Segmentation of CT images

Inspiration

Computer vision with Deep Learning has rich applications in the field of Medical Imaging. We found this project as a great opportunity to learn and understand this area of Deep Learning.

What it does

We tried segmenting Alpha Dart seeds from the CT images and then trying to make 3D image segmentation of the seeds.

How we built it

We used U-Net architecture to segment the seeds and then trying to predict the mask for the test data.

Challenges we ran into

The data was of very large magnitude, so we couldn't load all the slices of the images into our session, so we had to make them go on the fly(using a generator) and then train the model for that data. This is the major challenge that we faced, the seeds are of very small size and hence they occupy a very small region in the CT image, so when we tried to train the model, it was not able to learn the features of the seeds, and even after trying to do thresholding and pre-processing and using focal loss function, it couldn't learn the seeds.

Accomplishments that we're proud of

We were able to pre-process this huge data and without scaling the images we were able to predict the mask with better intensity and lower intensity for the background(Bones etc).

What we learned

We learned a new architecture U-Net and new segmentation approaches.

What's next for Segmentation of CT images

Trying different architecture to better understand the regions of the mask and using better pre-processing techniques, like passing the sliced image first through a high pass filter and then passing it to our network. Using Mask RCNN for instance segmentation and after the object is located calculating the co-ordinates and orientation by that.