BrainCTImageStrokeDetection-Segmentation

This project firstly aims to classify brain CT images into two classes namely 'Stroke' and 'Non-Stroke' using convolutional neural networks. In the second stage, the task is making the segmentation with Unet model.

The dataset used in this project is taken from Teknofest2021-AI in Medicine competition. It contains 6000 CT images.

Detection Accuracy: %94.57 (900 Test images) | Segmentation IOU Score: %71.04 (440 Test images)

The model architecture used for the detection task: VGG16

The model architecture used for segmentation task: UNET

Project includes 8 different phases:

1-Working with DICOM files, getting the images in a correct way to be able to classify/segment it easily with Deep Learning methods.

2-Creating datasets with the correct structure to use them in the project.

3-Making some required operations to images such as contouring-cropping, removing noise and centering brains to convert them into a standart format.

4-Making the operations suitable to the specific task , e.g centering is not suitable to segmentation task.

5-Making different augmentations for the two separate task.

6-Constructing the models.

7-Training/testing the models.

8-Choosing the best model.

ACCURACY VERSUS EPOCHS (visualization of accuracy over training epochs)

EXAMPLES

Segmentation of hemorrhagical stroke.

Segmentation of ischemic stroke.