Evaluating White Matter Lesion Segmentations with Refined Sørensen-Dice Analysis

From SPECT/CT towards absolute quantification? - the case of unilateral condylar hyperplasia of the mandible

BACKGROUND

Unilateral condylar hyperplasia (UCH) of the mandible is a rare condition characterized by asymmetric growth of the mandibular condyles. Bone scintigraphy with SPECT(/CT) is commonly used to diagnose UCH and guide treatment. Still, varying results have been reported using the traditional threshold of 55%:45% in relative tracer uptake. While absolute quantification of uptake on SPECT/CT could improve results, optimal correction and reconstruction settings are currently unknown.

METHODS

Three anthropomorphic phantoms representing UCH were developed from patient CT volumes and produced using 3D printing technology. Fillable spherical inserts of different sizes (Ø: 8-15 mm) were placed in the condylar positions representing symmetrical and asymmetrical distributions. Recovery coefficients were determined for SPECT/CT using various reconstruction corrections, including attenuation and scatter correction (ACSC), resolution modeling (RM), and partial volume correction (PVC) using phantom measurements. Uptake ratios between condyles and condyle to clivus were evaluated. Finally, the impact of these correction techniques on absolute activity and diagnostic accuracy was assessed in a retrospective patient cohort for the diagnostic threshold of 55%:45%.

RESULTS

The activity was only partially recovered in all spherical inserts (range: 22.5-64.9%). However, RM improved relative recovery by 20.2-62.3% compared to ACSC. In the symmetric phantoms, the 95% confidence interval (CI) of condyle ratios included the diagnostic threshold (57.6%:42.4%) for UCH when using ACSC potentially leading to false positives, but not for ACSCRM datasets. Partial volume corrections coefficients from the NEMA IQ phantom was positionally dependent, with improvements seen performing PVC using coefficients derived from anthropomorphic phantoms. Retrospective application in a patient cohort showed only a weak linear correlation (R²: 0.25-0.67) and large limits of agreement (9.6-12.5%) between different reconstructions. Up to 44% of patients were reclassified using the 55%:45% threshold. Using clinical outcome data, ACSCRM had highest sensitivity (91%; 95% CI 59-100%) and specificity (66%; 95% CI 47-81%), significantly improving specificity (P = 0.038).

CONCLUSIONS

Anthropomorphic phantoms were shown to be essential in determining optimal settings for acquisition, reconstruction, and analysis. SPECT/CT reconstructions with attenuation and scatter correction and resolution modeling are recommended and could improve specificity when using the 55%:45% threshold to assess condylar growth.

CellSNAP: a fast, accurate algorithm for 3D cell segmentation in quantitative phase imaging

Significance

Three-dimensional quantitative phase imaging (QPI) has rapidly emerged as a complementary tool to fluorescence imaging, as it provides an objective measure of cell morphology and dynamics, free of variability due to contrast agents. It has opened up new directions of investigation by providing systematic and correlative analysis of various cellular parameters without limitations of photobleaching and phototoxicity. While current QPI systems allow the rapid acquisition of tomographic images, the pipeline to analyze these raw three-dimensional (3D) tomograms is not well-developed. We focus on a critical, yet often underappreciated, step of the analysis pipeline that of 3D cell segmentation from the acquired tomograms.

Aim

We report the CellSNAP (Cell Segmentation via Novel Algorithm for Phase Imaging) algorithm for the 3D segmentation of QPI images.

Approach

The cell segmentation algorithm mimics the gemstone extraction process, initiating with a coarse 3D extrusion from a two-dimensional (2D) segmented mask to outline the cell structure. A 2D image is generated, and a segmentation algorithm identifies the boundary in the x - y plane. Leveraging cell continuity in consecutive z -stacks, a refined 3D segmentation, akin to fine chiseling in gemstone carving, completes the process.

Results

The CellSNAP algorithm outstrips the current gold standard in terms of speed, robustness, and implementation, achieving cell segmentation under 2 s per cell on a single-core processor. The implementation of CellSNAP can easily be parallelized on a multi-core system for further speed improvements. For the cases where segmentation is possible with the existing standard method, our algorithm displays an average difference of 5% for dry mass and 8% for volume measurements. We also show that CellSNAP can handle challenging image datasets where cells are clumped and marred by interferogram drifts, which pose major difficulties for all QPI-focused AI-based segmentation tools.

Conclusion

Our proposed method is less memory intensive and significantly faster than existing methods. The method can be easily implemented on a student laptop. Since the approach is rule-based, there is no need to collect a lot of imaging data and manually annotate them to perform machine learning based training of the model. We envision our work will lead to broader adoption of QPI imaging for high-throughput analysis, which has, in part, been stymied by a lack of suitable image segmentation tools.

U-Net enhanced real-time LED-based photoacoustic imaging

Photoacoustic (PA) imaging is hybrid imaging modality with good optical contrast and spatial resolution. Portable, cost-effective, smaller footprint light emitting diodes (LEDs) are rapidly becoming important PA optical sources. However, the key challenge faced by the LED-based systems is the low light fluence that is generally compensated by high frame averaging, consequently reducing acquisition frame-rate. In this study, we present a simple deep learning U-Net framework that enhances the signal-to-noise ratio (SNR) and contrast of PA image obtained by averaging low number of frames. The SNR increased by approximately four-fold for both in-class in vitro phantoms (4.39 ± 2.55) and out-of-class in vivo models (4.27 ± 0.87). We also demonstrate the noise invariancy of the network and discuss the downsides (blurry outcome and failure to reduce the salt & pepper noise). Overall, the developed U-Net framework can provide a real-time image enhancement platform for clinically translatable low-cost and low-energy light source-based PA imaging systems.

GradHC: highly reliable gradual hash-based clustering for DNA storage systems

MOTIVATION

As data storage challenges grow and existing technologies approach their limits, synthetic DNA emerges as a promising storage solution due to its remarkable density and durability advantages. While cost remains a concern, emerging sequencing and synthetic technologies aim to mitigate it, yet introduce challenges such as errors in the storage and retrieval process. One crucial task in a DNA storage system is clustering numerous DNA reads into groups that represent the original input strands.

RESULTS

In this paper, we review different methods for evaluating clustering algorithms and introduce a novel clustering algorithm for DNA storage systems, named Gradual Hash-based clustering (GradHC). The primary strength of GradHC lies in its capability to cluster with excellent accuracy various types of designs, including varying strand lengths, cluster sizes (including extremely small clusters), and different error ranges. Benchmark analysis demonstrates that GradHC is significantly more stable and robust than other clustering algorithms previously proposed for DNA storage, while also producing highly reliable clustering results.

AVAILABILITY AND IMPLEMENTATION

https://github.com/bensdvir/GradHC.

Spleen volume is independently associated with non-alcoholic fatty liver disease, liver volume and liver fibrosis

Non-alcoholic fatty liver disease (NAFLD) can lead to irreversible liver damage manifesting in systemic effects (e.g., elevated portal vein pressure and splenomegaly) with increased risk of deadly outcomes. However, the association of spleen volume with NAFLD and related type 2-diabetes (T2D) is not fully understood. The UK Biobank contains comprehensive health-data of 500,000 participants, including clinical data and MR images of >40,000 individuals. The present study estimated the spleen volume of 37,066 participants through automated deep learning-based image segmentation of neck-to-knee MR images. The aim was to investigate the associations of spleen volume with NAFLD, T2D and liver fibrosis, while adjusting for natural confounders. The recent redefinition and new designation of NAFLD to metabolic dysfunction-associated steatotic liver disease (MASLD), promoted by major organisations of studies on liver disease, was not employed as introduced after the conduct of this study. The results showed that spleen volume decreased with age, correlated positively with body size and was smaller in females compared to males. Larger spleens were observed in subjects with NAFLD and T2D compared to controls. Spleen volume was also positively and independently associated with liver fat fraction, liver volume and the fibrosis-4 score, with notable volumetric increases already at low liver fat fractions and volumes, but not independently associated with T2D. These results suggest a link between spleen volume and NAFLD already at an early stage of the disease, potentially due to initial rise in portal vein pressure.

Image harmonization improves consistency of intra-rater delineations of MS lesions in heterogeneous MRI

Clinical magnetic resonance images (MRIs) lack a standard intensity scale due to differences in scanner hardware and the pulse sequences used to acquire the images. When MRIs are used for quantification, as in the evaluation of white matter lesions (WMLs) in multiple sclerosis, this lack of intensity standardization becomes a critical problem affecting both the staging and tracking of the disease and its treatment. This paper presents a study of harmonization on WML segmentation consistency, which is evaluated using an object detection classification scheme that incorporates manual delineations from both the original and harmonized MRIs. A cohort of ten people scanned on two different imaging platforms was studied. An expert rater, blinded to the image source, manually delineated WMLs on images from both scanners before and after harmonization. It was found that there is closer agreement in both global and per-lesion WML volume and spatial distribution after harmonization, demonstrating the importance of image harmonization prior to the creation of manual delineations. These results could lead to better truth models in both the development and evaluation of automated lesion segmentation algorithms.

