Generalizability of Deep Learning Classification of Spinal Osteoporotic Compression Fractures on Radiographs Using an Adaptation of the Modified-2 Algorithm-Based Qualitative Criteria

RATIONALE AND OBJECTIVES

Spinal osteoporotic compression fractures (OCFs) can be an early biomarker for osteoporosis but are often subtle, incidental, and underreported. To ensure early diagnosis and treatment of osteoporosis, we aimed to build a deep learning vertebral body classifier for OCFs as a critical component of our future automated opportunistic screening tool.

MATERIALS AND METHODS

We retrospectively assembled a local dataset, including 1790 subjects and 15,050 vertebral bodies (thoracic and lumbar). Each vertebral body was annotated using an adaption of the modified-2 algorithm-based qualitative criteria. The Osteoporotic Fractures in Men (MrOS) Study dataset provided thoracic and lumbar spine radiographs of 5994 men from six clinical centers. Using both datasets, five deep learning algorithms were trained to classify each individual vertebral body of the spine radiographs. Classification performance was compared for these models using multiple metrics, including the area under the receiver operating characteristic curve (AUC-ROC), sensitivity, specificity, and positive predictive value (PPV).

RESULTS

Our best model, built with ensemble averaging, achieved an AUC-ROC of 0.948 and 0.936 on the local dataset's test set and the MrOS dataset's test set, respectively. After setting the cutoff threshold to prioritize PPV, this model achieved a sensitivity of 54.5% and 47.8%, a specificity of 99.7% and 99.6%, and a PPV of 89.8% and 94.8%.

CONCLUSION

Our model achieved an AUC-ROC>0.90 on both datasets. This testing shows some generalizability to real-world clinical datasets and a suitable performance for a future opportunistic osteoporosis screening tool.

Integrated multimodal cell atlas of Alzheimer's disease

Alzheimer's disease (AD) is the most common cause of dementia in older adults. Neuropathological and imaging studies have demonstrated a progressive and stereotyped accumulation of protein aggregates, but the underlying molecular and cellular mechanisms driving AD progression and vulnerable cell populations affected by disease remain coarsely understood. The current study harnesses single cell and spatial genomics tools and knowledge from the BRAIN Initiative Cell Census Network to understand the impact of disease progression on middle temporal gyrus cell types. We used image-based quantitative neuropathology to place 84 donors spanning the spectrum of AD pathology along a continuous disease pseudoprogression score and multiomic technologies to profile single nuclei from each donor, mapping their transcriptomes, epigenomes, and spatial coordinates to a common cell type reference with unprecedented resolution. Temporal analysis of cell-type proportions indicated an early reduction of Somatostatin-expressing neuronal subtypes and a late decrease of supragranular intratelencephalic-projecting excitatory and Parvalbumin-expressing neurons, with increases in disease-associated microglial and astrocytic states. We found complex gene expression differences, ranging from global to cell type-specific effects. These effects showed different temporal patterns indicating diverse cellular perturbations as a function of disease progression. A subset of donors showed a particularly severe cellular and molecular phenotype, which correlated with steeper cognitive decline. We have created a freely available public resource to explore these data and to accelerate progress in AD research at SEA-AD.org.

Deep Learning Classification of Spinal Osteoporotic Compression Fractures on Radiographs using an Adaptation of the Genant Semiquantitative Criteria

RATIONALE AND OBJECTIVES

Osteoporosis affects 9% of individuals over 50 in the United States and 200 million women globally. Spinal osteoporotic compression fractures (OCFs), an osteoporosis biomarker, are often incidental and under-reported. Accurate automated opportunistic OCF screening can increase the diagnosis rate and ensure adequate treatment. We aimed to develop a deep learning classifier for OCFs, a critical component of our future automated opportunistic screening tool.

MATERIALS AND METHODS

The dataset from the Osteoporotic Fractures in Men Study comprised 4461 subjects and 15,524 spine radiographs. This dataset was split by subject: 76.5% training, 8.5% validation, and 15% testing. From the radiographs, 100,409 vertebral bodies were extracted, each assigned one of two labels adapted from the Genant semiquantitative system: moderate to severe fracture vs. normal/trace/mild fracture. GoogLeNet, a deep learning model, was trained to classify the vertebral bodies. The classification threshold on the predicted probability of OCF outputted by GoogLeNet was set to prioritize the positive predictive value (PPV) while balancing it with the sensitivity. Vertebral bodies with the top 0.75% predicted probabilities were classified as moderate to severe fracture.

