Recently, the 2022 Medical Artificial Intelligence Conference (CMAI 2022) and the 2nd “China Medical Academic Journal Development” high-end forum were held.
This summit forum invited several directors of radiology departments of top hospitals and authoritative experts of artificial intelligence technology to discuss the clinical application and scientific research progress of artificial intelligence technology in medical imaging, and share their research experience.
Leifeng.com’s “Medical Health AI Nuggets”, as the support unit of this conference, participated in the speech content and in-depth reports of the guests throughout the whole process.
As a guest of this conference, Professor Lu Jie, Vice President of Xuanwu Hospital of Capital Medical University, delivered a speech on the topic of “Application Research of Artificial Intelligence in MRI Imaging of Brain Demyelinating Diseases”.
The following is the content of Professor Lu Jie’s speech. Leifeng.com’s “Medical Health AI Nuggets” has been sorted out without changing the original intention.
Hello everyone, the topic of today’s report is the application of artificial intelligence in magnetic resonance imaging of brain demyelinating disease.
Brain demyelinating disease is a neuroimmune disease in which demyelination of nerves is the main or primary lesion, followed by damage to axons, cell bodies and glia, mainly including multiple sclerosis (MS) and neuromyelitis optica (NMO). ).
There are about 2.8 million patients with multiple sclerosis in the world, and there are about 42,000 cases in China.
Its pathological characteristic changes are multiple demyelinating plaques in white matter with reactive gliosis and axonal damage. In the early stage of the disease, the main manifestations are the disintegration and demyelination and the activation of microglia, and the axons are relatively intact. Late lesions manifest as axonal disintegration, neuronal loss, and formation of glial sclerotic plaques.
This is the 2017 version of the multiple sclerosis diagnostic criteria, and we can see that we need both objective clinical evidence and some additional evidence in the diagnosis. Among them, magnetic resonance imaging is the most important assessment method to assist in the diagnosis of multiple sclerosis.
Conventional magnetic resonance imaging used in clinical practice can provide multi-dimensional information for multiple sclerosis, including the number, location, and volume of lesions, as well as the characteristics of enhancement and the progression of lesions. Moreover, the sensitivity of conventional MRI for the diagnosis of multiple sclerosis is high, reaching about 95%.
Common locations of multiple sclerosis lesions are periventricular, subcortical, U-fiber, brainstem, and cerebellum, and are oval or finger-shaped. In the acute stage, the lesions have a sense of swelling, showing “fried egg sign”, “open ring”, “C-type” enhancement, and the duration of enhancement is less than 90 days. In the chronic phase, T1 showed hypointense lesions with a “black hole” shape.
This is the typical location of multiple sclerosis lesions in the brain, paraventricular, infratentorial, and corpus callosum.
Typical manifestations are U-shaped fibrous lesions, Dawson’s hand sign, and oval lesion morphology.
The characteristics of lesions in different periods are also different, showing heterogeneity in time. If the patient is followed up clinically, it may be found that the patient has new lesions, and there are often more than one lesions, and there are both enhancing and non-enhancing lesions, that is to say, there are both acute phase lesions and chronic phase lesions.
Neuromyelitis optica is also a common inflammatory demyelinating disease in the central nervous system. Its incidence is second only to multiple sclerosis. It is more common in Asia, especially in women, with a male-to-female ratio of about 1:4. The disease mainly involves the optic nerve, spinal cord, and brain, and serum-specific autoantibodies are positive for AQP4 antibody. 85%-90% of patients will relapse, and the prognosis is worse than multiple sclerosis.
The diagnostic criteria for neuromyelitis optica have undergone several evolutions over the years. The concept of NMO was first proposed in 1894; the first diagnostic criteria for NMO appeared in 1999; after the discovery of AQP4-IgG in 2004, NMO was considered to be an independent disease different from MS In 2007, a new concept of NMOSD (neuromyelitis optica spectrum disorder) was proposed; in 2015, the latest consensus on the diagnosis of NMOSD appeared.
Neuromyelitis optica spectrum disorders can be diagnosed based on AQP4 antibody negative and AQP4 antibody positive, and magnetic resonance assessment is an important part of the diagnosis.
Intracerebral lesions of NMOSD are not uncommon, and 43%-70% of patients can develop intracerebral lesions at the first onset, mainly including periependymal lesions surrounding the ventricular system, periventricular and corpus callosum lesions, corticospinal tract lesions, and hemisphere lesions. White matter lesions and nonspecific lesions.
