Scribd Logo
Upload

Welcome to Scribd!

  • 5 views

    1.icicv2020-103 PPT

    Uploaded by

    Manoj hv
    The document presents a deep learning approach to predict brain tumors using MRI scans. It aims to classify scans as cancerous or non-cancerous with zero false negatives. The methodology uses transfer learning with pre-trained CNN models like Inception-V3, VGG-16 and ResNet-50. ResNet-50 achieved the best performance with an F1-score of 0.952, 95% accuracy and zero false negatives.

    Copyright:

    © All Rights Reserved
    Download as PDF, TXT or read online from Scribd

    1.icicv2020-103 PPT

    Uploaded by

    Manoj hv
    5 views19 pages
    The document presents a deep learning approach to predict brain tumors using MRI scans. It aims to classify scans as cancerous or non-cancerous with zero false negatives. The methodology uses transfer learning with pre-trained CNN models like Inception-V3, VGG-16 and ResNet-50. ResNet-50 achieved the best performance with an F1-score of 0.952, 95% accuracy and zero false negatives.

    Original Title

    1.ICICV2020-103_PPT

    Copyright

    © © All Rights Reserved

    Available Formats

    PDF, TXT or read online from Scribd

    Did you find this document useful?

    Is this content inappropriate?

    The document presents a deep learning approach to predict brain tumors using MRI scans. It aims to classify scans as cancerous or non-cancerous with zero false negatives. The methodology uses transfer learning with pre-trained CNN models like Inception-V3, VGG-16 and ResNet-50. ResNet-50 achieved the best performance with an F1-score of 0.952, 95% accuracy and zero false negatives.

    Copyright:

    © All Rights Reserved
    Download as PDF, TXT or read online from Scribd
    Download as pdf or txt
    5 views19 pages

    1.icicv2020-103 PPT

    Uploaded by

    Manoj hv
    The document presents a deep learning approach to predict brain tumors using MRI scans. It aims to classify scans as cancerous or non-cancerous with zero false negatives. The methodology uses transfer learning with pre-trained CNN models like Inception-V3, VGG-16 and ResNet-50. ResNet-50 achieved the best performance with an F1-score of 0.952, 95% accuracy and zero false negatives.

    Copyright:

    © All Rights Reserved
    Download as PDF, TXT or read online from Scribd
    Download as pdf or txt
    of 19

    See discussions, stats, and author profiles for this publication at: https://www.researchgate.

    net/publication/338686192

    Predictive modeling of brain tumor: a deep learning approach

    Presentation · January 2020


    DOI: 10.13140/RG.2.2.22911.76963

    CITATIONS READS
    49 556

    3 authors, including:

    Priyansh Saxena Saumil Maheshwari


    ABV-Indian Institute of Information Technology and Management Gwalior Madhav Institute of Technology & Science Gwalior
    15 PUBLICATIONS 145 CITATIONS 22 PUBLICATIONS 401 CITATIONS

    SEE PROFILE SEE PROFILE

    All content following this page was uploaded by Priyansh Saxena on 19 January 2020.

    The user has requested enhancement of the downloaded file.


    JANUARY 18, 2020

    International Conference on
    Innovations in Computational Intelligence and Computer Vision
    (ICICV-2020)

    PREDICTIVE MODELING OF
    BRAIN TUMOR:
    A DEEP LEARNING APPROACH
    Priyansh Saxena, ABV-IIITM
    Akshat Maheshwari, ABV-IIITM
    Saumil Maheshwari, ABV-IIITM
    noissucsid rof scipot yeK
    Presentation
    Outline
    Introduction
    Objectives
    Related Works
    Research Methodology
    Results & Discussion
    Conclusion
    References

    01
    Introduction

    Brain tumor accounted for 85% to 90% of all primary CNS tumors.

    As of 2019 statistics, the total annual number of deaths from brain


    cancer across all ages, and both sexes is 247,143[1].
    Recent advancements in Deep Learning, provides state of the art
    solutions for detecting a brain tumor in early stages and can thus
    save lives by taking preemptive measures.

    02
    OBJECTIVES To develop a deep learning-based model
    which can classify the brain MRI scans as
    cancerous or non-cancerous with zero false
    negatives rate.

    To further enhance the performance by using


    CNN based transfer learning approach, which
    not only increases accuracy but help reducing
    the false negatives significantly.

    03
    Related Works

    Emblem Ke et al. applied SVMs on perfusion MRI[2] and achieved sensitivity &
    specificity of 0.76 and 0.82, respectively.

    El-Dahshan et al.[3] suggested Feedback Pulse Coupling Network (FPCNN) for


    image segmentation in the brain tumor diagnosis systems. It uses DWT, PCA &
    ANN as feature extractor and selector and classification, respectively.

    Rahmani & Akbarizadeh[4] proposed an unsupervised feature learning


    technique which was related to the collaboration of spectral clustering &
    sparse coding for segmentation of SAR images.

    04
    Related Works (Contd.)

    Chang[5] devised an algorithm which contained both FCN and Conditional


    Random Fields (CRF).

    Lai[6] used the depth information by implementing a 3D convolution model


    which utilized the correlation between slices. The 3D convolution network
    requires a large number of parameters. Moreover, when a dataset with fewer
    samples is used, 3D CNN often overfit.

    05
    RESEARCH
    METHODOLOGY

    Proposed methodology is organized into 3 different steps.