RESULTS

Our model yielded a sensitivity of 59.8%, a PPV of 91.2%, and an F1 score of 0.72. The areas under the receiver operating characteristic curve (AUC-ROC) and the precision-recall curve were 0.99 and 0.82, respectively.

CONCLUSION

Our model classified vertebral bodies with an AUC-ROC of 0.99, providing a critical component for our future automated opportunistic screening tool. This could lead to earlier detection and treatment of OCFs.

DicomAnnotator: a Configurable Open-Source Software Program for Efficient DICOM Image Annotation

Modern, supervised machine learning approaches to medical image classification, image segmentation, and object detection usually require many annotated images. As manual annotation is usually labor-intensive and time-consuming, a well-designed software program can aid and expedite the annotation process. Ideally, this program should be configurable for various annotation tasks, enable efficient placement of several types of annotations on an image or a region of an image, attribute annotations to individual annotators, and be able to display Digital Imaging and Communications in Medicine (DICOM)-formatted images. No current open-source software program fulfills these requirements. To fill this gap, we developed DicomAnnotator, a configurable open-source software program for DICOM image annotation. This program fulfills the above requirements and provides user-friendly features to aid the annotation process. In this paper, we present the design and implementation of DicomAnnotator. Using spine image annotation as a test case, our evaluation showed that annotators with various backgrounds can use DicomAnnotator to annotate DICOM images efficiently. DicomAnnotator is freely available at https://github.com/UW-CLEAR-Center/DICOM-Annotator under the GPLv3 license.

Radiogenomic modeling predicts survival-associated prognostic groups in glioblastoma

Background

Combined whole-exome sequencing (WES) and somatic copy number alteration (SCNA) information can separate isocitrate dehydrogenase (IDH)1/2-wildtype glioblastoma into two prognostic molecular subtypes, which cannot be distinguished by epigenetic or clinical features. The potential for radiographic features to discriminate between these molecular subtypes has yet to be established.

Methods

Radiologic features (n = 35 340) were extracted from 46 multisequence, pre-operative magnetic resonance imaging (MRI) scans of IDH1/2-wildtype glioblastoma patients from The Cancer Imaging Archive (TCIA), all of whom have corresponding WES/SCNA data. We developed a novel feature selection method that leverages the structure of extracted MRI features to mitigate the dimensionality challenge posed by the disparity between a large number of features and the limited patients in our cohort. Six traditional machine learning classifiers were trained to distinguish molecular subtypes using our feature selection method, which was compared to least absolute shrinkage and selection operator (LASSO) feature selection, recursive feature elimination, and variance thresholding.

Results

We were able to classify glioblastomas into two prognostic subgroups with a cross-validated area under the curve score of 0.80 (±0.03) using ridge logistic regression on the 15-dimensional principle component analysis (PCA) embedding of the features selected by our novel feature selection method. An interrogation of the selected features suggested that features describing contours in the T2 signal abnormality region on the T2-weighted fluid-attenuated inversion recovery (FLAIR) MRI sequence may best distinguish these two groups from one another.

Conclusions

We successfully trained a machine learning model that allows for relevant targeted feature extraction from standard MRI to accurately predict molecularly-defined risk-stratifying IDH1/2-wildtype glioblastoma patient groups.

NIMG-46. RADIOGENOMIC FEATURES PREDICT CLINICALLY RELEVANT GENOME-WIDE ALTERATION SIGNATURES IN GLIOBLASTOMA

BACKGROUND

Previously, we have shown that combined whole-exome sequencing (WES) and genome-wide somatic copy number alteration (SCNA) information can separate IDH1/2-wildtype glioblastoma into two prognostic molecular subtypes (Group 1 and Group 2) and that these subtypes cannot be distinguished by epigenetic or clinical features. However, the potential for radiographic features to discriminate between these molecular subtypes has not been established.

METHODS

Radiogenomic features (n=35,400) were extracted from 46 multiparametric, pre-operative magnetic resonance imaging (MRI) of IDH1/2-wildtype glioblastoma patients from The Cancer Imaging Archive, all of whom have corresponding WES and SCNA data in The Cancer Genome Atlas. We developed a novel feature selection method that leverages the structure of extracted radiogenomic MRI features to mitigate the dimensionality challenge posed by the disparity between the number of features and patients in our cohort. Seven traditional machine learning classifiers were trained to distinguish Group 1 versus Group 2 using our feature selection method. Our feature selection was compared to lasso feature selection, recursive feature elimination, and variance thresholding.

