Gaurav Sahadev
Indiana
Siliguri, West Bengal
Intracranial Hemorrhage is referred to the type of bleeding inside the cranium of the skull, resulting from accidents or stroke mainly. According to survey, about 2.5 per 10,000 people are affected by ICH every year. About 44% of those affected die every year. ...learn more
Project status: Under Development
Mobile, Artificial Intelligence
Groups
Student Developers for AI
Intel Technologies
Other
Brain haemorrhage can be detected by performing CT scans of the brain. In severe cases, MRI is done.
CT scan or Computed Tomography is a noninvasive diagnostic imaging procedure that uses special X-ray measurements to produce horizontal, or axial, images (often called slices) of the brain. During a brain CT, the X-ray beam moves in a circle around the body, allowing many different views of the brain. The X-ray information is sent to a computer that interprets the X-ray data and displays it in a two-dimensional (2D) form on a monitor.
Expertise is required to interpret these scans, and even highly trained experts may miss subtle life-threatening findings. For head CT, a unique challenge is to identify, with near-perfect sensitivity and very high specificity, small subtle abnormalities on a multi-slice cross-sectional (three-dimensional [3D] ) imaging modality.
Here, the model can be trained with CT images of patients with intracranial haemorrhage and and without haemorrhage. Then the trained model is used to predict whether that particular patient has haemorrhage or not by only uploading a jpg/png file of their CT imaging. This makes the job more accurate and time-saving.
The entire workflow is as follows:
Collecting the dataset and importing the libraries.
Preparing the dataset (Pre-processing and Augmentation).
Model Building
Model evaluation and Results.
The dataset has been collected from Kaggle. It has been trained in Google Colab since the dataset needed is large.
Here, two important concepts has been utilised which are data augmentation and windowing.
The dataset is quite small therefore, we used data augmentation to reduce overfitting and to help the model generalize better.
We used the brain window images. We converted them to 3 channel images to suit the model. 20 images are set aside as holdout test set. The remainder of the data is split into 85% train and 15% validation. Keras Densenet121 encoder with a Unet decoder - Adam optimizer and dice loss are used as model. Instead of using DenseNet pre-processing, we simply normalized the images by dividing by 255.
There is a scope of forming a tensorflow.js app which can be published in the market.
Numpy.
Tensorflow.
Densenet.
Google Colaboratory.
https://github.com/DEBANJANAB/Brain-Hemorrhage-detection-in-CT-imaging