A Comparison of 7 Tesla MR Spectroscopic Imaging and 3 Tesla MR Fingerprinting for Tumor Localization in Glioma Patients

This paper investigated the correlation between magnetic resonance spectroscopic imaging (MRSI) and magnetic resonance fingerprinting (MRF) in glioma patients by comparing neuro-oncological markers obtained from MRSI to T1/T2 maps from MRF. Data from 12 consenting patients with gliomas were analyzed by defining hotspots for T1, T2, and various metabolic ratios, and comparing them using Sørensen-Dice similarity coefficients (DSCs) and the distances between their centers of intensity (COIDs). The median DSCs between MRF and the tumor segmentation were 0.73 (T1) and 0.79 (T2). The DSCs between MRSI and MRF were the highest for Gln/tNAA (T1: 0.75, T2: 0.80, tumor: 0.78), followed by Gly/tNAA (T1: 0.57, T2: 0.62, tumor: 0.54) and tCho/tNAA (T1: 0.61, T2: 0.58, tumor: 0.45). The median values in the tumor hotspot were T1 = 1724 ms, T2 = 86 ms, Gln/tNAA = 0.61, Gly/tNAA = 0.28, Ins/tNAA = 1.15, and tCho/tNAA = 0.48, and, in the peritumoral region, were T1 = 1756 ms, T2 = 102 ms, Gln/tNAA = 0.38, Gly/tNAA = 0.20, Ins/tNAA = 1.06, and tCho/tNAA = 0.38, and, in the NAWM, were T1 = 950 ms, T2 = 43 ms, Gln/tNAA = 0.16, Gly/tNAA = 0.07, Ins/tNAA = 0.54, and tCho/tNAA = 0.20. The results of this study constitute the first comparison of 7T MRSI and 3T MRF, showing a good correspondence between these methods.

Metrics reloaded: recommendations for image analysis validation

Increasing evidence shows that flaws in machine learning (ML) algorithm validation are an underestimated global problem. In biomedical image analysis, chosen performance metrics often do not reflect the domain interest, and thus fail to adequately measure scientific progress and hinder translation of ML techniques into practice. To overcome this, we created Metrics Reloaded, a comprehensive framework guiding researchers in the problem-aware selection of metrics. Developed by a large international consortium in a multistage Delphi process, it is based on the novel concept of a problem fingerprint-a structured representation of the given problem that captures all aspects that are relevant for metric selection, from the domain interest to the properties of the target structure(s), dataset and algorithm output. On the basis of the problem fingerprint, users are guided through the process of choosing and applying appropriate validation metrics while being made aware of potential pitfalls. Metrics Reloaded targets image analysis problems that can be interpreted as classification tasks at image, object or pixel level, namely image-level classification, object detection, semantic segmentation and instance segmentation tasks. To improve the user experience, we implemented the framework in the Metrics Reloaded online tool. Following the convergence of ML methodology across application domains, Metrics Reloaded fosters the convergence of validation methodology. Its applicability is demonstrated for various biomedical use cases.

Development of RLK-Unet: a clinically favorable deep learning algorithm for brain metastasis detection and treatment response assessment

Purpose/objectives

Previous deep learning (DL) algorithms for brain metastasis (BM) detection and segmentation have not been commonly used in clinics because they produce false-positive findings, require multiple sequences, and do not reflect physiological properties such as necrosis. The aim of this study was to develop a more clinically favorable DL algorithm (RLK-Unet) using a single sequence reflecting necrosis and apply it to automated treatment response assessment.

Methods and materials

A total of 128 patients with 1339 BMs, who underwent BM magnetic resonance imaging using the contrast-enhanced 3D T1 weighted (T1WI) turbo spin-echo black blood sequence, were included in the development of the DL algorithm. Fifty-eight patients with 629 BMs were assessed for treatment response. The detection sensitivity, precision, Dice similarity coefficient (DSC), and agreement of treatment response assessments between neuroradiologists and RLK-Unet were assessed.

Results

RLK-Unet demonstrated a sensitivity of 86.9% and a precision of 79.6% for BMs and had a DSC of 0.663. Segmentation performance was better in the subgroup with larger BMs (DSC, 0.843). The agreement in the response assessment for BMs between the radiologists and RLK-Unet was excellent (intraclass correlation, 0.84).

Conclusion

RLK-Unet yielded accurate detection and segmentation of BM and could assist clinicians in treatment response assessment.

A fully automated machine learning-based methodology for personalized radiation dose assessment in thoracic and abdomen CT

PURPOSE

To develop a machine learning-based methodology for patient-specific radiation dosimetry in thoracic and abdomen CT.

METHODS

Three hundred and thirty-one thoracoabdominal radiotherapy-planning CT examinations with the respective organ/patient contours were collected retrospectively for the development and validation of segmentation 3D-UNets. Moreover, 97 diagnostic thoracic and 89 diagnostic abdomen CT examinations were collected retrospectively. For each of the diagnostic CT examinations, personalized MC dosimetry was performed. The data derived from MC simulations along with the respective CT data were used for the training and validation of a dose prediction deep neural network (DNN). An algorithm was developed to utilize the trained models and perform patient-specific organ dose estimates for thoracic and abdomen CT examinations. The doses estimated with the DNN were compared with the respective doses derived from MC simulations. A paired t-test was conducted between the DNN and MC results. Furthermore, the time efficiency of the proposed methodology was assessed.

RESULTS

The mean percentage differences (range) between DNN and MC dose estimates for the lungs, liver, spleen, stomach, and kidneys were 7.2 % (0.2-24.1 %), 5.5 % (0.4-23.0 %), 7.9 % (0.6-22.3 %), 6.9 % (0.0-23.0 %) and 6.7 % (0.3-22.6 %) respectively. The differences between DNN and MC dose estimates were not significant (p-value = 0.12). Moreover, the mean processing time of the proposed workflow was 99 % lower than the respective time needed for MC-based dosimetry.

CONCLUSIONS

The proposed methodology can be used for rapid and accurate patient-specific dosimetry in chest and abdomen CT.

Experimental Examination of Conventional, Semi-Automatic, and Automatic Volumetry Tools for Segmentation of Pulmonary Nodules in a Phantom Study

The aim of this study is to examine the precision of semi-automatic, conventional and automatic volumetry tools for pulmonary nodules in chest CT with phantom N1 LUNGMAN. The phantom is a life-size anatomical chest model with pulmonary nodules representing solid and subsolid metastases. Gross tumor volumes (GTVis) were contoured using various approaches: manually (0); as a means of semi-automated, conventional contouring with (I) adaptive-brush function; (II) flood-fill function; and (III) image-thresholding function. Furthermore, a deep-learning algorithm for automatic contouring was applied (IV). An intermodality comparison of the above-mentioned strategies for contouring GTVis was performed. For the mean GTVref (standard deviation (SD)), the interquartile range (IQR)) was 0.68 mL (0.33; 0.34-1.1). GTV segmentation was distributed as follows: (I) 0.61 mL (0.27; 0.36-0.92); (II) 0.41 mL (0.28; 0.23-0.63); (III) 0.65 mL (0.35; 0.32-0.90); and (IV) 0.61 mL (0.29; 0.33-0.95). GTVref was found to be significantly correlated with GTVis (I) p < 0.001, r = 0.989 (III) p = 0.001, r = 0.916, and (IV) p < 0.001, r = 0.986, but not with (II) p = 0.091, r = 0.595. The Sørensen-Dice indices for the semi-automatic tools were 0.74 (I), 0.57 (II) and 0.71 (III). For the semi-automatic, conventional segmentation tools evaluated, the adaptive-brush function (I) performed closest to the reference standard (0). The automatic deep learning tool (IV) showed high performance for auto-segmentation and was close to the reference standard. For high precision radiation therapy, visual control, and, where necessary, manual correction, are mandatory for all evaluated tools.

Creating high-resolution 3D cranial implant geometry using deep learning techniques

Creating a personalized implant for cranioplasty can be costly and aesthetically challenging, particularly for comminuted fractures that affect a wide area. Despite significant advances in deep learning techniques for 2D image completion, generating a 3D shape inpainting remains challenging due to the higher dimensionality and computational demands for 3D skull models. Here, we present a practical deep-learning approach to generate implant geometry from defective 3D skull models created from CT scans. Our proposed 3D reconstruction system comprises two neural networks that produce high-quality implant models suitable for clinical use while reducing training time. The first network repairs low-resolution defective models, while the second network enhances the volumetric resolution of the repaired model. We have tested our method in simulations and real-life surgical practices, producing implants that fit naturally and precisely match defect boundaries, particularly for skull defects above the Frankfort horizontal plane.