RESULTS

We are able to classify Group 1 versus Group 2 glioblastomas with a cross-validated area under the curve (AUC) score of 0.82 using ridge logistic regression and our proposed feature selection method, which reduces the size of our feature set from 35,400 to 288. An interrogation of the selected features suggests that features describing contours in the T2 abnormality region on the FLAIR MRI modality may best distinguish these two groups from one another.

CONCLUSIONS

We successfully trained a machine learning model that allows for relevant targeted feature extraction from standard MRI to accurately predict molecularly-defined risk-stratifying IDH1/2-wildtype glioblastoma patient groups. This algorithm may be applied to future prospective studies to assess the utility of MRI as a surrogate for costly prognostic genomic studies.

Lumbar Spinal Stenosis Severity by CT or MRI Does Not Predict Response to Epidural Corticosteroid versus Lidocaine Injections

BACKGROUND AND PURPOSE

Epidural steroid injections may offer little-to-no short-term benefit in the overall population of patients with symptomatic spinal stenosis compared with lidocaine alone. We investigated whether imaging could identify subgroups of patients who might benefit most.

MATERIALS AND METHODS

A secondary analysis of the Lumbar Epidural Steroid Injections for Spinal Stenosis prospective, double-blind trial was performed, and patients were randomized to receive an epidural injection of lidocaine with or without corticosteroids. Patients (n = 350) were evaluated for qualitative and quantitative MR imaging or CT measures of lumbar spinal stenosis. The primary clinical end points were the Roland-Morris Disability Questionnaire and the leg pain numeric rating scale at 3 weeks following injection. ANCOVA was used to assess the significance of interaction terms between imaging measures of spinal stenosis and injectate type on clinical improvement.

RESULTS

There was no difference in the improvement of disability or leg pain scores at 3 weeks between patients injected with epidural lidocaine alone compared with corticosteroid and lidocaine when accounting for the primary imaging measures of qualitative spinal stenosis assessment (interaction coefficients for disability score, -0.1; 95% CI, -1.3 to 1.2; P = .90; and for the leg pain score, 0.1; 95% CI, -0.6 to 0.8; P = .81) or the quantitative minimum thecal sac cross-sectional area (interaction coefficients for disability score, 0.01; 95% CI, -0.01 to 0.03; P = .40; and for the leg pain score, 0.01; 95% CI, -0.01 to 0.03; P = .33).

CONCLUSIONS

Imaging measures of spinal stenosis are not associated with differential clinical responses following epidural corticosteroid injection.

Protein Interaction Networks

The study of protein interactions is playing an ever increasing role in our attempts to understand cells and diseases on a system-wide level. This article reviews several experimental approaches that are currently being used to measure protein-protein, protein-DNA and gene-gene interactions. These techniques have now been scaled up to produce extensive genome-wide data sets that are providing us with a first glimpse of global interaction networks. Complementing these experimental approaches, several computational methodologies to predict protein interactions are also reviewed. Existing databases that serve as repositories for protein interaction information and how such databases are used to analyze high-throughput data from a pathway perspective is also addressed. Finally, current efforts to combine multiple data types to obtain more accurate and comprehensive models of protein interactions are discussed. It is clear that the evolution of these experimental and computational approaches is rapidly changing our view of biology, and promises to provide us with an unprecedented ability to model cells and organisms at a system-wide level.

Multi-rigid image segmentation and registration for the analysis of joint motion from three-dimensional magnetic resonance imaging

We report an image segmentation and registration method for studying joint morphology and kinematics from in vivo magnetic resonance imaging (MRI) scans and its application to the analysis of foot and ankle joint motion. Using an MRI-compatible positioning device, a foot was scanned in a single neutral and seven other positions ranging from maximum plantar flexion, inversion, and internal rotation to maximum dorsiflexion, eversion, and external rotation. A segmentation method combining graph cuts and level set was developed. In the subsequent registration step, a separate rigid body transformation for each bone was obtained by registering the neutral position dataset to each of the other ones, which produced an accurate description of the motion between them. The segmentation algorithm allowed a user to interactively delineate 14 foot bones in the neutral position volume in less than 30 min total (user and computer processing unit [CPU]) time. Registration to the seven other positions took approximately 10 additional minutes of user time and 5.25 h of CPU time. For validation, our results were compared with those obtained from 3DViewnix, a semiautomatic segmentation program. We achieved excellent agreement, with volume overlap ratios greater than 88% for all bones excluding the intermediate cuneiform and the lesser metatarsals. For the registration of the neutral scan to the seven other positions, the average overlap ratio is 94.25%, while the minimum overlap ratio is 89.49% for the tibia between the neutral position and position 1, which might be due to different fields of view (FOV). To process a single foot in eight positions, our tool requires only minimal user interaction time (less than 30 min total), a level of improvement that has the potential to make joint motion analysis from MRI practical in research and clinical applications.

