3D-UNet-based model with the filling technique achieved the best performance, with a Dice score (Dice score) of 0.7989 ± 0.0398 and a mean intersection over union (mIoU) of 0.6669 ± 0.0540, which were greater than those of the original 3D-UNet-based model. In addition, the oversegmentation ratio (OSR), average surface distance (ASD) and Hausdorff distance (HD) were 0.0666 ± 0.0351, 0.5733 ± 0.1018 and 5.1235 ± 1.4397, respectively, which were better than those of the other models. In addition, when the size of the input image was set to 48 × 48 × 48, 64 × 64 × 64 and 96 × 96 × 96, the model performance gradually improved, and the Dice scores of the proposed model reached 0.7989 ± 0.0398, 0.8371 ± 0.0254 and 0.8674 ± 0.0257, respectively. In addition, the mIoUs reached 0.6669 ± 0.0540, 0.7207 ± 0.0370 and 0.7668 ± 0.0392, respectively. CONCLUSION: The proposed hippocampus segmentation model constructed by introducing the filling technique into a segmentation network performed better than models built solely on the original network and can improve the efficiency of diagnostic analysis.

Improving lung nodule segmentation in thoracic CT scans through the ensemble of 3D U-Net models.

PURPOSE: The current study explores the application of 3D U-Net architectures combined with Inception and ResNet modules for precise lung nodule detection through deep learning-based segmentation technique. This investigation is motivated by the objective of developing a Computer-Aided Diagnosis (CAD) system for effective diagnosis and prognostication of lung nodules in clinical settings. METHODS: The proposed method trained four different 3D U-Net models on the retrospective dataset obtained from AIIMS Delhi. To augment the training dataset, affine transformations and intensity transforms were utilized. Preprocessing steps included CT scan voxel resampling, intensity normalization, and lung parenchyma segmentation. Model optimization utilized a hybrid loss function that combined Dice Loss and Focal Loss. The model performance of all four 3D U-Nets was evaluated patient-wise using dice coefficient and Jaccard coefficient, then averaged to obtain the average volumetric dice coefficient (DSCavg) and average Jaccard coefficient (IoUavg) on a test dataset comprising 53 CT scans. Additionally, an ensemble approach (Model-V) was utilized featuring 3D U-Net (Model-I), ResNet (Model-II), and Inception (Model-III) 3D U-Net architectures, combined with two distinct patch sizes for further investigation. RESULTS: The ensemble of models obtained the highest DSCavg of 0.84 ± 0.05 and IoUavg of 0.74 ± 0.06 on the test dataset, compared against individual models. It mitigated false positives, overestimations, and underestimations observed in individual U-Net models. Moreover, the ensemble of models reduced average false positives per scan in the test dataset (1.57 nodules/scan) compared to individual models (2.69-3.39 nodules/scan). CONCLUSIONS: The suggested ensemble approach presents a strong and effective strategy for automatically detecting and delineating lung nodules, potentially aiding CAD systems in clinical settings. This approach could assist radiologists in laborious and meticulous lung nodule detection tasks in CT scans, improving lung cancer diagnosis and treatment planning.

Deep learning-based correction for time truncation in cerebral computed tomography perfusion.

Cerebral computed tomography perfusion (CTP) imaging requires complete acquisition of contrast bolus inflow and washout in the brain parenchyma; however, time truncation undoubtedly occurs in clinical practice. To overcome this issue, we proposed a three-dimensional (two-dimensional + time) convolutional neural network (CNN)-based approach to predict missing CTP image frames at the end of the series from earlier acquired image frames. Moreover, we evaluated three strategies for predicting multiple time points. Seventy-two CTP scans with 89 frames and eight slices from a publicly available dataset were used to train and test the CNN models capable of predicting the last 10 image frames. The prediction strategies were single-shot prediction, recursive multi-step prediction, and direct-recursive hybrid prediction.Single-shot prediction predicted all frames simultaneously, while recursive multi-step prediction used prior predictions as input for subsequent steps, and direct-recursive hybrid prediction employed separate models for each step with prior predictions as input for the next step. The accuracies of the predicted image frames were evaluated in terms of image quality, bolus shape, and clinical perfusion parameters. We found that the image quality metrics were superior when multiple CTP images were predicted simultaneously rather than recursively. The bolus shape also showed the highest correlation (r = 0.990, p < 0.001) and the lowest variance (95% confidence interval, -453.26-445.53) in the single-shot prediction. For all perfusion parameters, the single-shot prediction had the smallest absolute differences from ground truth. Our proposed approach can potentially minimize time truncation errors and support the accurate quantification of ischemic stroke.