Reproducibility evaluation of the effects of MRI defacing on brain segmentation

Purpose

Recent advances in magnetic resonance (MR) scanner quality and the rapidly improving nature of facial recognition software have necessitated the introduction of MR defacing algorithms to protect patient privacy. As a result, there are a number of MR defacing algorithms available to the neuroimaging community, with several appearing in just the last 5 years. While some qualities of these defacing algorithms, such as patient identifiability, have been explored in the previous works, the potential impact of defacing on neuroimage processing has yet to be explored.

Approach

We qualitatively evaluate eight MR defacing algorithms on 179 subjects from the OASIS-3 cohort and 21 subjects from the Kirby-21 dataset. We also evaluate the effects of defacing on two neuroimaging pipelines-SLANT and FreeSurfer-by comparing the segmentation consistency between the original and defaced images.

Results

Defacing can alter brain segmentation and even lead to catastrophic failures, which are more frequent with some algorithms, such as Quickshear, MRI_Deface, and FSL_deface. Compared to FreeSurfer, SLANT is less affected by defacing. On outputs that pass the quality check, the effects of defacing are less pronounced than those of rescanning, as measured by the Dice similarity coefficient.

Conclusions

The effects of defacing are noticeable and should not be disregarded. Extra attention, in particular, should be paid to the possibility of catastrophic failures. It is crucial to adopt a robust defacing algorithm and perform a thorough quality check before releasing defaced datasets. To improve the reliability of analysis in scenarios involving defaced MRIs, it is encouraged to include multiple brain segmentation pipelines.

Lung nodule segmentation via semi-residual multi-resolution neural networks

The integration of deep neural networks and cloud computing has become increasingly prevalent within the domain of medical image processing, facilitated by the recent strides in neural network theory and the advent of the internet of things (IoTs). This juncture has led to the emergence of numerous image segmentation networks and innovative solutions that facilitate medical practitioners in diagnosing lung cancer. Within the contours of this study, we present an end-to-end neural network model, christened as the "semi-residual Multi-resolution Convolutional Neural Network" (semi-residual MCNN), devised to engender precise lung nodule segmentation maps within the milieu of cloud computing. Central to the architecture are three pivotal features, each coalescing to effectuate a notable enhancement in predictive accuracy: the incorporation of semi-residual building blocks, the deployment of group normalization techniques, and the orchestration of multi-resolution output heads. This innovative model is systematically subjected to rigorous training and testing regimes, using the LIDC-IDRI dataset - a widely embraced and accessible repository - comprising a diverse ensemble of 1,018 distinct lung CT images tailored to the realm of lung nodule segmentation.

Automated Triage of Screening Breast MRI Examinations in High-Risk Women Using an Ensemble Deep Learning Model

OBJECTIVES

The aim of the study is to develop and evaluate the performance of a deep learning (DL) model to triage breast magnetic resonance imaging (MRI) findings in high-risk patients without missing any cancers.

MATERIALS AND METHODS

In this retrospective study, 16,535 consecutive contrast-enhanced MRIs performed in 8354 women from January 2013 to January 2019 were collected. From 3 New York imaging sites, 14,768 MRIs were used for the training and validation data set, and 80 randomly selected MRIs were used for a reader study test data set. From 3 New Jersey imaging sites, 1687 MRIs (1441 screening MRIs and 246 MRIs performed in recently diagnosed breast cancer patients) were used for an external validation data set. The DL model was trained to classify maximum intensity projection images as "extremely low suspicion" or "possibly suspicious." Deep learning model evaluation (workload reduction, sensitivity, specificity) was performed on the external validation data set, using a histopathology reference standard. A reader study was performed to compare DL model performance to fellowship-trained breast imaging radiologists.

RESULTS

In the external validation data set, the DL model triaged 159/1441 of screening MRIs as "extremely low suspicion" without missing a single cancer, yielding a workload reduction of 11%, a specificity of 11.5%, and a sensitivity of 100%. The model correctly triaged 246/246 (100% sensitivity) of MRIs in recently diagnosed patients as "possibly suspicious." In the reader study, 2 readers classified MRIs with a specificity of 93.62% and 91.49%, respectively, and missed 0 and 1 cancer, respectively. On the other hand, the DL model classified MRIs with a specificity of 19.15% and missed 0 cancers, highlighting its potential use not as an independent reader but as a triage tool.

CONCLUSIONS

Our automated DL model triages a subset of screening breast MRIs as "extremely low suspicion" without misclassifying any cancer cases. This tool may be used to reduce workload in standalone mode, to shunt low suspicion cases to designated radiologists or to the end of the workday, or to serve as base model for other downstream AI tools.

Application of a 1H Brain MRS Benchmark Dataset to Deep Learning for Out-of-Voxel Artifacts

Neural networks are potentially valuable for many of the challenges associated with MRS data. The purpose of this manuscript is to describe the AGNOSTIC dataset, which contains 259,200 synthetic 1H MRS examples for training and testing neural networks. AGNOSTIC was created using 270 basis sets that were simulated across 18 field strengths and 15 echo times. The synthetic examples were produced to resemble in vivo brain data with combinations of metabolite, macromolecule, residual water signals, and noise. To demonstrate the utility, we apply AGNOSTIC to train two Convolutional Neural Networks (CNNs) to address out-of-voxel (OOV) echoes. A Detection Network was trained to identify the point-wise presence of OOV echoes, providing proof of concept for real-time detection. A Prediction Network was trained to reconstruct OOV echoes, allowing subtraction during post-processing. Complex OOV signals were mixed into 85% of synthetic examples to train two separate CNNs for the detection and prediction of OOV signals. AGNOSTIC is available through Dryad and all Python 3 code is available through GitHub. The Detection network was shown to perform well, identifying 95% of OOV echoes. Traditional modeling of these detected OOV signals was evaluated and may prove to be an effective method during linear-combination modeling. The Prediction Network greatly reduces OOV echoes within FIDs and achieved a median log10 normed-MSE of -1.79, an improvement of almost two orders of magnitude.

Segmenting functional tissue units across human organs using community-driven development of generalizable machine learning algorithms

The development of a reference atlas of the healthy human body requires automated image segmentation of major anatomical structures across multiple organs based on spatial bioimages generated from various sources with differences in sample preparation. We present the setup and results of the Hacking the Human Body machine learning algorithm development competition hosted by the Human Biomolecular Atlas (HuBMAP) and the Human Protein Atlas (HPA) teams on the Kaggle platform. We create a dataset containing 880 histology images with 12,901 segmented structures, engaging 1175 teams from 78 countries in community-driven, open-science development of machine learning models. Tissue variations in the dataset pose a major challenge to the teams which they overcome by using color normalization techniques and combining vision transformers with convolutional models. The best model will be productized in the HuBMAP portal to process tissue image datasets at scale in support of Human Reference Atlas construction.

Survey of methods and principles in three-dimensional reconstruction from two-dimensional medical images

Three-dimensional (3D) reconstruction of human organs has gained attention in recent years due to advances in the Internet and graphics processing units. In the coming years, most patient care will shift toward this new paradigm. However, development of fast and accurate 3D models from medical images or a set of medical scans remains a daunting task due to the number of pre-processing steps involved, most of which are dependent on human expertise. In this review, a survey of pre-processing steps was conducted, and reconstruction techniques for several organs in medical diagnosis were studied. Various methods and principles related to 3D reconstruction were highlighted. The usefulness of 3D reconstruction of organs in medical diagnosis was also highlighted.

Segmentation of human functional tissue units in support of a Human Reference Atlas

The Human BioMolecular Atlas Program (HuBMAP) aims to compile a Human Reference Atlas (HRA) for the healthy adult body at the cellular level. Functional tissue units (FTUs), relevant for HRA construction, are of pathobiological significance. Manual segmentation of FTUs does not scale; highly accurate and performant, open-source machine-learning algorithms are needed. We designed and hosted a Kaggle competition that focused on development of such algorithms and 1200 teams from 60 countries participated. We present the competition outcomes and an expanded analysis of the winning algorithms on additional kidney and colon tissue data, and conduct a pilot study to understand spatial location and density of FTUs across the kidney. The top algorithm from the competition, Tom, outperforms other algorithms in the expanded study, while using fewer computational resources. Tom was added to the HuBMAP infrastructure to run kidney FTU segmentation at scale-showcasing the value of Kaggle competitions for advancing research.