This is the MRI change of the diencephalon and cerebral damage in NMOSD, usually thalamus and hypothalamus, extensive subcortical white matter damage and enhancement; E is line-like damage along the long axis of the corpus callosum, F is along the cerebral peduncle, pontine corticospinal Longitudinal damage of bundles, and changes in acute periependymal cerebral white matter damage and enhancement.
In addition, we can also see the MRI manifestations of NMOSD dorsal medulla oblongata, extreme posterior region and brainstem lesions, showing lesions in the dorsal medulla oblongata, lesions in the extreme posterior region, lesions around the fourth ventricle and ventral pons, and dorsal midbrain. damage as well as damage around the fourth ventricle.
The characteristic lesions of NMOSD are periependymal lesions and corticospinal tract lesions.
This is the periventricular and corpus callosum lesions of NMO. The periventricular lesions are close to the lateral ventricle wall, distributed along the peritunical lining, commonly found around the anterior and posterior pedicles of the lateral ventricle, and rarely involving the lateral ventricle body. When the full thickness of the corpus callosum is involved, the lesion presents an “arch-bridge-like” change, whereas brain lesions at the corpus callosum-septal junction (CSI) are more common in MS.
Hemispheric white matter lesions can manifest as tumor-like demyelinating lesions, multiple sclerosis-like lesions, and acute disseminated encephalomyelitis-like lesions.
Differential diagnosis of MS and NMOSD is very challenging. According to previous reports, about 30% of MS patients will be misdiagnosed as NMOSD in the early stage of the disease. The clinical symptoms of the two diseases are relatively similar, and the laboratory test results also partially overlap, so the diagnosis cycle is long, and about 12% of patients need at least 6 years to be diagnosed. For doctors in primary hospitals and junior doctors, the diagnosis of these two diseases is more challenging.
Can artificial intelligence help us do some work? Artificial intelligence is the use of computer technology to simulate the thinking and learning process of human beings, making it competent for complex tasks that can only be done by human intelligence. Its development started from the interaction of computers and human intelligence in 1950, to the algorithmic improvements of machine learning in 1980, which enabled us to do big data processing, to the intelligent neural network in deep learning in 2010 to decode deeper image information .
Radiomics is a precise image analysis technology generated in the context of big data and artificial intelligence, which can mine high-dimensional quantitative features that cannot be recognized by the naked eye. Using advanced mathematical model algorithms to transform into high-dimensional data with high-resolution repeatability, low-redundancy and mining, quantitative analysis of features, so as to deeply explore the potential information contained in the image.
This is the specific workflow of radiomics, from image acquisition and segmentation, to the extraction of depth features, semantic features, morphological features and texture features, to feature screening and model building. There are many options for feature screening and modeling.
Let’s take a look at the nomogram of omics, which is based on multi-factor regression analysis, integrates multiple predictors, and then uses scaled line segments to draw on the same plane according to a certain proportion, so as to be used for Express the interrelationships between variables in a predictive model.
In the artificial intelligence application study of MRI imaging of brain demyelinating disease, we collected 150 patients, of which 73 had multiple sclerosis and 77 had neuromyelitis optica. We used 68 patients as the training set and 62 as the validation set, manually delineated the lesions, and extracted 273 quantitative omics features in the lesions on T2WI images based on the nomogram.
Finally, 11 omics features and 4 clinical features were screened out. It can be seen that the AUC of the training set and the AUC of the test set are relatively close.
Another project studied 189 patients, of which 95 were multiple sclerosis and 94 were neuromyelitis optica, of which 135 patients were used as a training set and 54 were used as a validation set. nomogram, and 485 quantitative omics features were extracted from lesions displayed on T2WI images of the cervical cord.
Finally, 9 omics features and 5 clinical features were screened for inclusion in the model, and their identification efficiency was high.
AI-assisted differential diagnosis is also challenging, mainly because based on a single clinical database, a single sequence, and a single magnetic resonance apparatus, the nomogram of the omics is still lacking in interpretability, and there is a lack of comparison between artificial intelligence models and clinicians’ interpretations .
