Tissue segmentation: a crucial tool for quantitative MRI and visualization of anatomical structures

Accurate diagnostic tissue segmentation and concurrent disease subtyping with small datasets

Purpose

To provide a flexible, end-to-end platform for visually distinguishing diseased from undiseased tissue in a medical image, in particular pathology slides, and classifying diseased regions by subtype. Highly accurate results are obtained using small training datasets and reduced-scale source images that can be easily shared.

Approach

An ensemble of lightweight convolutional neural networks (CNNs) is trained on different subsets of images derived from a relatively small number of annotated whole-slide histopathology images (WSIs). The WSIs are first reduced in scale in a manner that preserves anatomic features critical to analysis while also facilitating convenient handling and storage. The segmentation and subtyping tasks are performed sequentially on the reduced-scale images using the same basic workflow: generating and sifting tiles from the image, then classifying each tile with an ensemble of appropriately trained CNNs. For segmentation, the CNN predictions are combined using a function to favor a selected similarity metric, and a mask or map for a a candidate image is produced from tiles whose combined predictions exceed a decision boundary. For subtyping, the resulting mask is applied to the candidate image, and new tiles are derived from the unoccluded regions. These are classified by the subtyping CNNs to produce an overall subtype prediction.

Results and conclusion

This approach was applied successfully to two very different datasets of large WSIs, one (PAIP2020) involving multiple subtypes of colorectal cancer and the other (CAMELYON16) single-type breast cancer metastases. Scored using standard similarity metrics, the segmentations outperformed more complex models typifying the state of the art.

Evaluation of a deep learning method for the automated detection of supraspinatus tears on MRI

OBJECTIVE

To evaluate if deep learning is a feasible approach for automated detection of supraspinatus tears on MRI.

MATERIALS AND METHODS

A total of 200 shoulder MRI studies performed between 2015 and 2019 were retrospectively obtained from our institutional database using a balanced random sampling of studies containing a full-thickness tear, partial-thickness tear, or intact supraspinatus tendon. A 3-stage pipeline was developed comprised of a slice selection network based on a pre-trained residual neural network (ResNet); a segmentation network based on an encoder-decoder network (U-Net); and a custom multi-input convolutional neural network (CNN) classifier. Binary reference labels were created following review of radiologist reports and images by a radiology fellow and consensus validation by two musculoskeletal radiologists. Twenty percent of the data was reserved as a holdout test set with the remaining 80% used for training and optimization under a fivefold cross-validation strategy. Classification and segmentation accuracy were evaluated using area under the receiver operating characteristic curve (AUROC) and Dice similarity coefficient, respectively. Baseline characteristics in correctly versus incorrectly classified cases were compared using independent sample t-test and chi-squared.

RESULTS

Test sensitivity and specificity of the classifier at the optimal Youden's index were 85.0% (95% CI: 62.1-96.8%) and 85.0% (95% CI: 62.1-96.8%), respectively. AUROC was 0.943 (95% CI: 0.820-0.991). Dice segmentation accuracy was 0.814 (95% CI: 0.805-0.826). There was no significant difference in AUROC between 1.5 T and 3.0 T studies. Sub-analysis showed superior sensitivity on full-thickness (100%) versus partial-thickness (72.5%) subgroups.

DATA CONCLUSION

Deep learning is a feasible approach to detect supraspinatus tears on MRI.

Marmoset brain segmentation from deconvolved magnetic resonance images and estimated label maps

PURPOSE

The proposed method aims to create label maps that can be used for the segmentation of animal brain MR images without the need of a brain template. This is achieved by performing a joint deconvolution and segmentation of the brain MR images.

METHODS

It is based on modeling locally the image statistics using a generalized Gaussian distribution (GGD) and couples the deconvolved image and its corresponding labels map using the GGD-Potts model. Because of the complexity of the resulting Bayesian estimators of the unknown model parameters, a Gibbs sampler is used to generate samples following the desired posterior probability.

RESULTS

The performance of the proposed algorithm is assessed on simulated and real MR images by the segmentation of enhanced marmoset brain images into its main compartments using the corresponding label maps created. Quantitative assessment showed that this method presents results that are comparable to those obtained with the classical method-registering the volumes to a brain template.

CONCLUSION

The proposed method of using labels as prior information for brain segmentation provides a similar or a slightly better performance compared with the classical reference method based on a dedicated template.

