Description:

Introduction

Medical 3D imaging uses techniques like CT, MRI, PET and ultrasound to provide detailed spatial information about internal structures, aiding accurate diagnosis and treatment planning. Globally, an estimated 400 million of these scans are taken every year. 2D convolutions, designed for 2D images, don't work on 3D images due to the depth dimension. Specialized algorithms like 3D convolutions and deep learning architectures are used to efficiently process volumetric medical data and improve patient care.

Outline

• Handling medical image [20 mins]
  • Introduction to SimpleITK and Pydicom.
  • creating cache for training deep learning models.
• Image processing [20 mins]
  • Extending albumentations for 3D images.
  • Adding transforms specific to 3D images.
• Deep learning Architectures. [20 mins]
  • Resnet 2D to 3D.
  • RetinaNet 2D to 3D
  • UNet 2D to 3D.
• Implementing loss function for 3D image specific problems (using torch/numpy broadcasting) [10 mins]
  • detection and segmentation: 3D IOU
  • only detection: DIoU, CIoU, GIoU
  • only segmentation: Dice score
• Train a 3D segmentation model on segmentation Lung from Chest CT. [10 mins]
• Train a 3D object detection model on identifying lung nodules in Chest CT (we will use LUNA16 data) [10 mins]
• Implementing 3D metrics like mAP and mAR to support both 2D and 3D bounding boxes. [20 mins]

Extras - We will do this if time permits

• Implementing ViT for 3D medical images. [20 mins]
• Training Masked AutoEncoders on 3D medical images [20 mins]

Prerequisites:

• Deep learning - Train an image classification model using PyTorch.

Speaker Info:

I am Prakash, the Director of Data Science at Qure.AI. My main area of expertise is in deep learning, where I focus on developing applications for processing 3D medical data. Before my current role, I gained valuable experience in deploying image processing applications across various industries, including retail, satellite technology, and document analysis.

Speaker Links:

• https://medium.com/@14prakash
• https://github.com/prakashjayy/computer_vision/tree/master/papers
• https://www.linkedin.com/in/prakash-vanapalli-99909b3a/