• The performance of Deep learning in brain tumor detection
  • Faezeh pourshakori (corresponding author),1,* Mohammad Amoozadeh,2
    1. Student research committee, Anzali International Medical Campus, Guilan University of Medical Sciences, Guilan, Iran
    2. Student research committee, Anzali International Medical Campus, Guilan University of Medical Sciences, Guilan, Iran


  • Introduction: Brain tumor is considered as one of the deadliest diseases in the world due to its increase affect and mortality rate in all age groups. The World Health Organization (WHO) claims that about 10 million deaths are recorded every year because of brain cancer. An early tumor diagnosis implies a faster response in treatment, which helps to improve patient’s survival rate with correct treatment decision that extent from surgery to radiotherapy and chemotherapy. Localization and classification of brain tumors in large medical images databases, taken in routine clinical tasks by manual procedure such as Magnetic Resonance Imaging (MRI) and Computed Tomography (CT), have a high cost both in effort and time. An automatic detection, location, and classification procedure is desirable and worthwhile. Consequently, using artificial intelligence (AI) techniques has become necessary in the automated detection and segmentation however, the manual methods results are regarded as the ground truth for developing and evaluating automated methods including deep learning which is an efficient and popular branch of AI. Deep learning techniques are commonly adopted for brain tumor MRI image detection and segmentation therefore this article reviews some used deep learning techniques in brain tumor MRI image detection and discusses the impact performance of deep learning models.
  • Methods: This review was conducted through electrical scientific databases, including PubMed, Scopus, and Google Scholar using keywords such as “deep learning”, “brain tumor”, “detection”, “segmentation”. After these articles were reviewed, a general conclusion was extracted from all the articles.
  • Results: The studies utilized several datasets to assess the performance of different models therefore the proposed model is evaluated on different parameters including Accuracy, Precision, Recall, F1-measure. One of the models with satisfactory results was CNN-LSTM in terms of accuracy 99.1%, precision 98,8%, recall 98.8% and F-1 measure 99.0%. Another model which had a good performance used both LRelu and ReLu activation functions although this Dl model will stop growing when the ReLu problem reaches its end. The model EfficientNetB2 outpeformed the other variations significantly, achieving an F1-score of 98.71%, a test accuracy of 98.86%, precision of 98.65% and recall of 98.77%. Even the TumorDetNet Dl model had the great results for optimal accuracy of 100% for classifying brain tumors. If the type of tumor is considered in performance results; glioma achieved the highest test accuracy of 99.60% while the pituitary tumor closely followed with a test accuracy of 97.88% and with highest F1-score, and specificity, with values of 99.54%, 99.53%, 99.54% and 99.81%. The values obtained for all the segmentation metrics are remarkable with average values of Dice=0.828, Sensitivity=0.940, and pttas=0.967.
  • Conclusion: When compared to manual procedures, these technologies give enhaced accuracy, volume reduction, and speed, however Deep learning require a significant amount of data to train models, otherwise, the predictive performance may suffer. Over all the Deep learning can be used to assist medical doctors in the diagnostic of brain tumors and some successful methods were named such as multiscale CNN which uses three processing pathways, is able to successfully segment and classify the three kinds of brain tumors in the datas, we assume CNN-LlSTM is the best for detecting the MR brain images also the TumorDetNet model had a great performance for classifying brain tumors into malignant and benign. Despite the perfect performance of AI, more research needs to be conducted in the field to discuss clearly about its performance.
  • Keywords: Deep learning detection detection localization segmentation