Spatiochemical Characterization of the Pancreas Using Mass Spectrometry Imaging and Topological Data Analysis

Mass Spectrometry Imaging (MSI) is a technique used to identify the spatial distribution of molecules in tissues. An MSI experiment results in large amounts of high dimensional data, so efficient computational methods are needed to analyze the output. Topological Data Analysis (TDA) has proven to be effective in all kinds of applications. TDA focuses on the topology of the data in high dimensional space. Looking at the shape in a high dimensional data set can lead to new or different insights. In this work, we investigate the use of Mapper, a form of TDA, applied on MSI data. Mapper is used to find data clusters inside two healthy mouse pancreas data sets. The results are compared to previous work using UMAP for MSI data analysis on the same data sets. This work finds that the proposed technique discovers the same clusters in the data as UMAP and is also able to uncover new clusters, such as an additional ring structure inside the pancreatic islets and a better defined cluster containing blood vessels. The technique can be used for a large variety of data types and sizes and can be optimized for specific applications. It is also computationally similar to UMAP for clustering. Mapper is a very interesting method, especially its use in biomedical applications.

Super-Resolution Reconstruction of Multi-Slice T2-W FLAIR MRI Improves Multiple Sclerosis Lesion Segmentation

Due to acquisition time constraints, T2-w FLAIR MRI of Multiple Sclerosis (MS) patients is often acquired with multi-slice 2D protocols with a low through-plane resolution rather than with high-resolution 3D protocols. Automated lesion segmentation on such low-resolution (LR) images, however, performs poorly and leads to inaccurate lesion volume estimates. Super-resolution reconstruction (SRR) methods can then be used to obtain a high-resolution (HR) image from multiple LR images to serve as input for lesion segmentation. In this work, we evaluate the effect on MS lesion segmentation of three SRR approaches: one based on interpolation, a state-of-the-art self-supervised CNN-based strategy, and a recently proposed model-based SRR method. These SRR strategies were applied to LR acquisitions simulated from 3D T2-w FLAIR MRI of MS patients. Each SRR method was evaluated in terms of image reconstruction quality and subsequent lesion segmentation performance. When compared to segmentation on LR images, the three considered SRR strategies demonstrate improved lesion segmentation. Furthermore, in some scenarios, SRR achieves a similar segmentation performance compared to segmentation of HR images.Clinical relevance- This study demonstrates the positive impact of super-resolution reconstruction from T2-w FLAIR multi-slice MRI acquisitions on segmentation performance of MS lesions.

A reproducibility evaluation of the effects of MRI defacing on brain segmentation

Purpose

Recent advances in magnetic resonance (MR) scanner quality and the rapidly improving nature of facial recognition software have necessitated the introduction of MR defacing algorithms to protect patient privacy. As a result, there are a number of MR defacing algorithms available to the neuroimaging community, with several appearing in just the last five years. While some qualities of these defacing algorithms, such as patient identifiability, have been explored in previous works, the potential impact of defacing on neuroimage processing has yet to be explored.

Approach

We qualitatively evaluate eight MR defacing algorithms on 179 subjects from the OASIS-3 cohort and the 21 subjects from the Kirby-21 dataset. We also evaluate the effects of defacing on two neuroimaging pipelines-SLANT and FreeSurfer-by comparing the segmentation consistency between the original and defaced images.

Results

Defacing can alter brain segmentation and even lead to catastrophic failures, which are more frequent with some algorithms such as Quickshear, MRI_Deface, and FSL_deface. Compared to FreeSurfer, SLANT is less affected by defacing. On outputs that pass the quality check, the effects of defacing are less pronounced than those of rescanning, as measured by the Dice similarity coefficient.

Conclusions

The effects of defacing are noticeable and should not be disregarded. Extra attention, in particular, should be paid to the possibility of catastrophic failures. It is crucial to adopt a robust defacing algorithm and perform a thorough quality check before releasing defaced datasets. To improve the reliability of analysis in scenarios involving defaced MRIs, it's encouraged to include multiple brain segmentation pipelines.

Overview of the HECKTOR Challenge at MICCAI 2022: Automatic Head and Neck Tumor Segmentation and Outcome Prediction in PET/CT

This paper presents an overview of the third edition of the HEad and neCK TumOR segmentation and outcome prediction (HECKTOR) challenge, organized as a satellite event of the 25th International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI) 2022. The challenge comprises two tasks related to the automatic analysis of FDG-PET/CT images for patients with Head and Neck cancer (H&N), focusing on the oropharynx region. Task 1 is the fully automatic segmentation of H&N primary Gross Tumor Volume (GTVp) and metastatic lymph nodes (GTVn) from FDG-PET/CT images. Task 2 is the fully automatic prediction of Recurrence-Free Survival (RFS) from the same FDG-PET/CT and clinical data. The data were collected from nine centers for a total of 883 cases consisting of FDG-PET/CT images and clinical information, split into 524 training and 359 test cases. The best methods obtained an aggregated Dice Similarity Coefficient (DSCagg) of 0.788 in Task 1, and a Concordance index (C-index) of 0.682 in Task 2.

Improved Selectivity in 7 T Digit Mapping Using VASO-CBV

Functional magnetic resonance imaging (fMRI) at Ultra-high field (UHF, ≥ 7 T) benefits from significant gains in the BOLD contrast-to-noise ratio (CNR) and temporal signal-to-noise ratio (tSNR) compared to conventional field strengths (3 T). Although these improvements enabled researchers to study the human brain to unprecedented spatial resolution, the blood pooling effect reduces the spatial specificity of the widely-used gradient-echo BOLD acquisitions. In this context, vascular space occupancy (VASO-CBV) imaging may be advantageous since it is proposed to have a higher spatial specificity than BOLD. We hypothesized that the assumed higher specificity of VASO-CBV imaging would translate to reduced overlap in fine-scale digit representation maps compared to BOLD-based digit maps. We used sub-millimeter resolution VASO fMRI at 7 T to map VASO-CBV and BOLD responses simultaneously in the motor and somatosensory cortices during individual finger movement tasks. We assessed the cortical overlap in different ways, first by calculating similarity coefficient metrics (DICE and Jaccard) and second by calculating selectivity measures. In addition, we demonstrate a consistent topographical organization of the targeted digit representations (thumb-index-little finger) in the motor areas. We show that the VASO-CBV responses yielded less overlap between the digit clusters than BOLD, and other selectivity measures were higher for VASO-CBV too. In summary, these results were consistent across metrics and participants, confirming the higher spatial specificity of VASO-CBV compared to BOLD.

Fully automated detection and localization of clinically significant prostate cancer on MR images using a cascaded convolutional neural network

Purpose

To develop a cascaded deep learning model trained with apparent diffusion coefficient (ADC) and T2-weighted imaging (T2WI) for fully automated detection and localization of clinically significant prostate cancer (csPCa).

Methods

This retrospective study included 347 consecutive patients (235 csPCa, 112 non-csPCa) with high-quality prostate MRI data, which were randomly selected for training, validation, and testing. The ground truth was obtained using manual csPCa lesion segmentation, according to pathological results. The proposed cascaded model based on Res-UNet takes prostate MR images (T2WI+ADC or only ADC) as inputs and automatically segments the whole prostate gland, the anatomic zones, and the csPCa region step by step. The performance of the models was evaluated and compared with PI-RADS (version 2.1) assessment using sensitivity, specificity, accuracy, and Dice similarity coefficient (DSC) in the held-out test set.

Results

In the test set, the per-lesion sensitivity of the biparametric (ADC + T2WI) model, ADC model, and PI-RADS assessment were 95.5% (84/88), 94.3% (83/88), and 94.3% (83/88) respectively (all p > 0.05). Additionally, the mean DSC based on the csPCa lesions were 0.64 ± 0.24 and 0.66 ± 0.23 for the biparametric model and ADC model, respectively. The sensitivity, specificity, and accuracy of the biparametric model were 95.6% (108/113), 91.5% (665/727), and 92.0% (773/840) based on sextant, and were 98.6% (68/69), 64.8% (46/71), and 81.4% (114/140) based on patients. The biparametric model had a similar performance to PI-RADS assessment (p > 0.05) and had higher specificity than the ADC model (86.8% [631/727], p< 0.001) based on sextant.

Conclusion

The cascaded deep learning model trained with ADC and T2WI achieves good performance for automated csPCa detection and localization.

A joint ventricle and WMH segmentation from MRI for evaluation of healthy and pathological changes in the aging brain

Age-related changes in brain structure include atrophy of the brain parenchyma and white matter changes of presumed vascular origin. Enlargement of the ventricles may occur due to atrophy or impaired cerebrospinal fluid (CSF) circulation. The co-occurrence of these changes in neurodegenerative diseases and in aging brains often requires investigators to take both into account when studying the brain, however, automated segmentation of enlarged ventricles and white matter hyperintensities (WMHs) can be a challenging task. Here, we present a hybrid multi-atlas segmentation and convolutional autoencoder approach for joint ventricle parcellation and WMH segmentation from magnetic resonance images (MRIs). Our fully automated approach uses a convolutional autoencoder to generate a standardized image of grey matter, white matter, CSF, and WMHs, which, in conjunction with labels generated by a multi-atlas segmentation approach, is then fed into a convolutional neural network to parcellate the ventricular system. Hence, our approach does not depend on manually delineated training data for new data sets. The segmentation pipeline was validated on both healthy elderly subjects and subjects with normal pressure hydrocephalus using ground truth manual labels and compared with state-of-the-art segmentation methods. We then applied the method to a cohort of 2401 elderly brains to investigate associations of ventricle volume and WMH load with various demographics and clinical biomarkers, using a multiple regression model. Our results indicate that the ventricle volume and WMH load are both highly variable in a cohort of elderly subjects and there is an independent association between the two, which highlights the importance of taking both the possibility of enlarged ventricles and WMHs into account when studying the aging brain.