Digital genome-wide ncRNA expression, including SnoRNAs, across 11 human tissues using polyA-neutral amplification

Non-coding RNAs (ncRNAs) are an essential class of molecular species that have been difficult to monitor on high throughput platforms due to frequent lack of polyadenylation. Using a polyadenylation-neutral amplification protocol and next-generation sequencing, we explore ncRNA expression in eleven human tissues. ncRNAs 7SL, U2, 7SK, and HBII-52 are expressed at levels far exceeding mRNAs. C/D and H/ACA box snoRNAs are associated with rRNA methylation and pseudouridylation, respectively: spleen expresses both, hypothalamus expresses mainly C/D box snoRNAs, and testes show enriched expression of both H/ACA box snoRNAs and RNA telomerase TERC. Within the snoRNA 14q cluster, 14q(I-6) is expressed at much higher levels than other cluster members. More reads align to mitochondrial than nuclear tRNAs. Many lincRNAs are actively transcribed, particularly those overlapping known ncRNAs. Within the Prader-Willi syndrome loci, the snoRNA HBII-85 (group I) cluster is highly expressed in hypothalamus, greater than in other tissues and greater than group II or III. Additionally, within the disease locus we find novel transcription across a 400,000 nt span in ovaries. This genome-wide polyA-neutral expression compendium demonstrates the richness of ncRNA expression, their high expression patterns, their function-specific expression patterns, and is publicly available.

More lateral and anterior prefrontal coil location is associated with better repetitive transcranial magnetic stimulation antidepressant response

BACKGROUND

The left dorsolateral prefrontal cortex (DLPFC) is the most commonly used target for transcranial magnetic stimulation (TMS) in the treatment of depression. The "5-cm rule" is an empiric method used for probabilistic targeting of the DLPFC in most clinical trials. This rule may be suboptimal, as it does not account for differences in skull size or variations in prefrontal anatomy relative to motor cortex location. This study is a post hoc analysis of data from a large repetitive TMS (rTMS) trial in which we examined the variability of coil placement and how it affects antidepressant efficacy.

METHODS

Fifty-four depressed subjects enrolled in a randomized, single-site trial received either active rTMS or sham for 3 weeks. Prior to treatment initiation, investigators placed vitamin E capsules at the point of stimulation and used a high-resolution magnetic resonance imaging (MRI) scan to image these fiducials relative to anatomy. We employed a semiautomated imaging-processing algorithm to localize the cortical region stimulated.

RESULTS

Active TMS significantly reduced Hamilton Depression Rating Scale (HDRS) scores. A linear model for this improvement involving the coordinates of the stimulated cortex location, age, and treatment condition was highly significant. Specifically, individuals with more anterior and lateral stimulation sites were more likely to respond.

CONCLUSIONS

These results suggest that within the general anatomical area targeted by the 5-cm rule, placing the TMS coil more laterally and anteriorly is associated with improved response rates in TMS depression studies. Controlled studies testing this anatomical hypothesis are needed.

Thrombosis and hemorrhage in the acute period following Gamma Knife surgery for arteriovenous malformation

Bleeding of an arteriovenous malformation (AVM) following stereotactic radiosurgery (SRS) is a known risk during the latency interval, but hemorrhage in the 30-day period following radiosurgery rarely has been reported in the literature. The authors present the case of a 57-year-old man who underwent Gamma Knife surgery for a large AVM, and they provide radiographic documentation of a thrombus in the primary draining vein immediately preceding an AVM hemorrhage within 9 days after radiosurgery. They postulate that the pathophysiology of an AVM hemorrhage in the acute period following SRS is related to an association among tissue irradiation, acute inflammatory response, and vessel thrombosis.