Deep learning-based automated detection and segmentation of bone and traumatic bone marrow lesions from MRI following an acute ACL tear.

INTRODUCTION: Traumatic bone marrow lesions (BML) are frequently identified on knee MRI scans in patients following an acute full-thickness, complete ACL tear. BMLs coincide with regions of elevated localized bone loss, and studies suggest these may act as a precursor to the development of post-traumatic osteoarthritis. This study addresses the labour-intensive manual assessment of BMLs by using a 3D U-Net for automated identification and segmentation from MRI scans. METHODS: A multi-task learning approach was used to segment both bone and BML from T2 fat-suppressed (FS) fast spin echo (FSE) MRI sequences for BML assessment. Training and testing utilized datasets from individuals with complete ACL tears, employing a five-fold cross-validation approach and pre-processing involved image intensity normalization and data augmentation. A post-processing algorithm was developed to improve segmentation and remove outliers. Training and testing datasets were acquired from different studies with similar imaging protocol to assess the model's performance robustness across different populations and acquisition conditions. RESULTS: The 3D U-Net model exhibited effectiveness in semantic segmentation, while post-processing enhanced segmentation accuracy and precision through morphological operations. The trained model with post-processing achieved a Dice similarity coefficient (DSC) of 0.75 ± 0.08 (mean ± std) and a precision of 0.87 ± 0.07 for BML segmentation on testing data. Additionally, the trained model with post-processing achieved a DSC of 0.93 ± 0.02 and a precision of 0.92 ± 0.02 for bone segmentation on testing data. This demonstrates the approach's high accuracy for capturing true positives and effectively minimizing false positives in the identification and segmentation of bone structures. CONCLUSION: Automated segmentation methods are a valuable tool for clinicians and researchers, streamlining the assessment of BMLs and allowing for longitudinal assessments. This study presents a model with promising clinical efficacy and provides a quantitative approach for bone-related pathology research and diagnostics.

Clinical Use of Hematoma Volume Based On Automated Segmentation of Chronic Subdural Hematoma Using 3D U-Net.

PURPOSE: To propose a method for calculating hematoma volume based on automatic segmentation of chronic subdural hematoma (CSDH) using 3D U­net and investigate whether it can be used clinically to predict recurrence. METHODS: Hematoma volumes manually measured from pre- and postoperative computed tomography (CT) images were used as ground truth data to train 3D U­net in 200 patients (400 CT scans). A total of 215 patients (430 CT scans) were used as test data to output segmentation results from the trained 3D U­net model. The similarity with the ground truth data was evaluated using Dice scores for pre and postoperative separately. The recurrence prediction accuracy was evaluated by obtaining receiver operating characteristic (ROC) curves for the segmentation results. Using a typical mobile PC, the computation time per case was measured and the average time was calculated. RESULTS: The median Dice score of the test data were preoperative hematoma volume (Pre-HV): 0.764 and postoperative subdural cavity volume (Post-SCV): 0.741. In ROC analyses assessing recurrence prediction, the area under the curve (AUC) of the manual was 0.755 in Pre-HV, whereas the 3D U­net was 0.735. In Post-SCV, the manual AUC was 0.779; the 3D U­net was 0.736. No significant differences were found between manual and 3D U­net for all results. Using a mobile PC, the average time taken to output the test data results was 30 s per case. CONCLUSION: The proposed method is a simple, accurate, and clinically applicable; it can contribute to the widespread use of recurrence prediction scoring systems for CSDH.

MRUNet-3D: A multi-stride residual 3D UNet for lung nodule segmentation.

Obtaining an accurate segmentation of the pulmonary nodules in computed tomography (CT) images is challenging. This is due to: (1) the heterogeneous nature of the lung nodules; (2) comparable visual characteristics between the nodules and their surroundings. A robust multi-scale feature extraction mechanism that can effectively obtain multi-scale representations at a granular level can improve segmentation accuracy. As the most commonly used network in lung nodule segmentation, UNet, its variants, and other image segmentation methods lack this robust feature extraction mechanism. In this study, we propose a multi-stride residual 3D UNet (MRUNet-3D) to improve the segmentation accuracy of lung nodules in CT images. It incorporates a multi-slide Res2Net block (MSR), which replaces the simple sequence of convolution layers in each encoder stage to effectively extract multi-scale features at a granular level from different receptive fields and resolutions while conserving the strengths of 3D UNet. The proposed method has been extensively evaluated on the publicly available LUNA16 dataset. Experimental results show that it achieves competitive segmentation performance with an average dice similarity coefficient of 83.47 % and an average surface distance of 0.35 mm on the dataset. More notably, our method has proven to be robust to the heterogeneity of lung nodules. It has also proven to perform better at segmenting small lung nodules. Ablation studies have shown that the proposed MSR and RFIA modules are fundamental to improving the performance of the proposed model.

Classification of substances by health hazard using deep neural networks and molecular electron densities.

In this paper we present a method that allows leveraging 3D electron density information to train a deep neural network pipeline to segment regions of high, medium and low electronegativity and classify substances as health hazardous or non-hazardous. We show that this can be used for use-cases such as cosmetics and food products. For this purpose, we first generate 3D electron density cubes using semiempirical molecular calculations for a custom European Chemicals Agency (ECHA) subset consisting of substances labelled as hazardous and non-hazardous for cosmetic usage. Together with their 3-class electronegativity maps we train a modified 3D-UNet with electron density cubes to segment reactive sites in molecules and classify substances with an accuracy of 78.1%. We perform the same process on a custom food dataset (CompFood) consisting of hazardous and non-hazardous substances compiled from European Food Safety Authority (EFSA) OpenFoodTox, Food and Drug Administration (FDA) Generally Recognized as Safe (GRAS) and FooDB datasets to achieve a classification accuracy of 64.1%. Our results show that 3D electron densities and particularly masked electron densities, calculated by taking a product of original electron densities and regions of high and low electronegativity can be used to classify molecules for different use-cases and thus serve not only to guide safe-by-design product development but also aid in regulatory decisions. SCIENTIFIC CONTRIBUTION: We aim to contribute to the diverse 3D molecular representations used for training machine learning algorithms by showing that a deep learning network can be trained on 3D electron density representation of molecules. This approach has previously not been used to train machine learning models and it allows utilization of the true spatial domain of the molecule for prediction of properties such as their suitability for usage in cosmetics and food products and in future, to other molecular properties. The data and code used for training is accessible at https://github.com/s-singh-ivv/eDen-Substances .

A Comparative Study of Deep Learning Dose Prediction Models for Cervical Cancer Volumetric Modulated Arc Therapy.

Purpose: Deep learning (DL) is widely used in dose prediction for radiation oncology, multiple DL techniques comparison is often lacking in the literature. To compare the performance of 4 state-of-the-art DL models in predicting the voxel-level dose distribution for cervical cancer volumetric modulated arc therapy (VMAT). Methods and Materials: A total of 261 patients' plans for cervical cancer were retrieved in this retrospective study. A three-channel feature map, consisting of a planning target volume (PTV) mask, organs at risk (OARs) mask, and CT image was fed into the three-dimensional (3D) U-Net and its 3 variants models. The data set was randomly divided into 80% as training-validation and 20% as testing set, respectively. The model performance was evaluated on the 52 testing patients by comparing the generated dose distributions against the clinical approved ground truth (GT) using mean absolute error (MAE), dose map difference (GT-predicted), clinical dosimetric indices, and dice similarity coefficients (DSC). Results: The 3D U-Net and its 3 variants DL models exhibited promising performance with a maximum MAE within the PTV 0.83% ± 0.67% in the UNETR model. The maximum MAE among the OARs is the left femoral head, which reached 6.95% ± 6.55%. For the body, the maximum MAE was observed in UNETR, which is 1.19 ± 0.86%, and the minimum MAE was 0.94 ± 0.85% for 3D U-Net. The average error of the Dmean difference for different OARs is within 2.5 Gy. The average error of V40 difference for the bladder and rectum is about 5%. The mean DSC under different isodose volumes was above 90%. Conclusions: DL models can predict the voxel-level dose distribution accurately for cervical cancer VMAT treatment plans. All models demonstrated almost analogous performance for voxel-wise dose prediction maps. Considering all voxels within the body, 3D U-Net showed the best performance. The state-of-the-art DL models are of great significance for further clinical applications of cervical cancer VMAT.

DeepN4: Learning N4ITK Bias Field Correction for T1-weighted Images.

T1-weighted (T1w) MRI has low frequency intensity artifacts due to magnetic field inhomogeneities. Removal of these biases in T1w MRI images is a critical preprocessing step to ensure spatially consistent image interpretation. N4ITK bias field correction, the current state-of-the-art, is implemented in such a way that makes it difficult to port between different pipelines and workflows, thus making it hard to reimplement and reproduce results across local, cloud, and edge platforms. Moreover, N4ITK is opaque to optimization before and after its application, meaning that methodological development must work around the inhomogeneity correction step. Given the importance of bias fields correction in structural preprocessing and flexible implementation, we pursue a deep learning approximation / reinterpretation of the N4ITK bias fields correction to create a method which is portable, flexible, and fully differentiable. In this paper, we trained a deep learning network "DeepN4" on eight independent cohorts from 72 different scanners and age ranges with N4ITK-corrected T1w MRI and bias field for supervision in log space. We found that we can closely approximate N4ITK bias fields correction with naïve networks. We evaluate the peak signal to noise ratio (PSNR) in test dataset against the N4ITK corrected images. The median PSNR of corrected images between N4ITK and DeepN4 was 47.96 dB. In addition, we assess the DeepN4 model on eight additional external datasets and show the generalizability of the approach. This study establishes that incompatible N4ITK preprocessing steps can be closely approximated by naïve deep neural networks, facilitating more flexibility. All code and models are released at https://github.com/MASILab/DeepN4 .

Active Learning in Brain Tumor Segmentation with Uncertainty Sampling and Annotation Redundancy Restriction.

Deep learning models have demonstrated great potential in medical imaging but are limited by the expensive, large volume of annotations required. To address this, we compared different active learning strategies by training models on subsets of the most informative images using real-world clinical datasets for brain tumor segmentation and proposing a framework that minimizes the data needed while maintaining performance. Then, 638 multi-institutional brain tumor magnetic resonance imaging scans were used to train three-dimensional U-net models and compare active learning strategies. Uncertainty estimation techniques including Bayesian estimation with dropout, bootstrapping, and margins sampling were compared to random query. Strategies to avoid annotating similar images were also considered. We determined the minimum data necessary to achieve performance equivalent to the model trained on the full dataset (α = 0.05). Bayesian approximation with dropout at training and testing showed results equivalent to that of the full data model (target) with around 30% of the training data needed by random query to achieve target performance (p = 0.018). Annotation redundancy restriction techniques can reduce the training data needed by random query to achieve target performance by 20%. We investigated various active learning strategies to minimize the annotation burden for three-dimensional brain tumor segmentation. Dropout uncertainty estimation achieved target performance with the least annotated data.

TMTV-Net: fully automated total metabolic tumor volume segmentation in lymphoma PET/CT images - a multi-center generalizability analysis.

PURPOSE: Total metabolic tumor volume (TMTV) segmentation has significant value enabling quantitative imaging biomarkers for lymphoma management. In this work, we tackle the challenging task of automated tumor delineation in lymphoma from PET/CT scans using a cascaded approach. METHODS: Our study included 1418 2-[18F]FDG PET/CT scans from four different centers. The dataset was divided into 900 scans for development/validation/testing phases and 518 for multi-center external testing. The former consisted of 450 lymphoma, lung cancer, and melanoma scans, along with 450 negative scans, while the latter consisted of lymphoma patients from different centers with diffuse large B cell, primary mediastinal large B cell, and classic Hodgkin lymphoma cases. Our approach involves resampling PET/CT images into different voxel sizes in the first step, followed by training multi-resolution 3D U-Nets on each resampled dataset using a fivefold cross-validation scheme. The models trained on different data splits were ensemble. After applying soft voting to the predicted masks, in the second step, we input the probability-averaged predictions, along with the input imaging data, into another 3D U-Net. Models were trained with semi-supervised loss. We additionally considered the effectiveness of using test time augmentation (TTA) to improve the segmentation performance after training. In addition to quantitative analysis including Dice score (DSC) and TMTV comparisons, the qualitative evaluation was also conducted by nuclear medicine physicians. RESULTS: Our cascaded soft-voting guided approach resulted in performance with an average DSC of 0.68 ± 0.12 for the internal test data from developmental dataset, and an average DSC of 0.66 ± 0.18 on the multi-site external data (n = 518), significantly outperforming (p < 0.001) state-of-the-art (SOTA) approaches including nnU-Net and SWIN UNETR. While TTA yielded enhanced performance gains for some of the comparator methods, its impact on our cascaded approach was found to be negligible (DSC: 0.66 ± 0.16). Our approach reliably quantified TMTV, with a correlation of 0.89 with the ground truth (p < 0.001). Furthermore, in terms of visual assessment, concordance between quantitative evaluations and clinician feedback was observed in the majority of cases. The average relative error (ARE) and the absolute error (AE) in TMTV prediction on external multi-centric dataset were ARE = 0.43 ± 0.54 and AE = 157.32 ± 378.12 (mL) for all the external test data (n = 518), and ARE = 0.30 ± 0.22 and AE = 82.05 ± 99.78 (mL) when the 10% outliers (n = 53) were excluded. CONCLUSION: TMTV-Net demonstrates strong performance and generalizability in TMTV segmentation across multi-site external datasets, encompassing various lymphoma subtypes. A negligible reduction of 2% in overall performance during testing on external data highlights robust model generalizability across different centers and cancer types, likely attributable to its training with resampled inputs. Our model is publicly available, allowing easy multi-site evaluation and generalizability analysis on datasets from different institutions.

deepPGSegNet: MRI-based pituitary gland segmentation using deep learning.

Introduction: In clinical research on pituitary disorders, pituitary gland (PG) segmentation plays a pivotal role, which impacts the diagnosis and treatment of conditions such as endocrine dysfunctions and visual impairments. Manual segmentation, which is the traditional method, is tedious and susceptible to inter-observer differences. Thus, this study introduces an automated solution, utilizing deep learning, for PG segmentation from magnetic resonance imaging (MRI). Methods: A total of 153 university students were enrolled, and their MRI images were used to build a training dataset and ground truth data through manual segmentation of the PGs. A model was trained employing data augmentation and a three-dimensional U-Net architecture with a five-fold cross-validation. A predefined field of view was applied to highlight the PG region to optimize memory usage. The model's performance was tested on an independent dataset. The model's performance was tested on an independent dataset for evaluating accuracy, precision, recall, and an F1 score. Results and discussion: The model achieved a training accuracy, precision, recall, and an F1 score of 92.7%, 0.87, 0.91, and 0.89, respectively. Moreover, the study explored the relationship between PG morphology and age using the model. The results indicated a significant association between PG volume and midsagittal area with age. These findings suggest that a precise volumetric PG analysis through an automated segmentation can greatly enhance diagnostic accuracy and surveillance of pituitary disorders.

Knowledge distillation on individual vertebrae segmentation exploiting 3D U-Net.

Recent advances in medical imaging have highlighted the critical development of algorithms for individual vertebral segmentation on computed tomography (CT) scans. Essential for diagnostic accuracy and treatment planning in orthopaedics, neurosurgery and oncology, these algorithms face challenges in clinical implementation, including integration into healthcare systems. Consequently, our focus lies in exploring the application of knowledge distillation (KD) methods to train shallower networks capable of efficiently segmenting vertebrae in CT scans. This approach aims to reduce segmentation time, enhance suitability for emergency cases, and optimize computational and memory resource efficiency. Building upon prior research in the field, a two-step segmentation approach was employed. Firstly, the spine's location was determined by predicting a heatmap, indicating the probability of each voxel belonging to the spine. Subsequently, an iterative segmentation of vertebrae was performed from the top to the bottom of the CT volume over the located spine, using a memory instance to record the already segmented vertebrae. KD methods were implemented by training a teacher network with performance similar to that found in the literature, and this knowledge was distilled to a shallower network (student). Two KD methods were applied: (1) using the soft outputs of both networks and (2) matching logits. Two publicly available datasets, comprising 319 CT scans from 300 patients and a total of 611 cervical, 2387 thoracic, and 1507 lumbar vertebrae, were used. To ensure dataset balance and robustness, effective data augmentation methods were applied, including cleaning the memory instance to replicate the first vertebra segmentation. The teacher network achieved an average Dice similarity coefficient (DSC) of 88.22% and a Hausdorff distance (HD) of 7.71 mm, showcasing performance similar to other approaches in the literature. Through knowledge distillation from the teacher network, the student network's performance improved, with an average DSC increasing from 75.78% to 84.70% and an HD decreasing from 15.17 mm to 8.08 mm. Compared to other methods, our teacher network exhibited up to 99.09% fewer parameters, 90.02% faster inference time, 88.46% shorter total segmentation time, and 89.36% less associated carbon (CO2) emission rate. Regarding our student network, it featured 75.00% fewer parameters than our teacher, resulting in a 36.15% reduction in inference time, a 33.33% decrease in total segmentation time, and a 42.96% reduction in CO2 emissions. This study marks the first exploration of applying KD to the problem of individual vertebrae segmentation in CT, demonstrating the feasibility of achieving comparable performance to existing methods using smaller neural networks.

SSDL-an automated semi-supervised deep learning approach for patient-specific 3D reconstruction of proximal femur from QCT images.

Deep Learning (DL) techniques have recently been used in medical image segmentation and the reconstruction of 3D anatomies of a human body. In this work, we propose a semi-supervised DL (SSDL) approach utilizing a CNN-based 3D U-Net model for femur segmentation from sparsely annotated quantitative computed tomography (QCT) slices. Specifically, QCT slices at the proximal end of the femur forming ball and socket joint with acetabulum were annotated for precise segmentation, where a segmenting binary mask was generated using a 3D U-Net model to segment the femur accurately. A total of 5474 QCT slices were considered for training among which 2316 slices were annotated. 3D femurs were further reconstructed from segmented slices employing polynomial spline interpolation. Both qualitative and quantitative performance of segmentation and 3D reconstruction were satisfactory with more than 90% accuracy achieved for all of the standard performance metrics considered. The spatial overlap index and reproducibility validation metric for segmentation-Dice Similarity Coefficient was 91.8% for unseen patients and 99.2% for validated patients. An average relative error of 12.02% and 10.75% for volume and surface area, respectively, were computed for 3D reconstructed femurs. The proposed approach demonstrates its effectiveness in accurately segmenting and reconstructing 3D femur from QCT slices.

Contour attention network for cerebrovascular segmentation from TOF-MRA volumetric images.

BACKGROUND: Cerebrovascular segmentation is a crucial step in the computer-assisted diagnosis of cerebrovascular pathologies. However, accurate extraction of cerebral vessels from time-of-flight magnetic resonance angiography (TOF-MRA) data is still challenging due to the complex topology and slender shape. PURPOSE: The existing deep learning-based approaches pay more attention to the skeleton and ignore the contour, which limits the segmentation performance of the cerebrovascular structure. We aim to weight the contour of brain vessels in shallow features when concatenating with deep features. It helps to obtain more accurate cerebrovascular details and narrows the semantic gap between multilevel features. METHODS: This work proposes a novel framework for priority extraction of contours in cerebrovascular structures. We first design a neighborhood-based algorithm to generate the ground truth of the cerebrovascular contour from original annotations, which can introduce useful shape information for the segmentation network. Moreover, We propose an encoder-dual decoder-based contour attention network (CA-Net), which consists of the dilated asymmetry convolution block (DACB) and the Contour Attention Module (CAM). The ancillary decoder uses the DACB to obtain cerebrovascular contour features under the supervision of contour annotations. The CAM transforms these features into a spatial attention map to increase the weight of the contour voxels in main decoder to better restored the vessel contour details. RESULTS: The CA-Net is thoroughly validated using two publicly available datasets, and the experimental results demonstrate that our network outperforms the competitors for cerebrovascular segmentation. We achieved the average dice similarity coefficient ( D S C $DSC$ ) of 68.15 and 99.92% in natural and synthetic datasets. Our method segments cerebrovascular structures with better completeness. CONCLUSIONS: We propose a new framework containing contour annotation generation and cerebrovascular segmentation network that better captures the tiny vessels and improve vessel connectivity.