UNet (and its variants) for Biomedical Image Segmentation

Sourodip Ghosh (~sourodip)




Different traditional methods in Deep Learning are used to detect the particular type of class/category and therefore, produce an output label. However semantic segmentations are often cumbersome and time-consuming. First introduced by O Ronneberger in 2015 (UNet paper), UNet is since then, used as a primary tool in the field of biomedical image segmentation. Since 2015, there has been a lot of gradients of UNet, fine-tuned on the particular base architecture, like UNet++, Nas-Unet, Vanilla U-Net, etc.

The aim is to discuss these sophisticated approaches for solving the segmentation problems of these target classes. The layout is structured as follows:

  1. Introduction: A little bit of description about the origin of architectures to analyze biomedical imaging (5 min)
  2. Difference between U-Net and a little bit of tweak in the algorithms (4 min)
  3. Detection and segmentation: A step further after detection (4 min)
  4. Use case of UNet vs variations: Using a particular dataset [Brain MRI] to explain the functionality (4 min)
  5. Results: Evaluation parameters (3 min)
  6. Model diversity and reliability (5 min)
  7. Questions! (5 min)

Link to presentation: https://1drv.ms/p/s!AmZTbGBdPZo9jieoZjz-g9W2TORf?e=hjDTPQ

Video: https://drive.google.com/file/d/1OMuNgGc3JqycYfz3W7RY-4u-pUDuELBd/view?usp=sharing


A little bit of basics in Machine Learning and Deep Learning is helpful. The rest would be covered in the course itself.

Speaker Info:

I am a senior year student, pursuing a Bachelor's in Technology, major in Electronics. I have been using Deep Learning for medical imaging for a while now. I have been publishing papers in the Biomedical image processing domain, using segmentation and analysis.

Recent articles:

Looking forward!

Speaker Links:

LinkedIn: https://www.linkedin.com/in/sourodip-ghosh-941424187/

Section: Data Science, Machine Learning and AI
Type: Talks
Target Audience: Intermediate
Last Updated: