Magnetic Resonance Imaging Classification Research Articles

Tumor Detection and Classification of MRI Brain Image using Support Vector Machine (SVM)

Brain is one of the most complex organs in the human body that works with billions of cells. A cerebral tumor occurs when there is an uncontrolled division of cells that form an abnormal group of cells around or within the brain. This cell group can affect the normal functioning of brain activity and can destroy healthy cells. Brain tumors are classified as benign or low-grade and malignant tumors or high-grade. Benign tumors are non-cancerous tumor and they do not spread to other tissues or organs. Malignant tumors are cancerous tissue and they can easily spread to other tissues or organs. Proposed system is to differentiate between normal brain and tumor brain (benign or malign). Also, the proposed system predicts brain tumor from MRI image classification system is based on extracting useful MRI features for diagnosing the medical MRI images. The benefits of using SVM is nevertheless of the image brightness or rotation of the MRI image, it also provides huge number of strong features that can be automatically prepared well to be suitable for MRI classification. Support Vector Machine (SVM) algorithm is used to predict the diseases accurately from MRI (Magnetic Resonance Imaging) scan images. SVM algorithm is the used for the purpose of classifying the image datasets and to predict the disease by itself for those matching the images to enhance a comprehensive set of quantitative measurements among several influential on various brain image databases.

Study and Analysis of Various Automatic Brain Tumour Segmentation andClassification: A Challenging Overview

Early prediction of Brain Tumour (BT) plays a vital function for providing treatment plan and patient care at initial stage. The manual identification and classification of brain tumour is difficult to realize and time consuming due to the same structures, and also the scarcity of available radiology expertise. Generally, the BT classification is performed in two ways, such as:realize whether the image as either normal or abnormal (ii) Classification of abnormal region based on the various kinds of tumours. Thus, the manual classification of brain tumour from MR images are time consuming, nonreproducible and impractical due to the vast quantity of MRI dataset. To overcome these issues, automatic classification is a proper solution to classify the tumour from Magnetic Resonance (MR) images with less intervention of radiologists. The key challenge in classification of MR image is the semantic gap among low level visual data gathered through MRI machine and high level information perceived through human evaluator. The conventional machine learning-based classification techniques focused only in either low level or high level or handcrafted features for diminishing this gap and achieving better feature extraction and classification methods. Thus, this survey analyzes several techniques intended to BT segmentation and classification through images from various sources. This study utilized25 research papers concentrated on various techniques, and the review of various researches technique-wiseis also to be provided.Finally, the analysis discussed in this survey based on the publication year, research technique, implementation tools, performance measures and achievement of the research methodologies towards BT segmentation and classification using different datasets. At the end, the research gaps, issues of the techniques and then, the motivation for developing an effective method for brain tumour segmentation and classification is also to be revealed.

Cerebrospinal Fluid Drop Metastases of Canine Glioma: Magnetic Resonance Imaging Classification.

Dissemination of glioma in humans can occur as leptomeningeal nodules, diffuse leptomeningeal lesions, or ependymal lesions. Cerebrospinal fluid (CSF) drop metastasis of glioma is not well-recognized in dogs. Ten dogs with at least two anatomically distinct and histologically confirmed foci of glioma were included in this study. The 10 dogs underwent 28 magnetic resonance imaging (MRI) examinations, with distant CSF drop metastasis revealed in 13 MRIs. The CSF drop metastases appeared as leptomeningeal nodules in four dogs, diffuse leptomeningeal lesions in six dogs, and ependymal lesions in seven dogs; six dogs had a combination of lesion types. Primary tumors were generally T2-heterogeneous and contrast-enhancing. Many metastases were T2-homogeneous and non-enhancing. Diffuse leptomeningeal lesions were seen as widespread extra-axial contrast-enhancement, again very dissimilar to the intra-axial primary mass. Primary masses were rostrotentorial, whereas metastases generally occurred in the direction of CSF flow, in ventricles, CSF cisterns, and the central canal or leptomeninges of the cervical or thoracolumbar spinal cord. Seven of the dogs had received therapy limited to the primary mass, such as surgery or stereotactic radiation, then developed metastasis in the following months. CSF drop metastasis of glioma may take a very different appearance on MRI to the primary mass, including periventricular lesions that are more homogeneous and less contrast-enhancing, rostral horn signal changes, or leptomeningeal enhancement ventral to the brainstem or encircling the spinal cord.

