Brain Tumor Segmentation Based on Deep Learning's Feature Representation.

Area of Science:

• Neuroscience
• Medical Imaging
• Artificial Intelligence

Background:

• Brain tumors are a significant cause of mortality worldwide, necessitating early detection.
• Existing tumor prediction and segmentation methods often require specialist intervention, long runtimes, and complex feature extraction.
• Addressing these limitations is crucial for improving patient outcomes and diagnostic efficiency.

Purpose of the Study:

• To develop and evaluate a novel deep learning approach for simultaneous prediction and segmentation of cerebral tumors.
• To overcome the limitations of traditional methods by reducing reliance on specialist-labeled data and simplifying feature extraction.
• To enhance the accuracy and efficiency of brain tumor detection and delineation.

Main Methods:

• A two-phase approach utilizing a convolutional neural network (CNN) architecture was proposed.
• Phase one employed simple binary annotations (tumor presence/absence) to avoid specialist intervention.
• Phase two involved feeding prepared image data into the deep learning model for classification and subsequent segmentation based on CNN feature representations.

Main Results:

• The proposed model demonstrated high performance on the BraTS 2017 dataset, which includes various glioma types.
• Achieved an accuracy of 91% for brain tumor classification.
• Obtained a Dice similarity coefficient of 82.35% for brain tumor segmentation.

Conclusions:

• The developed CNN-based approach effectively achieves simultaneous prediction and segmentation of brain tumors.
• The method shows promising results in accuracy and segmentation performance, offering a potential improvement over existing techniques.
• This approach could facilitate earlier and more efficient detection and delineation of brain tumors.