Kermany, D. S., Goldbaum, M., Cai, W., Valentim, C. C. S., Liang, H., Baxter, S. L., … Zhang, K. (2018). Identifying Medical Diagnoses and Treatable Diseases by Image-Based Deep Learning. Cell, 172(5), 1122–1124.e9. https://doi.org/10.1016/j.cell.2018.02.010
Ting, D. S. W., Liu, Y., Burlina, P., Xu, X., Bressler, N. M., & Wong, T. Y. (2018). AI for medical imaging goes deep. Nature Medicine, 24(5), 539–540. https://doi.org/10.1038/s41591-018-0029-3
Deep learning is a form of AI which is designed to mimic the layers of neurons in the human brain to process and extract information, allowing computers to learn without being explicitly programmed. This technique has been used to detect diseases, including retinal diseases, TB from chest radiographs and malignant melanoma from skin images. In the February issue in Cell, Kermany et al. present an AI approach for detecting several medical conditions, including diabetic macular edema (DME), choroidal neovascularization (CNV), drusen and pediatric pneumonia, from patient images, with promising diagnostic performance.
Transfer learning is a method for building an AI system using convolutional networks that have already been pretrained using a large data set in the public domain. Transfer learning allows the knowledge gained during the training process to recognize animals in images to be used in recognizing retinal diseases from optical coherence tomography (OCT) images. A ConvNet consists of multiple layers of “neurons” with trainable weights that are thus able to learn features and patterns. In medical imaging, many publicly available ConvNet models (VGGNet, ResNet, Inception V3 and DenseNet) have been used.
In this study, the authors trained the deep learning framework on 37,000 images of CNV, ∼ 11,000 images of DME, ∼ 9,000 images of drusen and ∼ 51,000 images from unaffected individuals using Inception V3; images were obtained from 4,686 individuals in total. This was followed by validation on 1,000 images, consisting of 250 images of CNV, 250 images of DME, 250 images of drusen and 250 normal images. The authors evaluated the AI performance in three models—multiclass comparison, limited model and binary classifiers. The diagnostic performance of deep learning for all three models was > 90% accuracy in differentiating CNV, DME, drusen and normal images, with the best outcome achieved in the binary classifiers model (an accuracy of > 98%).
Although it had a slight drop in accuracy, the limited model was able to achieve an accuracy of > 90% even though the training set was 100 times smaller than the full data set. In comparison to six human experts, the deep learning system showed similar outcomes in identifying individuals requiring urgent referral as determined on the basis of their OCT images.
As further validation of the authors learning approach for medical diagnoses was conducted on a set of children CXR consisting of 5,232 training images from 5,826 patients (2,538 bacterial pneumonia, 1,345 viral pneumonia and 1,349 healthy). They achieved an accuracy of 92.8%.
The authors note that future studies might address challenges in medical imaging, such as how and/or when machines and human adjudication differ, and design methods that quantitatively and qualitatively assess and explain sources of error for both humans and machines. The authors also providing their data and codes in a publicly available database, and they hope that other biomedical researchers use their work as a resource to improve the performance of future models and help drive the field forward.
Chen, S.-P., Chou, K.-H., Fuh, J.-L., Huang, Y.-H., Huang, C.-C., Lirng, J.-F., … Wang, S.-J. (2018). Dynamic Changes in White Matter Hyperintensities in Reversible Cerebral Vasoconstriction Syndrome. JAMA Neurology, [Epub ahead of print]. https://doi.org/10.1001/jamaneurol.2018.1321
Reversible cerebral vasoconstriction syndrome (RCVS) is a neurovascular disorder characterized by abrupt, severe headaches and reversible segmental vasoconstriction of cerebral arteries. Headaches in RCVS usually repetitively occur in 2-week to 3-week periods after disease onset, whereas the vasoconstrictions might take up to 3 months to resolve. Up to one-third of patients with RCVS experience complications, such as posterior reversible encephalopathy syndrome (PRES), ischemic stroke, intracranial (ie, cortical, intracerebral, or subdural) hemorrhage, or cervical artery dissection. Although WMHs on neuroimaging are more commonly seen than ischemic or hemorrhagic complications in patients with RCVS, there have not yet been any studies that have focused on investigating their nature.
The authors prospectively recruited patients with RCVS over a 3-year period from January 2010 through December 2012. Patients received serial isotropic 3-dimension FLAIR imaging (1-mm slice thickness) at 3T as well as transcranial and extracranial color-coded sonography on registration and during follow-ups. Sixty-five patients with RCVS completed the study and underwent a total of 162 MRI examinations. Of the 65 included patients, 58 (89%) were women, and the mean (SD) age was 50.1 years. The total mean WMH load peaked at 3.2cm3 in the third week post onset and fell to 0.8cm3 in the fourth week. White matter hyperintensities were predominantly frontal and periventricular.