The authors also review the literature on risk factors for hemorrhage due to untreated and radiosurgically treated AVMs. Recent evidence on the role of inflammation in the pathogenesis of AVMs and the pathophysiology of AVM rupture is presented. Inflammatory markers have been demonstrated in brain AVM tissue, and the association between inflammation and AVM hemorrhage has been established. There is an acute inflammatory response following tissue irradiation, resulting in structural and functional vascular changes that can lead to vessel thrombosis. Early hemorrhage following radiosurgical treatment of AVMs may be related to the acute inflammatory response and associated vascular changes that occur in irradiated tissue. In the first stage of a planned 2-stage Gamma Knife treatment for a large AVM in the featured case, the superior posteromedial portion of the primary draining vein was included in the treatment field. The authors present the planning images and subsequent CT scans demonstrating a new venous thrombus in the primary draining vein. An acute inflammatory response following radiosurgery with resultant acute venous thrombus formation and venous obstruction is proposed as one mechanism of an AVM hemorrhage in this patient. Radiographic evidence of the time course of thrombosis and hemorrhage supports the hypothesis that acute venous obstruction is a cause of intracranial hemorrhage.

Neuroinformatics for genome-wide 3D gene expression mapping in the mouse brain

Large scale gene expression studies in the mammalian brain offer the promise of understanding the topology, networks and ultimately the function of its complex anatomy, opening previously unexplored avenues in neuroscience. High-throughput methods permit genome-wide searches to discover genes that are uniquely expressed in brain circuits and regions that control behavior. Previous gene expression mapping studies in model organisms have employed situ hybridization (ISH), a technique that uses labeled nucleic acid probes to bind to specific mRNA transcripts in tissue sections. A key requirement for this effort is the development of fast and robust algorithms for anatomically mapping and quantifying gene expression for ISH. We describe a neuroinformatics pipeline for automatically mapping expression profiles of ISH data and its use to produce the first genomic scale 3-D mapping of gene expression in a mammalian brain. The pipeline is fully automated and adaptable to other organisms and tissues. Our automated study of over 20,000 genes indicates that at least 78.8 percent are expressed at some level in the adult C56BL/6J mouse brain. In addition to providing a platform for genomic scale search, high-resolution images and visualization tools for expression analysis are available at the Allen Brain Atlas web site (http://www.brain-map.org).

Impact of basilar artery vasospasm on outcome in patients with severe cerebral vasospasm after aneurysmal subarachnoid hemorrhage

BACKGROUND AND PURPOSE

The purpose of the present study was to evaluate the impact of basilar artery (BA) vasospasm on outcome in patients with severe vasospasm after aneurysmal subarachnoid hemorrhage (aSAH).

METHODS

Sixty-five patients with clinically suspect severe cerebral vasospasm after aSAH underwent cerebral angiography before endovascular treatment. Vasospasm severity was assessed for each patient by transcranial Doppler measurements, angiography, and (99m)Tc-ethylcysteinate dimer single-photon emission computed tomography (ECD-SPECT) imaging. Percentage of BA narrowing was calculated in reference to the baseline angiogram.

RESULTS

BA narrowing >or=25% was found in 23 of 65 patients, and delayed brain stem (BS) hypoperfusion, as estimated by ECD-SPECT, was found in 16. Fourteen of 23 patients with BA narrowing >or=25% experienced BS hypoperfusion, whereas only 2 of 42 patients with >or=25% BA narrowing experienced BS ischemia (P<0.001). Stepwise logistic regression after adjusting for age with Hunt and Hess grade, Fisher grade, hydrocephalus, and aneurysmal location as covariables revealed BA narrowing >or=25% and delayed BS hypoperfusion to be significantly and independently associated with unfavorable 3-month outcome (P=0.0001; odds ratio, 10.1; 95% CI, 2.5 to 40.8; and P=0.007; odds ratio, 13.8, 95% CI, 2.18 to 91.9, respectively).

CONCLUSIONS

These findings suggest for the first time that BA vasospasm after aSAH is an independent and significant prognostic factor associated with poor outcome in patients with severe cerebral vasospasm requiring endovascular therapy. Further study should be done to evaluate the role of interventional therapy on outcome in patients with posterior circulation vasospasm.

Modelling the network of cell cycle transcription factors in the yeast Saccharomyces cerevisiae

BACKGROUND

Reverse-engineering regulatory networks is one of the central challenges for computational biology. Many techniques have been developed to accomplish this by utilizing transcription factor binding data in conjunction with expression data. Of these approaches, several have focused on the reconstruction of the cell cycle regulatory network of Saccharomyces cerevisiae. The emphasis of these studies has been to model the relationships between transcription factors and their target genes. In contrast, here we focus on reverse-engineering the network of relationships among transcription factors that regulate the cell cycle in S. cerevisiae.