Localization of the Epileptogenic Zone by Multimodal Neuroimaging and High-Frequency Oscillation

Accurate localization of the epileptogenic zone (EZ) is a key factor to obtain good surgical outcome for refractory epilepsy patients. However, no technique, so far, can precisely locate the EZ, and there are barely any reports on the combined application of multiple technologies to improve the localization accuracy of the EZ. In this study, we aimed to explore the use of a multimodal method combining PET-MRI, fluid and white matter suppression (FLAWS)—a novel MRI sequence, and high-frequency oscillation (HFO) automated analysis to delineate EZ. We retrospectively collected 15 patients with refractory epilepsy who underwent surgery and used the above three methods to detect abnormal brain areas of all patients. We compared the PET-MRI, FLAWS, and HFO results with traditional methods to evaluate their diagnostic value. The sensitivities, specificities of locating the EZ, and marking extent removed versus not removed [RatioChann(ev)] of each method were compared with surgical outcome. We also tested the possibility of using different combinations to locate the EZ. The marked areas in every patient established using each method were also compared to determine the correlations among the three methods. The results showed that PET-MRI, FLAWS, and HFOs can provide more information about potential epileptic areas than traditional methods. When detecting the EZs, the sensitivities of PET-MRI, FLAWS, and HFOs were 68.75, 53.85, and 87.50%, and the specificities were 80.00, 33.33, and 100.00%. The RatioChann(ev) of HFO-marked contacts was significantly higher in patients with good outcome than those with poor outcome (p< 0.05). When intracranial electrodes covered all the abnormal areas indicated by neuroimaging with the overlapping EZs being completely removed referred to HFO analysis, patients could reach seizure-free (p < 0.01). The periphery of the lesion marked by neuroimaging may be epileptic, but not every lesion contributes to seizures. Therefore, approaches in multimodality can detect EZ more accurately, and HFO analysis may help in defining real epileptic areas that may be missed in the neuroimaging results. The implantation of intracranial electrodes guided by non-invasive PET-MRI and FLAWS findings as well as HFO analysis would be an optimized multimodal approach for locating EZ.

Novel stochastic framework for automatic segmentation of human thigh MRI volumes and its applications in spinal cord injured individuals

Severe spinal cord injury (SCI) leads to skeletal muscle atrophy and adipose tissue infiltration in the skeletal muscle, which can result in compromised muscle mechanical output and lead to health-related complications. In this study, we developed a novel automatic 3-D approach for volumetric segmentation and quantitative assessment of thigh Magnetic Resonance Imaging (MRI) volumes in individuals with chronic SCI as well as non-disabled individuals. In this framework, subcutaneous adipose tissue, inter-muscular adipose tissue and total muscle tissue are segmented using linear combination of discrete Gaussians algorithm. Also, three thigh muscle groups were segmented utilizing the proposed 3-D Joint Markov Gibbs Random Field model that integrates first order appearance model, spatial information, and shape model to localize the muscle groups. The accuracy of the automatic segmentation method was tested both on SCI (N = 16) and on non-disabled (N = 14) individuals, showing an overall 0.93±0.06 accuracy for adipose tissue and muscle compartments segmentation based on Dice Similarity Coefficient. The proposed framework for muscle compartment segmentation showed an overall higher accuracy compared to ANTs and STAPLE, two previously validated atlas-based segmentation methods. Also, the framework proposed in this study showed similar Dice accuracy and better Hausdorff distance measure to that obtained using DeepMedic Convolutional Neural Network structure, a well-known deep learning network for 3-D medical image segmentation. The automatic segmentation method proposed in this study can provide fast and accurate quantification of adipose and muscle tissues, which have important health and functional implications in the SCI population.

Are Magnetic Resonance Imaging Technologies Crucial to Our Understanding of Spinal Conditions?

Persistent spinal (traumatic and nontraumatic) pain is common and contributes to high societal and personal costs globally. There is an acknowledged urgency for new and interdisciplinary approaches to the condition, and soft tissues, including skeletal muscles, the spinal cord, and the brain, are rightly receiving increased attention as important biological contributors. In reaction to the recent suspicion and questioned value of imaging-based findings, this paper serves to recognize the promise that the technological evolution of imaging techniques, and particularly magnetic resonance imaging, is allowing in characterizing previously less visible morphology. We emphasize the value of quantification and data analysis of several contributors in the biopsychosocial model for understanding spinal pain. Further, we highlight emerging evidence regarding the pathobiology of changes to muscle composition (eg, atrophy, fatty infiltration), as well as advancements in neuroimaging and musculoskeletal imaging techniques (eg, fat-water imaging, functional magnetic resonance imaging, diffusion imaging, magnetization transfer imaging) for these important soft tissues. These noninvasive and objective data sources may complement known prognostic factors of poor recovery, patient self-report, diagnostic tests, and the "-omics" fields. When combined, advanced "big-data" analyses may assist in identifying associations previously not considered. Our clinical commentary is supported by empirical findings that may orient future efforts toward collaborative conversation, hypothesis generation, interdisciplinary research, and translation across a number of health fields. Our emphasis is that magnetic resonance imaging technologies and research are crucial to the advancement of our understanding of the complexities of spinal conditions. J Orthop Sports Phys Ther 2019;49(5):320-329. Epub 26 Mar 2019. doi:10.2519/jospt.2019.8793.