Optimized convolutional neural network by firefly algorithm for magnetic resonance image classification of glioma brain tumor grade

The most frequent brain tumor types are gliomas. The magnetic resonance imaging technique helps to make the diagnosis of brain tumors. It is hard to get the diagnosis in the early stages of the glioma brain tumor, although the specialist has a lot of experience. Therefore, for the magnetic resonance imaging interpretation, a reliable and efficient system is required which helps the doctor to make the diagnosis in early stages. To make classification of the images, to which class the glioma belongs, convolutional neural networks, which proved that they can obtain an excellent performance in the image classification tasks, can be used. Convolutional network hyperparameters’ tuning is a very important issue in this domain for achieving high accuracy on the image classification; however, this task takes a lot of computational time. Approaching this issue, in this manuscript, we propose a metaheuristics method to automatically find the near-optimal values of convolutional neural network hyperparameters based on a modified firefly algorithm and develop a system for automatic image classification of glioma brain tumor grades from magnetic resonance imaging. First, we have tested the proposed modified algorithm on the set of standard unconstrained benchmark functions and the performance is compared to the original algorithm and other modified variants. Upon verifying the efficiency of the proposed approach in general, it is applied for hyperparameters’ optimization of the convolutional neural network. The IXI dataset and the cancer imaging archive with more collections of data are used for evaluation purposes, and additionally, the method is evaluated on the axial brain tumor images. The obtained experimental results and comparative analysis with other state-of-the-art algorithms tested under the same conditions show the robustness and efficiency of the proposed method.

Magnetic Resonance Imaging Classifications of Rotator Cuff Tear Are Associated with Different Shoulder Outcome Scores

Objective. Rotator cuff tear (RCT) accounts for 50% of shoulder injuries, leading to chronic pain and disability in the upper extremity. The study is conducted to investigate the association between resonance imaging (MRI) classifications of patients with RCT and different shoulder outcome scores. Methods. From September 2018 to October 2019, 112 patients underwent shoulder MRI at our institution and selected as eligible study subjects according to inclusion and exclusion criteria. Among these 112 patients, 69 cases had confirmed history of shoulder trauma and 43 cases were due to chronic shoulder joint pain. The shoulder function of patients was evaluated by University of California Los Angeles Shoulder (UCLA) score, Constant-Murley score, Shoulder Pain and Disability Index (SPADI), and simple shoulder test (SST). Results. Among the 112 patients, there were 34 cases, 58 cases, and 20 cases with MRI classifications at grades I, II, and III, respectively. There was no significant difference in the injured tendons in patients with different MRI classifications ( P > 0.05 ). The injury at the supraspinatus was more common. The scores of UCLA, Constant-Murley, and SST in patients with MRI grading at grade I were significantly higher than those at grades II and III ( P < 0.05 ), which were significantly higher in patients at grade II than those at grade III ( P < 0.05 ). SPADI score in patients with MRI grading at grade I was significantly lower than that at grades II and III ( P < 0.05 ), while there was no significant difference in SPADI score between patients at grades II and III ( P > 0.05 ). MRI classifications were negatively correlated with scores of UCLA, Constant-Murley, and SST ( P > 0.05 ). There was no significant correlation between MRI grade and SPADI scores ( P > 0.05 ). Conclusion. The supraspinatus tendon injury is more common in patients with RCT. MRI classifications were negatively correlated with scores of UCLA, Constant-Murley, and SST.

Correlation between magnetic resonance imaging classification of endolymphatic hydrops and clinical manifestations and audiovestibular test results in patients with definite Ménière's disease

ObjectivesThe purpose was to evaluate magnetic resonance imaging (MRI) classification of endolymphatic hydrops with clinical features, audiological and vestibular tests in patients with definite unilateral Ménière's disease (MD). MethodsThirty-eight patients were enrolled in this study. The severity of the main clinical symptoms, audiovestibular tests, and MRI, performed 4 hours after intravenous injection of gadobutrol to visualize inner ear compartments, were evaluated. Endolymphatic space dilatation was assessed using Barath and Bernaerts grading systems, and the correlation between the grade of the hydrops and clinical features was evaluated. ResultsUsing the Barath system, cochlear hydrops was visualized in 81.6% of affected ears, while vestibular was 63.2%. Sensitivity increased to 94.7% using Bernaerts' modification. Vestibular hydrops involving the utricle was present only among patients with cochlear and saccular endolymphatic space dilatation. There was a significant relationship between the hearing level and the vestibular hydrops degree in the Bernaerts scale. The grade of the hydrops correlated neither with the duration of MD nor with the severity of main clinical symptoms. Our study proved MRI to be a sensitive diagnostic tool in MD. The endolymphatic hydrops' grade correlates with the hearing level, which confirms endolymphatic space dilatation's role in hearing loss. ConclusionsIn our study, two similar MRI grading systems were used; however, several differences were found compared to one another. The Bernaerts scale was more sensitive than the Barath scale, and several relationships between the radiological and clinical data were found. Therefore, several MRI evaluating scales and correlating them with the clinical features are needed. The increased perilymphatic enhancement of the cochlea and an extra low-grade vestibular hydrops distinguished in the Bernaerts scale may increase MD diagnosis sensitivity. Magnetic resonance findings in MD support the clinical diagnosis and may help to understand MD pathophysiology better. This study adds to the knowledge and diagnostics in MD for healthcare to improve patients' treatment.