White matter hyperintensity lesions in RCVS have a disease specific dynamic spatiotemporal evolution that parallels severity of vasoconstrictions. The pathogenesis of WMHs in patients with RCVS may be associated with regional hypoperfusion, impaired dampening capacity of cerebral arteries or arterioles to central pulsatile flow, and microvascular damage. White matter hyperintensities in patients with RCVS are partially reversible.
de Wilde, A., van der Flier, W. M., Pelkmans, W., Bouwman, F., Verwer, J., Groot, C., … Scheltens, P. (2018). Association of Amyloid Positron Emission Tomography With Changes in Diagnosis and Patient Treatment in an Unselected Memory Clinic Cohort. JAMA Neurology. https://doi.org/10.1001/jamaneurol.2018.1346
To evaluate the association of amyloid PET with changes in diagnosis, diagnostic confidence, treatment, and patients’ experiences in an unselected memory clinic cohort (i.e., real life).
Amyloid PET using fluoride-18 florbetaben was offered to 866 patients who visited the tertiary memory clinic at the VU University Medical Center between January 2015 and December 2016 as part of their routine diagnostic dementia workup. 476 (55%) were included, 32 (4%) were excluded, and 358 (41%) did not participate. To enrich this sample, 31 patients with mild cognitive impairment from the University Medical Center Utrecht memory clinic were included. For each patient, neurologists determined a preamyloid and postamyloid PET diagnosis that existed of both a clinical syndrome (dementia, mild cognitive impairment, or subjective cognitive decline) and a suspected etiology (Alzheimer disease [AD] or non-AD), with a confidence level ranging from 0% to 100%. In addition, the neurologist determined patient treatment in terms of ancillary investigations, medication, and care. Each patient received a clinical follow-up 1 year after being scanned.
Amyloid PET results were positive for 242 patients (48%). The suspected etiology changed for 125 patients (25%) after undergoing amyloid PET, more often due to a negative (82 of 265 [31%]) than a positive (43 of 242 [18%]) PET result. Post-PET changes in suspected etiology occurred more frequently in patients older (>65 years) than younger (<65 years) than the typical age at onset of 65 years. Mean diagnostic confidence increased from 80 to 89. In 123 patients (24%), there was a change in patient treatment post-PET, mostly related to additional investigations and therapy. In this prospective diagnostic study, the authors attempted to bridge the gap between validating amyloid PET in a research setting and implementing this diagnostic tool in daily clinical practice. The authors offered amyloid PET to all patients who visited a memory clinic and observed that both amyloid-positive and amyloid-negative results had an important association with changes in diagnosis and treatment in patients with and without dementia.
Steffensen, A. B., Oernbo, E. K., Stoica, A., Gerkau, N. J., Barbuskaite, D., Tritsaris, K., … MacAulay, N. (2018). Cotransporter-mediated water transport underlying cerebrospinal fluid formation. Nature Communications, 9(1), 2167. https://doi.org/10.1038/s41467-018-04677-9
The mammalian brain is bathed in CSF, which is continuously produced at a rate of approximately 500 ml fluid per day in the adult human. Prior to exiting the brain, the CSF travels through the ventricular system and part of it re-enters the brain via the para-vascular route along the large arteries and penetrating arterioles. CSF is predominantly produced by the choroid plexus, an epithelial monolayer resting on highly vascularized connective tissue and located at the base of each of the four ventricles. The molecular mechanisms underlying this choroidal fluid production remain unresolved.
The CSF production is generally assumed to take place by transport of osmotically active ions (e.g. sodium by the Na+/K+ -ATPase) followed by osmotically obliged, passive movement of water, partly via the water channel aquaporin 1 (AQP1) expressed at the luminal membrane of the choroid plexus. However, several observations suggest that such a simple osmotic model is not be adequate: (1) The CSF production declined by a mere 20% in the AQP1 knock-out mice. (2) With the known osmotic water permeability across the choroid plexus, detailed calculations have demonstrated that the osmolarity of the CSF must exceed that of the plasma by as much as 250 mOsm in order for the CSF to be produced at the observed rate by simple osmosis (in contrast to the measured difference in osmolarity of 5−10 mOsm) (3) The choroid plexus has the ability to produce CSF against an oppositely directed osmotic gradient.