Bayesian deep learning outperforms clinical trial estimators of intracerebral and intraventricular hemorrhage volume

BACKGROUND AND PURPOSE

Intracerebral hemorrhage (ICH) and intraventricular hemorrhage (IVH) clinical trials rely on manual linear and semi-quantitative (LSQ) estimators like the ABC/2, modified Graeb and IVH scores for timely volumetric estimation from CT. Deep learning (DL) volumetrics of ICH have recently approached the accuracy of gold-standard planimetry. However, DL and LSQ strategies have been limited by unquantified uncertainty, in particular when ICH and IVH estimates intersect. Bayesian deep learning methods can be used to approximate uncertainty, presenting an opportunity to improve quality assurance in clinical trials.

METHODS

A DL model was trained to simultaneously segment ICH and IVH using diagnostic CT data from the Minimally Invasive Surgery Plus Alteplase for ICH Evacuation (MISTIE) III and Clot Lysis: Evaluating Accelerated Resolution of IVH (CLEAR) III clinical trials. Bayesian uncertainty approximation was performed using Monte-Carlo dropout. We compared the performance of our model with estimators used in the CLEAR IVH and MISTIE II trials. The reliability of planimetry, DL, and LSQ volumetrics in the setting of high ICH and IVH intersection is quantified using consensus estimates.

RESULTS

Our DL model produced volume correlations and median Dice scores of .994 and .946 for ICH in MISTIE II, and .980 and .863 for IVH in CLEAR IVH, respectively, outperforming LSQ estimates from the clinical trials. We found significant linear relationships between ICH uncertainty, Dice scores (r = -.849), and relative volume difference (r = .735).

CONCLUSION

In our validation clinical trial dataset, DL models with Bayesian uncertainty approximation provided superior volumetric estimates to LSQ methods with real-time estimates of model uncertainty.

Differential Multimolecule Fingerprint for Similarity Search─Making Use of Active and Inactive Compound Sets in Virtual Screening

In conventional fingerprint methods, the similarity between two molecules is calculated using the Tanimoto index as a numerical criterion. Thus, the query molecules in virtual screening should be most representative of the wanted compound class at hand. In the concept introduced here, all available active molecules form a multimolecule fingerprint in which the appearing features are weighted according to their respective frequency. The features of inactive molecules are treated likewise and the resulting values are subtracted from those of the active ones. The obtained differential multimolecule fingerprint (DMMFP) is thus specific for the respective class of compounds. To account for the noninteger representation within this fingerprint, a modified Sørensen-Dice coefficient is used to compute the similarity. Potentially active molecules yield positive scores, whereas presumably inactive ones are denoted by negative values. The concept was applied to Angiotensin-converting enzyme (ACE) inhibitors, β2-adrenoceptor ligands, leukotriene A4 hydrolase inhibitors, dopamine D3 antagonists, and cytochrome CYP2C9 substrates, for which experimental binding affinities are known and was tested against decoys from DUD-E and a further background database consisting of molecules from the dark chemical matter, which comprises compounds that appear as frequent hitters across multiple assays. Using the 166 publicly available keys of the MACCS fingerprint and the larger PubChem fingerprint, actives were recovered with very high sensitivity. Furthermore, three marketed ACE inhibitors as well as the carbonic anhydrase II inhibitor dorzolamide were detected in the dark chemical matter data set. For comparison, the DMMFP was also used with a Bayesian classifier, for which the specificity (correctly classified inactives) and likewise the accuracy was superior. Conversely, the similarity score produced by the Sørensen-Dice coefficient showed its potential for the early recognition of (potentially) active molecules.

An Ensemble of U-Net Models for Kidney Tumor Segmentation With CT Images

We present here the Arkansas AI-Campus solution method for the 2019 Kidney Tumor Segmentation Challenge (KiTS19). Our Arkansas AI-Campus team participated the KiTS19 Challenge for four months, from March to July of 2019. This paper provides a summary of our methods, training, testing and validation results for this grand challenge in biomedical imaging analysis. Our deep learning model is an ensemble of U-Net models developed after testing many model variations. Our model has consistent performance on the local test dataset and the final competition independent test dataset. The model achieved local test Dice scores of 0.949 for kidney and tumor segmentation, and 0.601 for tumor segmentation, and the final competition test earned Dice scores 0.9470 and 0.6099 respectively. The Arkansas AI-Campus team solution with a composite DICE score of 0.7784 has achieved a final ranking of top fifty worldwide, and top five among the United States teams in the KiTS19 Competition.

166Holmium-99mTechnetium dual-isotope imaging: scatter compensation and automatic healthy-liver segmentation for 166Holmium radioembolization dosimetry

Background

Partition modeling allows personalized activity calculation for holmium-166 (¹⁶⁶Ho) radioembolization. However, it requires the definition of tumor and non-tumorous liver, by segmentation and registration of a separately acquired CT, which is time-consuming and prone to error. A protocol including ¹⁶⁶Ho-scout, for treatment simulation, and technetium-99m (^99mTc) stannous phytate for healthy-liver delineation was proposed. This study assessed the accuracy of automatic healthy-liver segmentation using ^99mTc images derived from a phantom experiment. In addition, together with data from a patient study, the effect of different ^99mTc activities on the ¹⁶⁶Ho-scout images was investigated. To reproduce a typical scout procedure, the liver compartment, including two tumors, of an anthropomorphic phantom was filled with 250 MBq of ¹⁶⁶Ho-chloride, with a tumor to non-tumorous liver activity concentration ratio of 10. Eight SPECT/CT scans were acquired, with varying levels of ^99mTc added to the non-tumorous liver compartment (ranging from 25 to 126 MBq). For comparison, forty-two scans were performed in presence of only ^99mTc from 8 to 240 MBq. ^99mTc image quality was assessed by cold-sphere (tumor) contrast recovery coefficients. Automatic healthy-liver segmentation, obtained by thresholding ^99mTc images, was evaluated by recovered volume and Sørensen–Dice index. The impact of ^99mTc on ¹⁶⁶Ho images and the role of the downscatter correction were evaluated on phantom scans and twenty-six patients’ scans by considering the reconstructed ¹⁶⁶Ho count density in the healthy-liver.

Results

All ^99mTc image reconstructions were found to be independent of the ¹⁶⁶Ho activity present during the acquisition. In addition, cold-sphere contrast recovery coefficients were independent of ^99mTc activity. The segmented healthy-liver volume was recovered fully, independent of ^99mTc activity as well. The reconstructed ¹⁶⁶Ho count density was not influenced by ^99mTc activity, as long as an adequate downscatter correction was applied.

Conclusion

The ^99mTc image reconstructions of the phantom scans all performed equally well for the purpose of automatic healthy-liver segmentation, for activities down to 8 MBq. Furthermore, ^99mTc could be injected up to at least 126 MBq without compromising ¹⁶⁶Ho image quality.

Clinical trials The clinical study mentioned is registered with Clinicaltrials.gov (NCT02067988) on February 20, 2014.

Correlation and colocalization of HIF-1α and pimonidazole staining for hypoxia in laryngeal squamous cell carcinomas: A digital, single-cell-based analysis

OBJECTIVE

Tumor hypoxia results in worse local control and patient survival. We performed a digital, single-cell-based analysis to compare two biomarkers for hypoxia (hypoxia-inducible factor 1-alpha [HIF-1α] and pimonidazole [PIMO]) and their effect on outcome in laryngeal cancer patients treated with accelerated radiotherapy with or without carbogen breathing and nicotinamide (AR versus ARCON).

MATERIALS AND METHODS

Immunohistochemical staining was performed for HIF-1α and PIMO in consecutive sections of 44 laryngeal cancer patients randomized between AR and ARCON. HIF-1α expression and PIMO-binding were correlated using digital image analysis in QuPath. High-density areas for each biomarker were automatically annotated and staining overlap was analyzed. Kaplan-Meier survival analyses for local control, regional control and disease-free survival were performed to predict a response benefit of ARCON over AR alone for each biomarker.