    The models are trained & tested using Brain Tumor by
    Navoneel[7].
    253 samples of brain MRI images out of which, 155 samples
    are of malignant tumor and 98 samples benign tumor.

    06
    DATA PREPROCESSING
    CROP NORMALIZATION
    Approach of determining of extreme points in contours.
    Determines the farthest north, south, east, and west (x, y)-coordinates along a given contour.
    Only the portion of the image containing the brain is cropped out [8].

    RESIZING OF IMAGES
    Input dataset contains images with a different dimensions and with different aspect ratio.
    Images in the dataset are resized to a preset format, since the pretrained models used require the
    images to be 224×224×3 dimensions.

    07
    ARTIFICIAL DATA
    AUGMENTATION
    This technique is used to artificially enhance the size of image training data by generating modified
    images using the original dataset, which enhance the capability of the model to learn and generalize
    better on future unseen data.
    Figure shows the first image as the original image present in the dataset and the remaining images
    are generated by this technique[9].

    08
    TRANSFER LEARNING

    Uses 3 pre-trained CNN models,


    Inception-V3 which acts as feature extractors.
    VGG-16
    Inception-V3
    Resnet-50

    09
    VGG-16

    Resnet-50
    RESULTS & DISCUSSION
    The dataset of 253 images is divided into-183 training images, 50 images for
    the validation set, and 20 testing images.
    The performance for the proposed methodology was measured in terms of
    Cohen’s kappa(κ), F1-score, area under the ROC curve (AUC-ROC) and test
    accuracy.

    10
    RESULTS & DISCUSSION (Contd.)
    Resnet-50 achieved the highest F1-score, accuracy & meager false-negative
    rateas zero on the test data.

    Performance of Inception-V3 was very close to a random classifier since its


    AUC is 0.55 as it suffered from overfitting.

    11
    RESULTS & DISCUSSION (Contd.)

    Inception-V3 VGG-16 ResNet-50

    TRAINING AND VALIDATION AC C URAC Y ANALYS IS


    Inception-V3 suffers from overfitting.
    VGG-16 achieves a decent accuracy.
    ResNet-50 outperforms the other two.
    12
    RESULTS & DISCUSSION (Contd.)

    Inception-V3 VGG-16 ResNet-50

    TRAINING AND VALIDATION LOS S ANALYS IS


    Inception-V3 suffers from high validation loss.
    VGG-16 shows substantially less validation loss.
    ResNet-50 shows the least validation loss out of the above two.
    13
    Conclusion & future scope

    Resnet-50 achieved the best F1 score (0.952) and highest test accuracy (95%)
    with the zero false-negative rate (primary concern in an automatic diagnostic
    system and thus suitable for practical applications).
    Inception-V3 suffered from overfitting and is just slightly better than a random
    classifier with an accuracy 0.55.
    The future possibilities for increasing the accuracy further can be done by using
    new state of the art architectures by using transfer learning technique.
    Extensive hyper-parameter tuning, and a better preprocessing technique can be
    devised to enhance the model performance further.

    14
    References
    1. Ritchie, Hannah. "How Many People in the World Die from Cancer?" Our World in Data, Institute for
    Health Metrics and Evaluation (IHME) , 1 Feb. 2018, https://ourworldindata.org/how-many-people-in-
    the-world-die-from-cancer.
    2. Emblem, Kyrre E., et al. "Predictive modeling in glioma grading from MR perfusion images using
    support vector machines." Magnetic Resonance in Medicine: An Official Journal of the International
    Society for Magnetic Resonance in Medicine 60.4 (2008): 945-952.
    3. El-Dahshan, El-Sayed A., et al. "Computer-aided diagnosis of human brain tumor through MRI: A
    survey and a new algorithm." Expert systems with Applications 41.11 (2014): 5526-5545.
    4. Rahmani, Masoumeh, and Gholamreza Akbarizadeh. "Unsupervised feature learning based on sparse
    coding and spectral clustering for segmentation of synthetic aperture radar images." IET Computer
    Vision 9.5 (2015): 629-638.
    5. Chang, Peter D. "Fully convolutional deep residual neural networks for brain tumor segmentation."

    15
    International Workshop on Brainlesion: Glioma, Multiple Sclerosis, Stroke and Traumatic Brain
    Injuries. Springer, Cham, 2016.
    References (Contd.)

    6. Lai, Matthew. "Deep learning for medical image segmentation." arXiv preprint arXiv:1505.02000
    (2015).
    7. Chakrabarty, Navoneel. "Brain MRI Images for Brain Tumor Detection". Kaggle, 14 Apr. 2019,
    https://www.kaggle.com/navoneel/brain-mri-images-forbrain-tumor-detection.
    8. Rosebrock, Adrian. Finding Extreme Points in Contours with OpenCV.âĂİ PyImageSearch, 11 2016,
    https://www.pyimagesearch.com/2016/04/11/findingextreme-points-in-contours-with-opencv/.
    9. Tanner, Martin A., and Wing Hung Wong. "The calculation of posterior distributions by data
    augmentation." Journal of the American statistical Association 82.398 (1987): 528-540.
    10. Krizhevsky, Alex, Ilya Sutskever, and Geoffrey E. Hinton. "Imagenet classification with deep
    convolutional neural networks." Advances in neural information processing systems. 2012.

    16
    View publication stats

    Thank you
    ICICV-2020

    You might also like