Diagnostic magnetic resonance imaging in synovial chondromatosis of the temporomandibular joint

The aim of this paper was to investigate the clinical and magnetic resonance imaging (MRI) features of synovial chondromatosis (SC) of the temporomandibular joint (TMJ). Fourteen patients with SC of the TMJ were included in the study. Clinical and MRI features were analysed and divided into three types based on MRI classification: type I with loose bodies, type II with homogeneous masses, and type III with a mixture of loose bodies and homogeneous masses. All SCs occurred in the superior compartment of the TMJ. There were two patients (14%) categorised as type I, five (36%) as type II and seven (50%) as type III. Four patients (29%) had disc perforation, and nine had bone erosion; among those nine, seven (78%) had type III and two (22%) type II. Histological examination showed inflammation and calcification in the synovial membrane and, and cartilage of the hyaline type in all cases. MRI has advantages in the diagnosis of SC.

The clinical impact of Prostate Imaging-Reporting and Data System classification in patients with haemospermia undergoing multiparametric magnetic resonance imaging of the prostate.

In this study, we evaluated the role of the Prostate Imaging-Reporting and Data System (PI-RADS) classification of multiparametric magnetic resonance imaging (mpMRI) to determine the likelihood of prostate cancer (PCa) in patients with haemospermia. Fifty-one patients presenting with haemospermia between 2018 and 2020 were included in this retrospective study. Forty-two of the patients (82.4%) were over 40years, and the median prostate-specific antigen (PSA) level was 1.4ng/ml. Fourteen of the patients (27.5%) had recurrent haemospermia. All patients underwent mpMRI, and assessments were classified according to PI-RADS v2. The mpMRI revealed PI-RADS one to four lesions in 10 (19.6%), 30 (58.8%), 6 (11.8%) and 5 (9.8%) patients respectively. One patient with PI-RADS 3 and five with PI-RADS 4 lesions underwent cognitive fusion prostate biopsy depending on MRI findings, and two patients with PI-RADS 4 lesions were diagnosed with PCa. Patients with haemospermia and risk factors, that is aged over 40years, a high PSA level or familial history of PCa, need a more thorough evaluation with mpMRI.

Standard MRI May Not Predict Specific Acute Anterior Cruciate Ligament Rupture Characteristics.

Background:There has been renewed interest in the concept of anterior cruciate ligament (ACL) suture repair (ACLSR). Morphologic characteristics of the ruptured ACL remnant play a role in deciding whether a patient is eligible for ACLSR. However, no classification of these characteristics of ACL rupture on magnetic resonance imaging (MRI) scans has yet been compared with intraoperative findings in the context of ACLSR.Purpose:To investigate the value of using preoperative MRI to predict specific characteristics of acute complete ACL rupture.Study Design:Cohort study (diagnostic); Level of evidence, 2.Methods:A total of 25 patients were included. Two radiologists classified ACL rupture location and pattern on preoperative 1.5-T MRI scans with a standard sequence; the results were compared with the corresponding findings at arthroscopy conducted by a single surgeon. The agreement between the MRI and surgical findings was calculated using Cohen κ values. Furthermore, the reliability coefficients of the MRI classifications within and between radiologists were calculated.Results:The agreement between MRI classification and arthroscopic findings for ACL rupture location was slight (Cohen κ, 0.016 [radiologist 1] and 0.087 [radiologist 2]), and for ACL rupture pattern, this was poor to slight (Cohen κ, <0 and 0.074). The intraobserver reliability of MRI classification for ACL rupture location was moderate for radiologist 1 and slight for radiologist 2 (Cohen κ, 0.526 and 0.061, respectively), and for ACL rupture pattern, this was slight for radiologist 1 and 2 (Cohen κ, 0.051 and 0.093, respectively). The interobserver reliability of MRI classification for ACL rupture location and pattern was slight between radiologists (Cohen κ, 0.172 and 0.040, respectively).Conclusion:In the current study, we found poor to slight agreement between MRI classification and arthroscopic findings of specific ACL rupture characteristics. In addition, the intra- and interobserver reliability for MRI classification of the ACL rupture characteristics was slight to moderate.

Study of the Feasibility of Conventional MR Images and Magnetic Resonance (MR)/Ultra-Short Echo Time (UTE) Technique in the Evaluation of Lumbar Disc Degeneration in the Axial Plane

The advantage and value of magnetic resonance (MR)-ultra-short echo time (UTE) technique in the displaying the structures of degenerative intervertebral discs were observed through evaluating the lumbar disc degeneration in the axial plane by conventional MR image and MR-UTE technique. A total of 160 examinees screened by inclusion and exclusion criteria were enrolled, and a total of 800 intervertebral disks were involved, and the degree of intervertebral disc degeneration was classified by conventional MR image classification method. After the first echo image of MR-UTE technique was integrated with the contour, the image entered UTE cartilage endplate, fibrous ring and osseous endplate. The integrity of the intervertebral disc from the above and below the intervertebral disc and the correlation of the structural endplate with the integrities of cartilage endplate, fibrous ring and osseous endplate under double echo sequence were analyzed, and the correlation analysis of the two categorical variables was conducted using rank correlation analysis. Under the MR-UTE sequence, there was no statistical significant difference in structural integrity constituent ratio (intact structure only on one side, incomplete structure on both side) between the non-low back pain group and the low back pain group (P &gt; 0.05), the proportion of complete structure disappearance was lowest, and the proportion of partial structure disappearance was highest; there was a significant difference in the structural integrity of fibrous ring between the non-low back pain group and the low back pain group (P &lt; 0.05), there was a statistical significant difference in structural integrity of fibrous rings between the two groups (P &lt; 0.05); there was a statistical significant difference in pain degree between the partial structure disappearance group and the complete structure disappearance group (P &lt; 0.05). There was no significant difference in pain degree between the groups (intact structure only on one side and incomplete structure on both side); conventional MR images showed no significant difference in the intervertebral disc degeneration degree between the non-low back pain group and low back pain group (P &gt; 0.05), and the intervertebral disc degeneration degree had no significant correlation with the lower back pain. MR-UTE sequence and T2WI sequence showed that there was a correlation between different intervertebral disc degeneration degrees and the structural integrity distributions of cartilage endplate, osseous endplate and fibrous rings in two groups of patients with or without low back pain (P &lt; 0.05). Whether low back pain exists or not has little effect on the intervertebral disc degeneration degree on T2WI. MR-UTE double echo sequence scan reveals that the intervertebral disc degeneration degree is positively related to the integrity and integrity distribution of cartilage endplate, fibrous ring and bone endplate.

Three‐class classification of brain magnetic resonance images using average‐pooling convolutional neural network

AbstractBrain tumor image classification is one of the predominant tasks of brain image processing. The three‐class brain tumor classification becomes a trivial task for researchers as each tumor exhibit distinct characteristics. Existing classification models use deep neural networks and suffer from high computational cost. We have proposed an eight‐layer average‐pooling convolutional neural network to address three‐class brain tumor classification. The proposed model uses three convolution blocks along with a dense layer and a softmax layer. We have utilized N‐adam optimizer with a sparse‐categorical cross‐entropy loss function to improve the learning rate. The proposed model has been evaluated using a dataset consists of 3064 brain tumor magnetic resonance images. The proposed model outperforms state‐of‐the‐art models with 97.42% accuracy and takes lesser computation time than its competitive models.

Transfer learning networks with skip connections for classification of brain tumors

AbstractThis article presents a transfer learning model via convolutional neural networks (CNNs) with skip connection topology, to avoid the vanishing gradient and time complexity, which are usually common in transfer learning networks. Three pretrained CNN architectures, namely AlexNet, VGG16 and GoogLeNet are employed to equip with skip connections. The transfer learning is implemented through fine‐tuning and freezing the CNN architectures with skip connections based on magnetic resonance imaging (MRI) slices of brain tumor dataset. Furthermore, in the preprocessing, a frequency‐domain information enhancement technique is employed for better image clarity. Performance evaluation is conducted on the transfer learning networks with skip connections to obtain improved accuracy in brain MRI classifications.