RESULTS

106 Tissue fragments of 44 patients were analyzed. A weak, significant positive correlation was observed between HIF-1α and PIMO positivity on fragment level, but not on patient level. A moderate strength correlation (r = 0.705, p < 0.001) was observed between the number of high-density staining areas for both biomarkers. Staining overlap was poor. HIF-1α expression, PIMO-binding or a combination could not predict a response benefit of ARCON over AR.

CONCLUSION

Digital image analysis to compare positive cell fractions and staining overlap between two hypoxia biomarkers using open-source software is feasible. Our results highlight that there are distinct differences between HIF-1α and PIMO as hypoxia biomarkers and therefore suggest co-existence of different forms of hypoxia within a single tumor.

DADP: Dynamic abnormality detection and progression for longitudinal knee magnetic resonance images from the Osteoarthritis Initiative

Osteoarthritis (OA) is the most common disabling joint disease. Magnetic resonance (MR) imaging has been commonly used to assess knee joint degeneration due to its distinct advantage in detecting morphologic cartilage changes. Although several statistical methods over conventional radiography have been developed to perform quantitative cartilage analyses, little work has been done capturing the development and progression of cartilage lesions (or abnormal regions) and how they naturally progress. There are two major challenges, including (i) the lack of building spatial-temporal correspondences and correlations in cartilage thickness and (ii) the spatio-temporal heterogeneity in abnormal regions. The goal of this work is to propose a dynamic abnormality detection and progression (DADP) framework for quantitative cartilage analysis, while addressing the two challenges. First, spatial correspondences are established on flattened 2D cartilage thickness maps extracted from 3D knee MR images both across time within each subject and across all subjects. Second, a dynamic functional mixed effects model (DFMEM) is proposed to quantify abnormality progression across time points and subjects, while accounting for the spatio-temporal heterogeneity. We systematically evaluate our DADP using simulations and real data from the Osteoarthritis Initiative (OAI). Our results show that DADP not only effectively detects subject-specific dynamic abnormal regions, but also provides population-level statistical disease mapping and subgroup analysis.

Study on Body Size Measurement Method of Goat and Cattle under Different Background Based on Deep Learning

The feasibility of using depth sensors to measure the body size of livestock has been extensively tested. Most existing methods are only capable of measuring the body size of specific livestock in a specific background. In this study, we proposed a unique method of livestock body size measurement using deep learning. By training the data of cattle and goat with same feature points, different animal sizes can be measured under different backgrounds. First, a novel penalty function and an autoregressive model were introduced to reconstruct the depth image with super-resolution, and the effect of distance and illumination on the depth image was reduced. Second, under the U-Net neural network, the characteristics exhibited by the attention module and the DropBlock were adopted to improve the robustness of the background and trunk segmentation. Lastly, this study initially exploited the idea of human joint point location to accurately locate the livestock body feature points, and the livestock was accurately measured. According to the results, the average accuracy of this method was 93.59%. The correct key points for detecting the points of withers, shoulder points, shallowest part of the chest, highest point of the hip bones and ischia tuberosity had the percentages of 96.7%, 89.3%, 95.6%, 90.5% and 94.5%, respectively. In addition, the mean relative errors of withers height, hip height, body length and chest depth were only 1.86%, 2.07%, 2.42% and 2.72%, respectively.

Evaluation of a pipeline for simulation, reconstruction, and classification in ultrasound-aided diffuse optical tomography of breast tumors

SIGNIFICANCE

Diffuse optical tomography is an ill-posed problem. Combination with ultrasound can improve the results of diffuse optical tomography applied to the diagnosis of breast cancer and allow for classification of lesions.

AIM

To provide a simulation pipeline for the assessment of reconstruction and classification methods for diffuse optical tomography with concurrent ultrasound information.

APPROACH

A set of breast digital phantoms with benign and malignant lesions was simulated building on the software VICTRE. Acoustic and optical properties were assigned to the phantoms for the generation of B-mode images and optical data. A reconstruction algorithm based on a two-region nonlinear fitting and incorporating the ultrasound information was tested. Machine learning classification methods were applied to the reconstructed values to discriminate lesions into benign and malignant after reconstruction.

RESULTS

The approach allowed us to generate realistic US and optical data and to test a two-region reconstruction method for a large number of realistic simulations. When information is extracted from ultrasound images, at least 75% of lesions are correctly classified. With ideal two-region separation, the accuracy is higher than 80%.

CONCLUSIONS

A pipeline for the generation of realistic ultrasound and diffuse optics data was implemented. Machine learning methods applied to a optical reconstruction with a nonlinear optical model and morphological information permit to discriminate malignant lesions from benign ones.

Bayesian spatial models for voxel-wise prostate cancer classification using multi-parametric magnetic resonance imaging data

Multi-parametric magnetic resonance imaging (mpMRI) has been playing an increasingly important role in the detection of prostate cancer (PCa). Various computer-aided detection algorithms were proposed for automated PCa detection by combining information in multiple mpMRI parameters. However, there are specific features of mpMRI, including between-voxel correlation within each prostate and heterogeneity across patients, that have not been fully explored but could potentially improve PCa detection if leveraged appropriately. This article proposes novel Bayesian approaches for voxel-wise PCa classification that accounts for spatial correlation and between-patient heterogeneity in the mpMRI data. Modeling the spatial correlation is challenging due to the extreme high dimensionality of the data, and we propose three scalable approaches based on Nearest Neighbor Gaussian Process (NNGP), reduced-rank approximation, and a conditional autoregressive (CAR) model that approximates a Gaussian Process with the Matérn covariance, respectively. Our simulation study shows that properly modeling the spatial correlation and between-patient heterogeneity can substantially improve PCa classification. Application to in vivo data illustrates that classification is improved by all three spatial modeling approaches considered, while modeling the between-patient heterogeneity does not further improve our classifiers. Among the proposed models, the NNGP-based model is recommended given its high classification accuracy and computational efficiency.

Deep Learning Networks for Automatic Retroperitoneal Sarcoma Segmentation in Computerized Tomography

The volume estimation of retroperitoneal sarcoma (RPS) is often difficult due to its huge dimensions and irregular shape; thus, it often requires manual segmentation, which is time-consuming and operator-dependent. This study aimed to evaluate two fully automated deep learning networks (ENet and ERFNet) for RPS segmentation. This retrospective study included 20 patients with RPS who received an abdominal computed tomography (CT) examination. Forty-nine CT examinations, with a total of 72 lesions, were included. Manual segmentation was performed by two radiologists in consensus, and automatic segmentation was performed using ENet and ERFNet. Significant differences between manual and automatic segmentation were tested using the analysis of variance (ANOVA). A set of performance indicators for the shape comparison (namely sensitivity), positive predictive value (PPV), dice similarity coefficient (DSC), volume overlap error (VOE), and volumetric differences (VD) were calculated. There were no significant differences found between the RPS volumes obtained using manual segmentation and ENet (p-value = 0.935), manual segmentation and ERFNet (p-value = 0.544), or ENet and ERFNet (p-value = 0.119). The sensitivity, PPV, DSC, VOE, and VD for ENet and ERFNet were 91.54% and 72.21%, 89.85% and 87.00%, 90.52% and 74.85%, 16.87% and 36.85%, and 2.11% and -14.80%, respectively. By using a dedicated GPU, ENet took around 15 s for segmentation versus 13 s for ERFNet. In the case of CPU, ENet took around 2 min versus 1 min for ERFNet. The manual approach required approximately one hour per segmentation. In conclusion, fully automatic deep learning networks are reliable methods for RPS volume assessment. ENet performs better than ERFNet for automatic segmentation, though it requires more time.

Metallic Component Preserving Algorithm Based on the Cerebral Computed Tomography Angiography in Aneurysm Surgery

The purpose of this study was to investigate the viability of the proposed method in preventing the loss of metallic components including the clip and coil in cerebral computed tomography angiography (CTA). Forty patients undergoing surgery for aneurysms carried metallic materials. The proposed method is based on conventional bone subtraction CTA (BS-CTA) system. Briefly, the position of metal components was determined using the threshold value and a region of interest (ROI). An appropriate threshold was used to separate the background from the target materials based on the Otsu method. A three-dimensional (3D) rendering was performed from the proposed BS-CTA data carrying the extracted target information. The accuracy of clip and coil region measured using the dice similarity coefficient (DSC) and bidirectional Hausdorff distance (HD) is reported. The metallic components of the proposed BS-CTA were significantly visualized in various patient cases. Quantitative evaluation using the proposed method is based on the mean DSC of 0.93 with a standard deviation (SD) of ±0.05 (e.g., maximum value = 0.99, minimum value = 0.75, 95% confidence interval (CI) = 0.91 to 0.95, and all p < 0.05). The mean HD was 1.50 voxels with an SD of ± 0.58 (e.g., maximum value = 5.95, minimum value = 0.12, 95% CI = 1.10 to 1.90, and all p < 0.05). The proposed method demonstrates effective segmentation of the metallic component and application to the existing conventional BS-CTA system.