RESULTS

We have developed a technique to reverse-engineer networks of the time-dependent activities of transcription factors that regulate the cell cycle in S. cerevisiae. The model utilizes linear regression to first estimate the activities of transcription factors from expression time series and genome-wide transcription factor binding data. We then use least squares to construct a model of the time evolution of the activities. We validate our approach in two ways: by demonstrating that it accurately models expression data and by demonstrating that our reconstructed model is similar to previously-published models of transcriptional regulation of the cell cycle.

CONCLUSION

Our regression-based approach allows us to build a general model of transcriptional regulation of the yeast cell cycle that includes additional factors and couplings not reported in previously-published models. Our model could serve as a starting point for targeted experiments that test the predicted interactions. In the future, we plan to apply our technique to reverse-engineer other systems where both genome-wide time series expression data and transcription factor binding data are available.

FDG-PET/CT-guided intensity modulated head and neck radiotherapy: a pilot investigation

BACKGROUND

2-deoxy-2[(18)F]fluoro-D-glucose-positron emission tomography (FDG-PET) imaging can be registered with CT images and can potentially improve neck staging sensitivity and specificity in patients with head and neck squamous cell cancer. The intent of this study was to examine the use of registered FDG-PET/CT imaging to guide head and neck intensity modulated radiotherapy (IMRT) planning.

METHODS

Twenty patients with squamous cell carcinoma of the oral cavity, oropharynx, larynx, or hypopharynx underwent FDG-PET and contrast-enhanced CT imaging of the head and neck before neck dissection surgery. Combined FDG-PET/CT images were created by use of a nonrigid image registration algorithm. All IMRT plans were theoretical and were not used for treatment. We prescribed 66 Gy in 30 fractions to FDG-avid CT abnormalities and nodal zones directly involved with disease, without prophylactic coverage of uninvolved neck levels. Matched CT-guided IMRT plans designed according to the specifications of Radiation Therapy Oncology Group (RTOG) H-0022 were available for comparison. We investigated the feasibility of FDG-PET/CT-directed IMRT dose escalation in five patients with FDG-avid disease located away from critical normal structures. After 66 Gy, FDG-avid disease with 0.5-cm margins was boosted in 220 cGy increments until dose-limiting criteria were reached.

RESULTS

Elimination of prophylactic coverage to FDG-PET/CT-negative neck levels markedly reduced mean dose (Dmean) to the contralateral parotid gland (p < .001) and Dmean to the laryngeal cartilage (p = .001). No FDG-PET/CT-directed plan missed pathologically verified nodal disease. During the dose escalation exercise, we successfully increased the dose to 95% of the planning target volume (PTV95%) to a mean of 7490 cGy (range, 7153-8098 cGy).

CONCLUSIONS

We demonstrate early proof of the principle that FDG-PET/CT-guided IMRT planning can selectively target and intensify treatment of head and neck disease while reducing critical normal tissue doses. Routine clinical use of such planning should not be engaged until the accuracy of FDG-PET/CT is fully validated. Future directions, including refinement of treatment to gross disease and radiologically uninvolved neck nodal levels, are discussed.

Characterization of genetically defined types of Charcot-Marie-Tooth neuropathies by using magnetic resonance neurography

Object. Charcot-Marie-Tooth (CMT) disease is a collection of related genetic disorders affecting peripheral nerves with an incidence of one in every 2500 individuals. A diagnosis of CMT disease has classically relied on a medical history, examination, and measurement of nerve conduction velocities. Advancements in genetic testing and magnetic resonance (MR) imaging techniques may provide clinicians with a more precise diagnostic armamentarium. The authors investigated MR neurography as a possible method to characterize CMT subtypes.

Methods. The authors performed MR neurography to evaluate sciatic nerves in the mid-thigh area of seven patients with genetically defined subtypes of CMT, one patient with chronic inflammatory demylinating polyneuropathy, and one patient without neuropathy. The authors correlate their findings with normal nerve conduction velocities (NCVs) and present their results as a descriptive case series.

Although MR neurography could not be used to distinguish subtypes of CMT disease on nerve area or fascicle number, it appears to characterize phenotypic features and disease progression noninvasively in patients with some subtypes.

Conclusions. In conjunction with NCV measurements, MR neurography may be useful in the diagnosis of CMT neuropathies and in monitoring disease progression.