Xuanwu Hospital collected 116 patients with cerebral demyelinating disease, of which 78 were MS and 38 were NMO, including two datasets, both 1.5T and 3T. We collected imaging data of T1-MPRAGE and T2WI sequences, as well as clinical information such as the patient’s disability score and disease course.
We extracted 1118 quantitative omics features from T2 lesions, constructed a random forest classification model based on multi-parameter image representation, and selected 9 omics features and 4 clinical features into the model.
And compared the correct rate of neuroimaging doctor’s diagnosis and radiomic model diagnosis. The accuracy of the doctor’s identification results was 0.709, the sensitivity was 0.615, and the specificity was 0.750. The accuracy, sensitivity, and specificity of the machine learning model identification results have been significantly improved, which are better than the results of the doctor’s naked eye identification.
Machine learning models are composed of high-dimensional omics features, so their interpretability is relatively poor and difficult to understand. SHAP is a popular model interpretation method, which determines the importance of each individual by calculating the contribution of each individual in the cooperation. This model is interpretable. For case A, the random forest model judges that it is 89% likely to be MS, and the most important contribution comes from the feature of H-MPR-Log_95. We can calculate the contribution value of each feature.
Topology is a method of abstracting entities into “points” that have nothing to do with their size and shape, and abstracting the lines connecting entities into “lines”, and then expressing the relationship between these points and lines in the form of a graph. The purpose is to study these points. , the connection between the lines. A topology diagram is a diagram that represents the relationship between points and lines. Topology can be applied to molecular structures, geographic maps, DNA structures, and knots.
We can also apply topology in the diagnosis of brain demyelinating diseases. Due to the different pathogenesis of MS and NMO, the spatial distribution, shape and size of brain lesions are different, so the topological properties of brain lesions in these two diseases are not the same. We hope to find the potential Differential diagnosis breakthrough point.
We analyzed T2WI data from 97 demyelinating cases, 66 with MS and 31 with NMO. We have not yet published the results of this part of the data. Specifically, the software was used to delineate the lesions, and the spatial patterns of MS multiple lesions were extracted. The whole DTA framework is composed of three modules, dynamic hierarchical network construction, dynamic topology quantification and topology pattern analysis.
We see the results of a topological study of the differential diagnosis that the lesions in the NMO patients are more tightly connected and the lesions are larger compared with the lesions in the MS patients. Topological studies have many advantages, visualization of models and results, visualization of network connectivity of lesions, visualization of lesion volumes.
Compared with the previous differential diagnosis models, the topological model has a higher AUC of 0.875, and the accuracy and specificity are also high.
In previous literature reports, the accuracy rate of MS MRI prediction studies was about 58%-70%, the clinical manifestations and imaging features usually do not match, the MRI lesions lack specificity, and the pathological characteristics of lesions on T2WI and T1WI. Also lacks specificity.
So can we help predict prognosis through topology-based artificial intelligence models.
Longitudinal follow-up studies using magnetic resonance imaging found that chronic lesions tended to consolidate, and the degree of clinical disability gradually increased. Therefore, we hypothesize that as the disease progresses, the topological properties of the lesions change; conversely, the changes in the topological structure of the lesions also have a potential predictive role.
This part of the study collected a total of 90 cases of progressive MS and 54 cases of non-progressive MS. If the disability scale increased by more than 1.5 points, it was defined as progression, and less than 1.5 points was defined as non-progression. T2WI data were used for training. Spatial patterns of multiple lesions were extracted by delineating the lesions.
The results showed that patients with advanced MS had denser lesion connections and lesions with larger lesion volume than non-progressed patients. At the same time, visualization of models and results, 3D visualization of lesions, visualization of lesion network connections, and visualization of lesion volumes are also realized.
Compared with previous research reports, the magnetic resonance prediction model also showed a higher AUC, reaching 0.752, with higher accuracy and specificity.
in conclusion:
Multiple sclerosis and neuromyelitis optica are common demyelinating diseases of the brain, but the differential diagnosis of the two is difficult; artificial intelligence can help to mine high-dimensional quantitative features that cannot be identified by the naked eye in imaging images; topology-based artificial intelligence models are used in prediction. Brain demyelinating disease has important value in the prognosis, and we hope to conduct more in-depth exploration in this area in the future.
This article is reprinted from: https://www.leiphone.com/category/healthai/yg2l8S5lCt8aQrtD.html
This site is for inclusion only, and the copyright belongs to the original author.