Scribd
Upload
13 views

Brain Tumor Segmentation Using Deep Learning

Uploaded by

Nitya
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
13 views

Brain Tumor Segmentation Using Deep Learning

Uploaded by

Nitya
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 9

Brain Tumor

Segmentation Using
Deep Learning
This paper explores the use of deep learning techniques for brain tumor
segmentation in Magnetic Resonance Imaging (MRI) scans. The focus
is on utilizing a hybrid model combining the UNet architecture with the
pre-trained DenseNet121 network to achieve accurate segmentation of
brain tumors, particularly in smaller sub-regions.
The Importance of Brain Tumor Segmentation
1 Early Detection 2 Treatment Planning
Accurate brain tumor segmentation is crucial for Segmentation provides valuable information for
early detection, which is vital for improving treatment planning, helping surgeons determine
survival rates. Early diagnosis allows for timely the tumor's precise location, size, and extent.
intervention and treatment, increasing the This information is essential for guiding surgical
chances of successful outcomes. procedures and radiation therapy.

3 Monitoring Tumor Growth 4 Computer-Aided Diagnosis


Segmentation plays a role in monitoring tumor Computer-aided systems based on deep learning
growth over time. By comparing segmented can assist medical professionals in detecting
images from different time points, doctors can tumors more efficiently and accurately, reducing
track tumor progression and adjust treatment the time and effort required for manual analysis.
strategies accordingly.
Challenges in Brain Tumor Segmentation
Fuzzy Borders Data Imbalance Pixel-Based Classification
Limitations
Infiltrated glioma tumors often have Brain tumor datasets often exhibit
fuzzy borders, making it difficult to data imbalance, where certain tumor Traditional pixel-based
distinguish between normal and sub-regions (e.g., enhancing classification methods struggle to
abnormal tissues. This poses a tumors) are smaller and less capture local dependencies and
significant challenge for accurate frequent than others. This can lead contextual information, which is
segmentation. to biased model training and crucial for accurate tumor
reduced accuracy for these smaller segmentation.
regions.
The Proposed Hybrid DenseNet121-UNet Model

Preprocessing DenseNet121 UNet Decoder Postprocessing


Encoder
The proposed model The UNet decoder After segmentation, the
incorporates a The pre-trained upscales the features smaller sub-regions are
preprocessing step to DenseNet121 network extracted by the reassembled based on
address data imbalance acts as the encoder, encoder, reconstructing their original
and improve extracting features from the image with high- coordinates to obtain the
segmentation accuracy. the input image. Dense resolution details. Skip final segmented image.
This involves dividing connections between connections between
the image into smaller layers help to alleviate encoder and decoder
sub-regions based on the vanishing gradient layers help to preserve
tumor size and center problem and enhance semantic information.
coordinates. feature propagation.
Dataset and Evaluation Metrics
Dataset BRATS 2019

Tumor Types Whole Tumor, Core Tumor, Enhancing Tumor

Evaluation Metrics Dice Similarity Coefficient (DSC), Jaccard Index, Mean


Intersection-Over-Union (IoU), Sensitivity, Specificity
Experimental Results and Discussion

Whole Tumor Core Tumor Enhancing Tumor


The proposed model achieved a Dice The model achieved a DSC of 0.943 for The model achieved a DSC of 0.892 for
Similarity Coefficient (DSC) of 0.959 core tumor segmentation, indicating enhancing tumor segmentation,
for whole tumor segmentation, strong performance in identifying the demonstrating good performance in
demonstrating high accuracy in core tumor region, which includes identifying the enhancing tumor region,
identifying the entire tumor region. necrotic tissue, diminishing tumor, and which is often smaller and more
enhancing tumor. challenging to segment.
Comparison with Other Methods
Model Whole Tumor (DSC) Core Tumor (DSC) Enhancing Tumor (DSC)

UNet 0.859 0.807 0.829

ResUNet 0.894 0.862 0.828

VGG16-UNet 0.936 0.896 0.845

ResNet50-UNet 0.914 0.882 0.839

Proposed DenseNet121- **0.959** **0.943** **0.892**


UNet
Advantages of the Hybrid Model
Improved Accuracy Reduced Training Time
The hybrid model achieved higher DSC values for all The use of pre-trained DenseNet121 and image
tumor types compared to other methods, demonstrating preprocessing techniques helped to reduce training time,
its effectiveness in segmenting brain tumors accurately. making the model more efficient.

Enhanced Feature Extraction Improved Handling of Data Imbalance


The DenseNet121 architecture with its dense The preprocessing step effectively addressed the data
connections facilitated the extraction of more imbalance problem, leading to improved segmentation
comprehensive and informative features from the input accuracy for smaller tumor sub-regions.
images.
Conclusion and Future
Directions
The proposed hybrid DenseNet121-UNet model demonstrated
significant improvements in brain tumor segmentation accuracy,
particularly for smaller tumor sub-regions. This model offers a
promising approach for computer-aided diagnosis and treatment
planning in clinical settings. Future research could explore the use of 3D
convolutional neural networks for volumetric segmentation,
incorporating multi-modal MRI data, and developing more robust
methods for handling complex tumor shapes and variations.

You might also like