FDG-PET/CT imaging for preradiotherapy staging of head-and-neck squamous cell carcinoma

PURPOSE

Image localization of head-and-neck squamous cell carcinoma lags behind current techniques to deliver a precise radiation dose with intensity-modulated radiotherapy. This pilot study prospectively examined the use of registered 18-F-fluorodeoxyglucose (FDG)-positron emission tomography (PET)/CT for preradiotherapy staging of the neck.

METHODS AND MATERIALS

Sixty-three patients with squamous cell carcinoma of the oral cavity, oropharynx, larynx, or hypopharynx were enrolled into an institutional FDG-PET imaging protocol between September 2000 and June 2003. Of these patients, 20 went on to immediate neck dissection surgery and were studied further. Of these 20, 17 (85%) had American Joint Committee on Cancer Stage III or IV disease. All patients underwent preoperative FDG-PET and contrast-enhanced CT of the head and neck. FDG-PET/CT images were created using a nonrigid image registration algorithm developed at the University of Washington. Alternate primary and nodal gross tumor volumes were contoured with radiotherapy treatment planning software, blinded to each other and to the pathology results. One set of volumes was designed with CT guidance alone and the other with the corresponding FDG-PET/CT images. Neck dissection specimens were subdivided into surgical nodal levels intraoperatively, and the histopathologic findings were correlated with the CT and FDG-PET/CT nodal level findings.

RESULTS

FDG-PET/CT detected 17 of 17 heminecks and 26 of 27 nodal zones histologically positive by dissection (100% and 96% sensitivity, respectively). The nodal level staging sensitivity and specificity for FDG-PET/CT was 96% (26 of 27) and 98.5% (68 of 69), respectively. FDG-PET/CT correctly detected nodal disease in 2 patients considered to have node-negative disease by CT alone. Agreement between the imaging results and pathology findings was stronger for FDG-PET/CT (kappa 0.95, 95% confidence interval 0.82-0.99) than for CT alone (kappa 0.81, 95% confidence interval 0.63-0.91; p = 0.06 by two-sided McNemar's testing).

CONCLUSION

These early findings suggest that FDG-PET/CT is superior to CT alone for geographic localization of diseased neck node levels. Confirmatory trials to substantiate the accuracy of FDG-PET/CT neck staging should be prioritized.

Application of a neural network to improve nodal staging accuracy with 18F-FDG PET in non-small cell lung cancer

We proposed to train a back-propagation artificial neural network (aNN) on a cohort of surgically proven non-small cell lung cancers (NSCLCs) and compare its accuracy with that of a trained (18)F-FDG PET reader. We plan to show that an aNN trained on (18)F-FDG PET- and CT-derived data is more accurate in predicting the true surgicopathologic nodal stage than a human reader.

METHODS

One hundred thirty-three NSCLC patients with surgically proven N status treated at the University of Washington Medical Center or the Veterans Affairs Puget Sound Health Care System between February 1998 and September 2002 were used as inputs for the creation of an aNN. From CT of the thorax and (18)F-FDG PET (neck to pelvis) performed before surgery, we extracted the primary tumor size and uptake (maximum pixel SUV [maxSUV]), normal lung and mediastinal uptake, and nodal uptake (maxSUV). Using the same 133 cases, the same output (surgical N status, N(0) to N(3)), and the same software configuration settings, scenarios were created to assess which input parameters were most influential in creating an optimal aNN. To compute this optimal aNN, cases were split randomly 100 times into a training subset of 103 cases and a testing subset of 30 cases having the same proportion of N(0), N(1), N(2), and N(3) cases. N status predicted by the aNN was compared with the proven surgical N status to calculate the aNN accuracy. The N status readings from (18)F-FDG PET were also compared with the surgical N status for the same cases to determine (18)F-FDG PET accuracy.

RESULTS

Statistical tests demonstrate that the best aNN accuracy is achieved by using N(1)-N(2)- N(3) nodal maxSUV divided by background uptake, the primary tumor size, and primary tumor maxSUV as inputs. The aNN correctly predicted the N stage in 87.3% of the testing cases compared with 73.5% for the (18)F-FDG PET expert reader. Accuracy of the aNN increased to 94.8% (PET, 89.4%) when comparing N(0) + N(1) with N(2) or N(3) status and to 94.9% (PET, 91.9%) when comparing N(0) + N(1) with N(2) + N(3) status.

CONCLUSION

A back-propagation aNN can be trained to predict hilar and mediastinal nodal involvement with greater accuracy than an expert (18)F-FDG PET reader. Such a tool could be used to improve clinical interpretations and for clinical training.