Neuroimaging Patterns and Function in Cerebral Palsy-Application of an MRI Classification.

Background: Cerebral palsy (CP) is a disorder of movement and posture and every child with CP has a unique composition of neurological symptoms, motor severity, and associated impairments, constituting the functional profile. Although not part of the CP definition, magnetic resonance imaging (MRI) sheds light on the localization, nature, and severity of brain compromise. The MRI classification system (MRICS), developed by the Surveillance of Cerebral Palsy in Europe (SCPE), describes typical MRI patterns associated with specific timing of vulnerability in different areas of the brain. The classification has proven to be reliable and easy to use.Aims: The aim of this study is to apply the MRICS on a large dataset and describe the functional profile associated with the different MRI patterns of the MRICS.Materials and Methods: Data on children with CP born in 1999–2009 with a post-neonatal MRI from 20 European registers in the JRC-SCPE Central Registry was included. The CP classification and the MRICS was applied, and The Gross Motor Function Classification (GMFCS) and the Bimanual Fine Motor Function (BFMF) classification were used. The following associated impairments were documented: intellectual impairment, active epilepsy, visual impairment, and hearing impairment. An impairment index was used to characterize severity of impairment load.Results: The study included 3,818 children with post-neonatal MRI. Distribution of CP type, motor, and associated impairments differed by neuroimaging patterns. Functional profiles associated with neuroimaging patterns were described, and the impairment index showed that bilateral findings were associated with a more severe outcome both regarding motor impairment and associated impairments than unilateral compromise. The results from this study, particularly the differences in functional severity regarding uni- and bilateral brain compromise, may support counseling and service planning of support of children with CP.

Classification of Pituitary Tumor and Multiple Sclerosis Brain Lesions through Convolutional Neural Networks

Automatic classification of Brain Tumor and brain Lesions has become a very important step in the field of medical image analytics. The machine learning/Deep learning approaches are playing a tremendous role in the field of medical imaging classification, due to the drastic changes in the field of computing power and image analytics techniques. The deep learning, which is the subfield of machine learning, is playing the major role in the automatic classification of Magnetic Resonance Images (MRIs) having various brain abnormalities. Convolutional Neural Networks are widely used for the classification and detection of various brain disorders. In this research paper, Convolutional Neural Networks are designed with considering various learning parameters for the classification of Multiple Sclerosis Brain Lesions and Pituitary Tumor. In the proposed research, T1-weighted Contrast-enhanced Magnetic Resonance images are preprocessed with various image-preprocessing approaches such as to resize the images, to convert the images into suitable image format so that the experimental work can be performed with deep learning in the Matlab environment. The Experiment is conducted with the dataset of Multiple Sclerosis and Pituitary Tumor each of having 718 and 930T1-weighted MRI images respectively. The experimental results we achieved 99.7% classification accuracy of pituitary Tumor, and 99.2% accuracy of Multiple Sclerosis brain Lesions. The average accuracy of both classifications is 99.55%. The precision of the classification of Pituitary Tumor is 99.7, recall value is 99.7 and the f1_score of the classification is 99.7%. Similarly, the Precision of the classification of Multiple Sclerosis Brain Lesions is 99.15%, the recall value is 99.15%, and the f1_score is 99.15%. The purposed approach of the Convolutional Neural Network architecture exhibited outstanding performance as compared to other research outcomes.

Two-Dimensional Stockwell Transform and Deep Convolutional Neural Network for Multi-Class Diagnosis of Pathological Brain.

Since the brain lesion detection and classification is a vital diagnosis task, in this paper, the problem of brain magnetic resonance imaging (MRI) classification is investigated. Recent advantages in machine learning and deep learning allows the researchers to develop the robust computer-aided diagnosis (CAD) tools for classification of brain lesions. Feature extraction is an essential step in any machine learning scheme. Time-frequency analysis methods provide localized information that makes them more attractive for image classification applications. Owing to the advantages of two-dimensional discrete orthonormal Stockwell transform (2D DOST), we propose to use it to extract the efficient features from brain MRIs and obtain the feature matrix. Since there are some irrelevant features, two-directional two-dimensional principal component analysis ((2D)2PCA) is used to reduce the dimension of the feature matrix. Finally, convolution neural networks (CNNs) are designed and trained for MRI classification. Simulation results indicate that the proposed CAD tool outperforms the recently introduced ones and can efficiently diagnose the MRI scans.