A number of cotransporter proteins have the inherent ability to cotransport water along with the ions/solutes in the translocation mechanism. The coupling between water translocation and substrate transport takes place within the protein itself in a manner that permits water to be transported independently of, and even against, an osmotic gradient. Examples of such water-translocating cotransporters are the Na+/K+/2Cl− cotransporter 1 (NKCC1) and the K+/Cl− cotransporters (KCCs). Isoforms of these transport proteins have been detected in the choroid plexus epithelium, although their exact isoform distribution, relative expression, and membrane targeting remain largely unknown, as are their ability to transport water independently of an osmotic gradient in the choroid plexus tissue and their contribution to CSF production in vivo. In the present study, the authors introduce the water-translocating cotransporter, NKCC1, as the main contributor to CSF formation in the mouse choroid plexus.
Through a series of experiments, of which only the subheading will be detailed here, the authors make their case:
NKCC1 and KCC1 are expressed in choroid plexus.
NKCC1 is localized to the luminal membrane in choroid
plexus.
NKCC1 is poised for outwardly directed transport.
NKCC1 significantly contributes to CSF production in vivo.
Live imaging shows NKCC1-mediated CSF production.
The authors conclude that have demonstrated, by complementary ex vivo and in vivo experimentation, that the high production rate of CSF is sustained by NKCC1 via its inherent ability to cotransport water along with its directional ion translocation in a manner independent of osmotic driving forces. This unconventional means of fluid secretion underlying CSF production represents a paradigm shift in the field and provides a long-needed rational therapeutic target towards brain pathologies involving disturbances in brain water homeostasis and increased intracranial pressure.
7 Figures….good luck with those.
Buchbinder, R., Johnston, R. V, Rischin, K. J., Homik, J., Jones, C. A., Golmohammadi, K., & Kallmes, D. F. (2018). Percutaneous vertebroplasty for osteoporotic vertebral compression fracture. The Cochrane Database of Systematic Reviews, 4(1), CD006349. https://doi.org/10.1002/14651858.CD006349.pub3
190 pages, but they do provide an executive type summary.
Twenty-one trials were included: five compared vertebroplasty with placebo (541 randomized participants), eight with usual care (1136 randomized participants), seven with kyphoplasty (968 randomized participants) and one compared vertebroplasty with facet joint steroid injection (217 randomized participants). Trial size varied from 46 to 404 participants, most participants were female, mean age ranged between 62.6 and 81 years, and mean symptom duration varied from a week to more than six months. Three placebo-controlled trials were at low risk of bias and two were possibly susceptible to performance and detection bias. Other trials were at risk of bias for several criteria, most notably due to lack of participant and personnel blinding. Compared with placebo, high- to moderate-quality evidence from five trials (one with incomplete data reported) indicates that vertebroplasty provides no clinically important benefits with respect to pain, disability, disease-specific or overall quality of life or treatment success at one month. Evidence was not downgraded for potential publication bias as only one placebo-controlled trial remains unreported. Mean pain (on a scale zero to 10, higher scores indicate more pain) was five points with placebo and 0.6 points better (0.2 better to 1 better) with vertebroplasty, an absolute pain reduction of 6% (2% better to 10% better, minimal clinical important difference is 15%) and relative reduction of 9% (3% better to 14% better) (five trials, 535 participants). Mean disability measured by the Roland-Morris Disability Questionnaire (scale range zero to 23, higher scores indicate worse disability) was 14.2 points in the placebo group and 1.7 points better (0.3 better to 3.1 better) in the vertebroplasty group.
They conclude that based upon high- to moderate-quality evidence, their updated review does not support a role for vertebroplasty for treating acute or subacute osteoporotic vertebral fractures in routine practice. They found no demonstrable clinically important benefits compared with placebo (sham procedure) and subgroup analyses indicated that the results did not differ according to duration of pain ≤ 6 weeks versus > 6 weeks.
Inglese, M., & Petracca, M. (2018). MRI in multiple sclerosis. Current Opinion in Neurology, 17 Suppl 1, 1. https://doi.org/10.1097/WCO.0000000000000559
Review of current imaging standards, and areas of future research.
The brain standardized protocol includes the following sequences: axial and sagittal 2D T2- weighted (proton-density/T2-FLAIR/T2-weighted) for lesion identification, being the sagittal view of choice to facilitate lesion detection in the corpus callosum and juxtacortical area; precontrast T1- weighted that, if acquired as 3D inversion recovery T1-weighted scan, can be used for volumetric assessment; postcontrast T1-weighted for identification of contrast-enhancing lesions. Additional suggested sequences are diffusion weighted sequence for detection of non MS disease (i.e. acute ischemia/infarction); precontrast or postcontrast axial T1 spin-echo for chronic black holes identification; SWI sequence for identification of lesional central vein; 2D and/or 3D dual inversion recovery sequence for identification of cortical lesions. All sequences should be acquired with nongapped slice thickness 3mm or less and an in-plane spatial resolution of 1X1mm, with the exception of DWI, for which a slice thickness 5mm or less is suggested. Spinal cord standardized protocol includes the following sequences: two sagittal sequences with different contrasts [T2 and proton density and/or short-tau inversion recovery (STIR)] to increase confidence in lesion detection; postcontrast T1-weighted spin-echo if T2 lesions are present. Axial imaging using 2D or 3D T2-weighted spin echo sequences are suggested if sagittal images are suboptimal or inconclusive.