Automatic classification of tissues on pelvic MRI based on relaxation times and support vector machine

Tissue segmentation and classification in MRI is a challenging task due to a lack of signal intensity standardization. MRI signal is dependent on the acquisition protocol, the coil profile, the scanner type, etc. While we can compute quantitative physical tissue properties independent of the hardware and the sequence parameters, it is still difficult to leverage these physical properties to segment and classify pelvic tissues. The proposed method integrates quantitative MRI values (T1 and T2 relaxation times and pure synthetic weighted images) and machine learning (Support Vector Machine (SVM)) to segment and classify tissues in the pelvic region, i.e.: fat, muscle, prostate, bone marrow, bladder, and air. Twenty-two men with a mean age of 30±14 years were included in this prospective study. The images were acquired with a 3 Tesla MRI scanner. An inversion recovery-prepared turbo spin echo sequence was used to obtain T1-weighted images at different inversion times with a TR of 14000 ms. A 32-echo spin echo sequence was used to obtain the T2-weighted images at different echo times with a TR of 5000 ms. T1 and T2 relaxation times, synthetic T1- and T2-weighted images and anatomical probabilistic maps were calculated and used as input features of a SVM for segmenting and classifying tissues within the pelvic region. The mean SVM classification accuracy across subjects was calculated for the different tissues: prostate (94.2%), fat (96.9%), muscle (95.8%), bone marrow (91%) and bladder (82.1%) indicating an excellent classification performance. However, the segmentation and classification for air (within the rectum) may not always be successful (mean SVM accuracy 47.5%) due to the lack of air data in the training and testing sets. Our findings suggest that SVM can reliably segment and classify tissues in the pelvic region.

Semi-automated segmentation of magnetic resonance images for thigh skeletal muscle and fat using threshold technique after spinal cord injury

Magnetic resonance imaging is considered the “gold standard” technique for quantifying thigh muscle and fat cross-sectional area. We have developed a semi-automated technique to segment seven thigh compartments in persons with spinal cord injury. Thigh magnetic resonance images from 18 men (18–50 years old) with traumatic motor-complete spinal cord injury were analyzed in a blinded fashion using the threshold technique. The cross-sectional area values acquired by thresholding were compared to the manual tracing technique. The percentage errors for thigh circumference were (threshold: 170.71 ± 38.67; manual: 169.45 ± 38.27 cm²) 0.74%, subcutaneous adipose tissue (threshold: 65.99±30.79; manual: 62.68 ± 30.22) 5.2%, whole muscle (threshold: 98.18 ± 20.19; manual: 98.20 ± 20.08 cm2) 0.13%, femoral bone (threshold: 6.53 ± 1.09; manual: 6.53 ± 1.09 cm²) 0.64%, bone marrow fat (threshold: 3.12 ± 1.12; manual: 3.1 ± 1.11 cm²) 0.36%, knee extensor (threshold: 43.98 ± 7.66; manual: 44.61 ± 7.81 cm²) 1.78% and % intramuscular fat (threshold: 10.45 ± 4.29; manual: 10.92 ± 8.35%) 0.47%. Collectively, these results suggest that the threshold technique provided a robust accuracy in measuring the seven main thigh compartments, while greatly reducing the analysis time.

Evaluation of an automated thresholding algorithm for the quantification of paraspinal muscle composition from MRI images

BACKGROUND

The imaging assessment of paraspinal muscle morphology and fatty infiltration has gained considerable attention in the past decades, with reports suggesting an association between muscle degenerative changes and low back pain (LBP). To date, qualitative and quantitative approaches have been used to assess paraspinal muscle composition. Though highly reliable, manual thresholding techniques are time consuming and not always feasible in a clinical setting. The tedious and rater-dependent nature of such manual thresholding techniques provides the impetus for the development of automated or semi-automated segmentation methods. The purpose of the present study was to develop and evaluate an automated thresholding algorithm for the assessment of paraspinal muscle composition. The reliability and validity of the muscle measurements using the new automated thresholding algorithm were investigated through repeated measurements and comparison with measurements from an established, highly reliable manual thresholding technique.

METHODS

Magnetic resonance images of 30 patients with LBP were randomly selected cohort of patients participating in a project on commonly diagnosed lumbar pathologies in patients attending spine surgeon clinics. A series of T2-weighted MR images were used to train the algorithm; preprocessing techniques including adaptive histogram equalization method image adjustment scheme were used to enhance the quality and contrast of the images. All muscle measurements were repeated twice using a manual thresholding technique and the novel automated thresholding algorithm, from axial T2-weigthed images, at least 5 days apart. The rater was blinded to all earlier measurements. Inter-method agreement and intra-rater reliability for each measurement method were assessed. The study did not received external funding and the authors have no disclosures.

RESULTS

There was excellent agreement between the two methods with inter-method reliability coefficients (intraclass correlation coefficients) varying from 0.79 to 0.99. Bland and Altman plots further confirmed the agreement between the two methods. Intra-rater reliability and standard error of measurements were comparable between methods, with reliability coefficient varying between 0.95 and 0.99 for the manual thresholding and 0.97-0.99 for the automated algorithm.

CONCLUSION

The proposed automated thresholding algorithm to assess paraspinal muscle size and composition measurements was highly reliable, with excellent agreement with the reference manual thresholding method.