Local Frequency Descriptor and Hybrid Features for Classification of Brain Magnetic Resonance Images using Ensemble Classifier

Local Frequency Descriptor and Hybrid Features for Classification of Brain Magnetic Resonance Images using Ensemble Classifier

Ensemble learning-based classification on local patches from magnetic resonance images to detect iron depositions in the brain

Iron deposition in the brain has been observed with normal aging and is associated with neurodegenerative diseases. The automated classification of brain magnetic resonance images (MRI) based on iron deposition in basal ganglia region of the brain has not been performed, to our knowledge. It is very difficult to analyse iron regions in brain using simple MRI techniques. The MRI sequence namely susceptibility weighted imaging (SWI) helps to distinguish brain iron regions. The objective of our work is to investigate the iron regions in selected areas of basal ganglia region of brain and classify MR images. The study included a total of 60 MRI images which consists of 40 subjects with iron region and 20 subjects of healthy controls. We performed Gaussian smoothing followed by construction of 40 localised patches of each MR image based on iron and normal regions. Grey level co-occurrence matrix (GLCM) features are extracted from the patches and fed to random forest (RF) classifier for patch-based classification of iron region. Training of data patch features was done by random forest classifier and the performance of classifier in terms of accuracy was measured. The experimental results show that the proposed localised patch-based approach for classification of brain iron images using random forest classifier achieved 96.25% classification accuracy in identifying normal and iron regions from brain MR sequences.

Ensemble learning-based classification on local patches from magnetic resonance images to detect iron depositions in the brain

Iron deposition in the brain has been observed with normal aging and is associated with neurodegenerative diseases. The automated classification of brain magnetic resonance images (MRI) based on iron deposition in basal ganglia region of the brain has not been performed, to our knowledge. It is very difficult to analyse iron regions in brain using simple MRI techniques. The MRI sequence namely susceptibility weighted imaging (SWI) helps to distinguish brain iron regions. The objective of our work is to investigate the iron regions in selected areas of basal ganglia region of brain and classify MR images. The study included a total of 60 MRI images which consists of 40 subjects with iron region and 20 subjects of healthy controls. We performed Gaussian smoothing followed by construction of 40 localised patches of each MR image based on iron and normal regions. Grey level co-occurrence matrix (GLCM) features are extracted from the patches and fed to random forest (RF) classifier for patch-based classification of iron region. Training of data patch features was done by random forest classifier and the performance of classifier in terms of accuracy was measured. The experimental results show that the proposed localised patch-based approach for classification of brain iron images using random forest classifier achieved 96.25% classification accuracy in identifying normal and iron regions from brain MR sequences.

MR-brain image classification system based on SWT-LBP and ensemble of SVMs

In this paper, we present an efficient magnetic resonance (MR) image classification system. At the first stage, the brain image is decomposed into several subbands using stationary wavelet transform (SWT). Then, local binary patterns (LBP) with reduced histograms are computed for each subband to form several primary feature vectors. Principal components analysis (PCA) followed by linear discriminant analysis (LDA) are then applied to each primary feature vector in order to transform them into new lower-dimension feature vectors. The third stage consists of using an ensemble of support vector machines (SVMs) in order to build voters and make the final decision on the requested image. The designed system is evaluated on 255 brain images with five-fold cross-validation. Experimental results show that the proposed system achieves a classification rate of 99.78% which outperforms the existing brain classification approaches.

A Multi-Task Network for Cardiac Magnetic Resonance Image Segmentation and Classification

Cardiomyopathy is a group of diseases that affect the heart and can cause serious health problems. Segmentation and classification are important for automating the clinical diagnosis and treatment planning for cardiomyopathy. However, this automation is difficult because of the poor quality of cardiac magnetic resonance (CMR) imaging data and varying dimensions caused by movement of the ventricle. To address these problems, a deep multi-task framework based on a convolutional neural network (CNN) is proposed to segment the left ventricle (LV) myocardium and classify cardiopathy simultaneously. The proposed model consists of a longitudinal encoder–decoder structure that obtains high- and low-level features at the same time. The encoder employs a feature pyramid module (FPM) and dense atrous convolution (DAC) to extract features from images with variable scales for classification. Meanwhile, the decoder leverages the subpixel layer to recover spatial information caused by downsampling in the encoder for segmentation. The approach was verified using 654 magnetic resonance images. It achieved a Dice similarity coefficient (DSC) metric of 82.14% on segmentation and a classification accuracy of 95.72%, with an area under the receiver operating characteristic curve (AUC) of 97.88%. The proposed method can aid in the segmentation of cardiac magnetic images and improve the classification accuracy of cardiopathy.