Multi-Path U-Net Architecture for Cell and Colony-Forming Unit Image Segmentation

U-Net is the most cited and widely-used deep learning model for biomedical image segmentation. In this paper, we propose a new enhanced version of a ubiquitous U-Net architecture, which improves upon the original one in terms of generalization capabilities, while addressing several immanent shortcomings, such as constrained resolution and non-resilient receptive fields of the main pathway. Our novel multi-path architecture introduces a notion of an individual receptive field pathway, which is merged with other pathways at the bottom-most layer by concatenation and subsequent application of Layer Normalization and Spatial Dropout, which can improve generalization performance for small datasets. In general, our experiments show that the proposed multi-path architecture outperforms other state-of-the-art approaches that embark on similar ideas of pyramid structures, skip-connections, and encoder-decoder pathways. A significant improvement of the Dice similarity coefficient is attained at our proprietary colony-forming unit dataset, where a score of 0.809 was achieved for the foreground class.

The Effect of Training Sample Size on the Prediction of White Matter Hyperintensity Volume in a Healthy Population Using BIANCA

Introduction: White matter hyperintensities of presumed vascular origin (WMH) are an important magnetic resonance imaging marker of cerebral small vessel disease and are associated with cognitive decline, stroke, and mortality. Their relevance in healthy individuals, however, is less clear. This is partly due to the methodological challenge of accurately measuring rare and small WMH with automated segmentation programs. In this study, we tested whether WMH volumetry with FMRIB software library v6.0 (FSL; https://fsl.fmrib.ox.ac.uk/fsl/fslwiki) Brain Intensity AbNormality Classification Algorithm (BIANCA), a customizable and trainable algorithm that quantifies WMH volume based on individual data training sets, can be optimized for a normal aging population.

Methods: We evaluated the effect of varying training sample sizes on the accuracy and the robustness of the predicted white matter hyperintensity volume in a population (n = 201) with a low prevalence of confluent WMH and a substantial proportion of participants without WMH. BIANCA was trained with seven different sample sizes between 10 and 40 with increments of 5. For each sample size, 100 random samples of T1w and FLAIR images were drawn and trained with manually delineated masks. For validation, we defined an internal and external validation set and compared the mean absolute error, resulting from the difference between manually delineated and predicted WMH volumes for each set. For spatial overlap, we calculated the Dice similarity index (SI) for the external validation cohort.

Results: The study population had a median WMH volume of 0.34 ml (IQR of 1.6 ml) and included n = 28 (18%) participants without any WMH. The mean absolute error of the difference between BIANCA prediction and manually delineated masks was minimized and became more robust with an increasing number of training participants. The lowest mean absolute error of 0.05 ml (SD of 0.24 ml) was identified in the external validation set with a training sample size of 35. Compared to the volumetric overlap, the spatial overlap was poor with an average Dice similarity index of 0.14 (SD 0.16) in the external cohort, driven by subjects with very low lesion volumes.

Discussion: We found that the performance of BIANCA, particularly the robustness of predictions, could be optimized for use in populations with a low WMH load by enlargement of the training sample size. Further work is needed to evaluate and potentially improve the prediction accuracy for low lesion volumes. These findings are important for current and future population-based studies with the majority of participants being normal aging people.

Dynamic topology analysis for spatial patterns of multifocal lesions on MRI

Quantitatively analysing the spatial patterns of multifocal lesions on clinical MRI is an important step towards a better understanding of the disease and for precision medicine, which is yet to be properly explored by feature engineering and deep learning methods. Network science addresses this issue by explicitly modeling the inter-lesion topology. However, the construction of the informative graph with optimal edge sparsity and quantification of community graph structures are the current challenges in network science. In this paper, we address these challenges with a novel Dynamic Topology Analysis framework on the basis of persistent homology, aiming to investigate the predictive values of global geometry and local clusters of multifocal lesions. Firstly, Dynamic Hierarchical Network is proposed to construct informative global and community-level topology over multi-scale networks from sparse to dense. Multi-scale global topology is constructed with a nested sequence of Rips complexes, from which a new K-simplex Filtration is designed to generate a higher-level topological abstraction for community identification based on the connectivity of k-simplices in the Rips Complex. Secondly, to quantify multi-scale community structures, we design a new Decomposed Community Persistence algorithm to track the dynamic evolution of communities, and then summarise the evolutionary communities incorporated with a customisable descriptor. The quantified community features are encapsulated with global geometric invariants for topological pattern analysis. The proposed framework was evaluated on both diagnostic differentiation and prognostic prediction for multiple sclerosis that is a typical multifocal disease, and achieved ROC_AUC 0.875 and 0.767, respectively, outperforming seven state-of-the-art persistent homology methods and the reported performance of six feature engineering and deep learning methods.

Automated segmentation by deep learning of loose connective tissue fibers to define safe dissection planes in robot-assisted gastrectomy

The prediction of anatomical structures within the surgical field by artificial intelligence (AI) is expected to support surgeons’ experience and cognitive skills. We aimed to develop a deep-learning model to automatically segment loose connective tissue fibers (LCTFs) that define a safe dissection plane. The annotation was performed on video frames capturing a robot-assisted gastrectomy performed by trained surgeons. A deep-learning model based on U-net was developed to output segmentation results. Twenty randomly sampled frames were provided to evaluate model performance by comparing Recall and F1/Dice scores with a ground truth and with a two-item questionnaire on sensitivity and misrecognition that was completed by 20 surgeons. The model produced high Recall scores (mean 0.606, maximum 0.861). Mean F1/Dice scores reached 0.549 (range 0.335–0.691), showing acceptable spatial overlap of the objects. Surgeon evaluators gave a mean sensitivity score of 3.52 (with 88.0% assigning the highest score of 4; range 2.45–3.95). The mean misrecognition score was a low 0.14 (range 0–0.7), indicating very few acknowledged over-detection failures. Thus, AI can be trained to predict fine, difficult-to-discern anatomical structures at a level convincing to expert surgeons. This technology may help reduce adverse events by determining safe dissection planes.

Realistic image normalization for multi-Domain segmentation

Image normalization is a building block in medical image analysis. Conventional approaches are customarily employed on a per-dataset basis. This strategy, however, prevents the current normalization algorithms from fully exploiting the complex joint information available across multiple datasets. Consequently, ignoring such joint information has a direct impact on the processing of segmentation algorithms. This paper proposes to revisit the conventional image normalization approach by, instead, learning a common normalizing function across multiple datasets. Jointly normalizing multiple datasets is shown to yield consistent normalized images as well as an improved image segmentation when intensity shifts are large. To do so, a fully automated adversarial and task-driven normalization approach is employed as it facilitates the training of realistic and interpretable images while keeping performance on par with the state-of-the-art. The adversarial training of our network aims at finding the optimal transfer function to improve both, jointly, the segmentation accuracy and the generation of realistic images. We have evaluated the performance of our normalizer on both infant and adult brain images from the iSEG, MRBrainS and ABIDE datasets. The results indicate that our contribution does provide an improved realism to the normalized images, while retaining a segmentation accuracy at par with the state-of-the-art learnable normalization approaches.

A Comprehensive Computer-Assisted Diagnosis System for Early Assessment of Renal Cancer Tumors

Renal cell carcinoma (RCC) is the most common and a highly aggressive type of malignant renal tumor. In this manuscript, we aim to identify and integrate the optimal discriminating morphological, textural, and functional features that best describe the malignancy status of a given renal tumor. The integrated discriminating features may lead to the development of a novel comprehensive renal cancer computer-assisted diagnosis (RC-CAD) system with the ability to discriminate between benign and malignant renal tumors and specify the malignancy subtypes for optimal medical management. Informed consent was obtained from a total of 140 biopsy-proven patients to participate in the study (male = 72 and female = 68, age range = 15 to 87 years). There were 70 patients who had RCC (40 clear cell RCC (ccRCC), 30 nonclear cell RCC (nccRCC)), while the other 70 had benign angiomyolipoma tumors. Contrast-enhanced computed tomography (CE-CT) images were acquired, and renal tumors were segmented for all patients to allow the extraction of discriminating imaging features. The RC-CAD system incorporates the following major steps: (i) applying a new parametric spherical harmonic technique to estimate the morphological features, (ii) modeling a novel angular invariant gray-level co-occurrence matrix to estimate the textural features, and (iii) constructing wash-in/wash-out slopes to estimate the functional features by quantifying enhancement variations across different CE-CT phases. These features were subsequently combined and processed using a two-stage multilayer perceptron artificial neural network (MLP-ANN) classifier to classify the renal tumor as benign or malignant and identify the malignancy subtype as well. Using the combined features and a leave-one-subject-out cross-validation approach, the developed RC-CAD system achieved a sensitivity of 95.3%±2.0%, a specificity of 99.9%±0.4%, and Dice similarity coefficient of 0.98±0.01 in differentiating malignant from benign tumors, as well as an overall accuracy of 89.6%±5.0% in discriminating ccRCC from nccRCC. The diagnostic abilities of the developed RC-CAD system were further validated using a randomly stratified 10-fold cross-validation approach. The obtained results using the proposed MLP-ANN classification model outperformed other machine learning classifiers (e.g., support vector machine, random forests, relational functional gradient boosting, etc.). Hence, integrating morphological, textural, and functional features enhances the diagnostic performance, making the proposal a reliable noninvasive diagnostic tool for renal tumors.