Intra-arterial urokinase for acute ischemic stroke: factors associated with complications

The authors abstracted the records of 43 patients treated with intra-arterial urokinase for acute ischemic stroke to identify predictors of serious complications. Sixteen (37%) had such a complication. Higher urokinase dose (>1.5 x 10(6) U), higher mean arterial blood pressure before treatment (>130 mm Hg), basilar occlusive strokes, and severe strokes were most predictive of these complications. Although urokinase is no longer manufactured, these findings identify patients at risk for complications from other intra-arterial thrombolytics.

Ghost imaging in MRI

Needle biopsies and other interventions done under MR Fluoroscopy sometimes do not show the target well, either because the rapid sequence does not have adequate contrast or because a contrast agent may have washed out of the target. In these cases, an image that shows the target can be saved and scaled to match the spatial parameters of the fluoroscopic sequence, and used as a virtual or ghost field upon which the fluoroscopic images are superimposed, thus providing a view of the target, useful for needle pre-localization and for monitoring its progress as it is inserted.

Creating fast finite element models from medical images

The procedure for creating a patient-specific virtual tissue model with finite element (FE) based haptic (force) feedback varies substantially from that which is required for generating a typical volumetric model. In addition to extracting geometrical and texture map data to provide visual realism, it is necessary to obtain information for supporting a FE model. Among many differences, FE-based VR environments require a FE model with appropriate material properties assigned. The FE equation must also be processed in a manner specific to the surgical task in order to maximize deformation and haptic computation speed. We are currently developing methodologies and support software for creating patient-specific models from medical images. The steps for creating such a model are as follows: 1) obtain medical images and texture maps of tissue structures; 2) extract tissue structure contours; 3) generate a 3D mesh from the tissue structure contours; 4) alter mesh based on simulation objectives; 5) assign material properties, boundary nodes and texture maps; 6) generate a fast (or real-time) FE model; and 7) support the tissue models with task-specific tools and training aids. This paper will elaborate on the above steps with particular reference to the creation of suturing simulation software, which will also be described.

Reducing negativity artifacts in emission tomography: post-processing filtered backprojection solutions

The problem of negative artifacts in emission tomography reconstructions computed by filtered backprojection (FBP) is of practical concern particularly in low count studies. Statistical reconstruction methods based on maximum likelihood (ML) are automatically constrained to be non-negative but their excessive computational overhead (orders of magnitude greater than FBP) has limited their use in operational settings. Motivated by the statistical character of the negativity artifact, the authors develop a simple post-processing technique that iteratively adjusts negative values by cancellation with positive values in a surrounding local neighborhood. The compute time of this approach is roughly equivalent to 2 applications of FBP. The approach was evaluated by numerical simulation in 1- and 2-dimensional (2D) settings. In 2D, the source distributions included the Hoffman, the Shepp and Vardi, and a digitized version of the Jaszczak cold spheres phantoms. The authors' studies compared smoothed versions of FBP, the post-processed FBP, and ML implemented by the expectation-maximization algorithm. The root mean square (RMS) error between the true and estimated source distribution was used to evaluate performance; in 2D, additional region-of-interest-based measures of reconstruction accuracy were also employed. In making comparisons between the different methods, the amount of smoothing applied to each reconstruction method was adapted to minimize the RMS error-this was found to be critical.

Brain tissue pH and ventilatory acclimatization to high altitude

31P nuclear magnetic resonance spectroscopy (31P-NMRS) was performed on brain cross sections of four human subjects before and after 7 days in a hypobaric chamber at 447 Torr to test the hypothesis that brain intracellular acidosis develops during acclimatization to high altitude and accounts for the progressively increasing ventilation that develops (ventilatory acclimatization). Arterial blood gas measurements confirmed increased ventilation. At the end of 1 wk of hypobaria, brain intracellular pH was 7.023 +/- 0.046 (SD), unchanged from preexposure pH of 6.998 +/- 0.029. After return to sea level, however, it decreased to 6.918 +/- 0.032 at 15 min (P less than 0.01) and 6.920 +/- 0.046 at 12 h (P less than 0.01). The ventilatory response to hypoxia increased [from 0.35 +/- 0.11 (l/min)/(-%O2 saturation) before exposure to 0.69 +/- 0.19 after, P = 0.06]. Brain intracellular acidosis is probably not a supplemental stimulus to ventilatory acclimatization to high altitude. However, brain intracellular acidosis develops on return to normoxia from chronic hypoxia, suggesting that brain pH may follow changes in blood and cerebrospinal fluid pH as they are altered by changes in ventilation.