Other topics discussed include MT, leptomeningeal inflammation, and spectroscopy.
2 Figures.
Suzuki, Y., Nakamura, Y., Yamada, K., Kurabe, S., Okamoto, K., Aoki, H., … Nakada, T. (2018). Aquaporin Positron Emission Tomography Differentiates Between Grade III and IV Human Astrocytoma. Neurosurgery, 82(6), 842–846. https://doi.org/10.1093/neuros/nyx314
Studies have reported a relationship between astrocytoma malignancy and the abnormal expression of AQPs in the CNS, particularly AQP1 and AQP4. These water channel proteins are highly expressed in malignant astrocytomas, and a positive correlation is observed between their expression levels and histological tumor grade. Both AQPs have significant roles in mesenchymal microvascular proliferation and infiltrative growth. Malignant astrocytoma is characterized by microvascular proliferation and diffuse, infiltrative growth, tumor characteristics that are believed to promote tumor progression and portend a poor prognosis. The evaluation of both the AQPs’ expression in malignant astrocytoma patients in vivo would lead to an early diagnosis of malignancy and an accurate evaluation of the extent of malignant tumor regions.
The authors examined the utility of aquaporin PET for differentiating between astrocytoma grade III and grade IV using the AQP radioligand [11C]TGN-020. Fifteen astrocytoma patients, grade III (n = 7) and grade IV (n = 8), and 10 healthy volunteers underwent [11C]TGN-020 aquaporin PET imaging. Surgical tissues of astrocytoma patients were examined for histopathological grading using the WHO classification standard and expression of AQP1 and AQP4 immunohistochemically. Mean standardized uptake values of astrocytoma grade III and IV were higher than normal white matter for both tumor grades. Importantly, mean standardized uptake values of astrocytoma grade IV were significantly higher than grade III.
The authors speculate that in addition to distinguishing between astrocytoma grades III and IV, the present PET study also suggests a therapeutic potential for AQP1 and AQP4 inhibitors in the treatment of astrocytoma grades III and IV. Given that proliferation, migration, and infiltration are major biological properties of malignant brain tumors and involve abnormal AQP distribution, AQP1 and AQP4 are considered potential drug targets for oncotherapy. Increased uptake of [11C]TGN-020 in astrocytoma grades III and IV relative to surrounding healthy tissue may also suggest that selective AQP inhibitors will be helpful in arresting tumor growth and infiltration without damaging healthy tissue.
3 Figures
Tumturk, A., Li, Y., Turan, Y., Cikla, U., Iskandar, B. J., & Baskaya, M. K. (2018). Emergency resection of brainstem cavernous malformations. Journal of Neurosurgery, 128(5), 1289–1296. https://doi.org/10.3171/2017.1.JNS161693
This is a series of 4 cases with emergency surgery of brainstem hemorrhages from CM. Brainstem and other deep-seated CMs may have higher rates of symptomatic hemorrhage and often present with higher rates of neurological compromise compared with their cerebral counterparts. In a meta-analysis, the overall estimated annual incidences of hemorrhage and recurrent hemorrhage were 0.3% and 6.3%, respectively, for non-brainstem CMs compared with 2.8% and 32.3% for brainstem CMs. Moreover, the outcomes of untreated brainstem CMs can be poor, as patients with recurrent neurological deficits are less likely to recover after multiple bouts of rehemorrhage.
The cases presented illustrate the potential value of emergency extirpative surgery in patients with symptomatic brainstem CMs presenting with rapidly progressive neurological deterioration.
The timing of surgery for brainstem CMs has been a controversial topic. On the one hand, surgery can result in symptom recovery and functional protection, while on the other, cerebral edema and acute hematoma can distort lesional and normal anatomical boundaries, making surgery more difficult. Many authors propose delaying surgery into the subacute phase when perilesional edema has subsided, hematoma has liquefied, and neurological function has stabilized, thereby facilitating surgical decision making. When surgery is delayed until the chronic stage, scar tissue and adhesions can become prominent, making surgical dissection more tedious.
8 Figures, 2 tables.