End-to-end prostate cancer detection in bpMRI via 3D CNNs: Effects of attention mechanisms, clinical priori and decoupled false positive reduction

We present a multi-stage 3D computer-aided detection and diagnosis (CAD) model² for automated localization of clinically significant prostate cancer (csPCa) in bi-parametric MR imaging (bpMRI). Deep attention mechanisms drive its detection network, targeting salient structures and highly discriminative feature dimensions across multiple resolutions. Its goal is to accurately identify csPCa lesions from indolent cancer and the wide range of benign pathology that can afflict the prostate gland. Simultaneously, a decoupled residual classifier is used to achieve consistent false positive reduction, without sacrificing high sensitivity or computational efficiency. In order to guide model generalization with domain-specific clinical knowledge, a probabilistic anatomical prior is used to encode the spatial prevalence and zonal distinction of csPCa. Using a large dataset of 1950 prostate bpMRI paired with radiologically-estimated annotations, we hypothesize that such CNN-based models can be trained to detect biopsy-confirmed malignancies in an independent cohort. For 486 institutional testing scans, the 3D CAD system achieves 83.69±5.22% and 93.19±2.96% detection sensitivity at 0.50 and 1.46 false positive(s) per patient, respectively, with 0.882±0.030 AUROC in patient-based diagnosis -significantly outperforming four state-of-the-art baseline architectures (U-SEResNet, UNet++, nnU-Net, Attention U-Net) from recent literature. For 296 external biopsy-confirmed testing scans, the ensembled CAD system shares moderate agreement with a consensus of expert radiologists (76.69%; kappa = 0.51±0.04) and independent pathologists (81.08%; kappa = 0.56±0.06); demonstrating strong generalization to histologically-confirmed csPCa diagnosis.

A novel computer-aided diagnostic system for accurate detection and grading of liver tumors

Liver cancer is a major cause of morbidity and mortality in the world. The primary goals of this manuscript are the identification of novel imaging markers (morphological, functional, and anatomical/textural), and development of a computer-aided diagnostic (CAD) system to accurately detect and grade liver tumors non-invasively. A total of 95 patients with liver tumors (M = 65, F = 30, age range = 34–82 years) were enrolled in the study after consents were obtained. 38 patients had benign tumors (LR1 = 19 and LR2 = 19), 19 patients had intermediate tumors (LR3), and 38 patients had hepatocellular carcinoma (HCC) malignant tumors (LR4 = 19 and LR5 = 19). A multi-phase contrast-enhanced magnetic resonance imaging (CE-MRI) was collected to extract the imaging markers. A comprehensive CAD system was developed, which includes the following main steps: i) estimation of morphological markers using a new parametric spherical harmonic model, ii) estimation of textural markers using a novel rotation invariant gray-level co-occurrence matrix (GLCM) and gray-level run-length matrix (GLRLM) models, and iii) calculation of the functional markers by estimating the wash-in/wash-out slopes, which enable quantification of the enhancement characteristics across different CE-MR phases. These markers were subsequently processed using a two-stages random forest-based classifier to classify the liver tumor as benign, intermediate, or malignant and determine the corresponding grade (LR1, LR2, LR3, LR4, or LR5). The overall CAD system using all the identified imaging markers achieved a sensitivity of 91.8%±0.9%, specificity of 91.2%±1.9%, and F\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{1}$$\end{document}1 score of 0.91±0.01, using the leave-one-subject-out (LOSO) cross-validation approach. Importantly, the CAD system achieved overall accuracies of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$88\%\pm 5\%$$\end{document}88%±5%, 85%±2%, 78%±3%, 83%±4%, and 79%±3% in grading liver tumors into LR1, LR2, LR3, LR4, and LR5, respectively. In addition to LOSO, the developed CAD system was tested using randomly stratified 10-fold and 5-fold cross-validation approaches. Alternative classification algorithms, including support vector machine, naive Bayes classifier, k-nearest neighbors, and linear discriminant analysis all produced inferior results compared to the proposed two stage random forest classification model. These experiments demonstrate the feasibility of the proposed CAD system as a novel tool to objectively assess liver tumors based on the new comprehensive imaging markers. The identified imaging markers and CAD system can be used as a non-invasive diagnostic tool for early and accurate detection and grading of liver cancer.

Quantitative analysis of bone microvasculature in a mouse model using the monogenic signal phase asymmetry and marker-controlled watershed

Thethree-dimensional (3D) imaging and quantitative analysis of bone microvasculature are important to describe angiogenesis involvement in bone metastatic processes. Here, we propose an algorithm based on marker-controlled watershed for the 3D segmentation of vessels and bone in mouse bone imaged with a contrast agent using synchrotron radiation micro-computed tomography (SR-μCT). Markers were generated using hysteresis thresholding and morphological filters, and the control surface was constructed using the monogenic signal phase asymmetry. The accuracy and robustness of the proposed method were evaluated on a series of synthetic volumes generated to mimic the vessel, bone and background structures. Different contrast between different structures, as well as different noise levels were considered. A series of multi-class synthetic volumes were segmented using the proposed method, and the overall segmentation quality was evaluated using the Matthews correlation coefficient (MCC) by comparing to the ground truth. Additionally, we evaluated the segmentation of thin structures under various levels of Gaussian noise. The simulation study indicated that the algorithm was performant in multi-class segmentation with different contrast, noise, and thickness. The algorithm was applied to images of bone from a mouse model of breast cancer bone metastasis acquired using SR-μCT. The segmentation quality was evaluated using the Dice coefficient and the MCC by comparing to manual segmentation. The proposed method performed better than hysteresis thresholding and marker-controlled watershed using the magnitude of the gradient as control surface. Several quantitative parameters on bone and vessels were extracted, including bone volume fraction (BV/TV), vessel volume fraction (VV/TV) and the mean vessel thickness (VTh). The bone volume fraction (BV/TV) was significantly lower in the metastatic group compared to the healthy group. This demonstrated the effectiveness of the algorithm for the study of bone and vessel microstructures in mouse model.

Automatic and unbiased segmentation and quantification of myofibers in skeletal muscle

Skeletal muscle has the remarkable ability to regenerate. However, with age and disease muscle strength and function decline. Myofiber size, which is affected by injury and disease, is a critical measurement to assess muscle health. Here, we test and apply Cellpose, a recently developed deep learning algorithm, to automatically segment myofibers within murine skeletal muscle. We first show that tissue fixation is necessary to preserve cellular structures such as primary cilia, small cellular antennae, and adipocyte lipid droplets. However, fixation generates heterogeneous myofiber labeling, which impedes intensity-based segmentation. We demonstrate that Cellpose efficiently delineates thousands of individual myofibers outlined by a variety of markers, even within fixed tissue with highly uneven myofiber staining. We created a novel ImageJ plugin (LabelsToRois) that allows processing of multiple Cellpose segmentation images in batch. The plugin also contains a semi-automatic erosion function to correct for the area bias introduced by the different stainings, thereby identifying myofibers as accurately as human experts. We successfully applied our segmentation pipeline to uncover myofiber regeneration differences between two different muscle injury models, cardiotoxin and glycerol. Thus, Cellpose combined with LabelsToRois allows for fast, unbiased, and reproducible myofiber quantification for a variety of staining and fixation conditions.

Orchard Mapping with Deep Learning Semantic Segmentation

This study aimed to propose an approach for orchard trees segmentation using aerial images based on a deep learning convolutional neural network variant, namely the U-net network. The purpose was the automated detection and localization of the canopy of orchard trees under various conditions (i.e., different seasons, different tree ages, different levels of weed coverage). The implemented dataset was composed of images from three different walnut orchards. The achieved variability of the dataset resulted in obtaining images that fell under seven different use cases. The best-trained model achieved 91%, 90%, and 87% accuracy for training, validation, and testing, respectively. The trained model was also tested on never-before-seen orthomosaic images or orchards based on two methods (oversampling and undersampling) in order to tackle issues with out-of-the-field boundary transparent pixels from the image. Even though the training dataset did not contain orthomosaic images, it achieved performance levels that reached up to 99%, demonstrating the robustness of the proposed approach.