AlgoSurg: Markerless Tracking for Surgical Navigation

Domains: Image processing, Computer vision, 3D geometry and point clouds, Surgical navigation

Project Overview: The project aims to develop markerless tracking algorithms for surgical navigation, utilizing Intel Realsense cameras to capture RGB and Depth images of a mock bone in a non-surgical environment. The process involves two real-time CNN-based algorithms for ROI detection and point cloud segmentation. The ultimate goal is to register a virtual 3D model of the bone with the physical bone using 3D-3D point registration algorithms.

The project is in its initial stages, with a focus on Phase 1 – ROI detection in RGB images. Relevant papers for reference have been provided, and off-the-shelf algorithms like ROINet and SegNet will be used and customized for the given problem.

Roles: 1 Image Processing Developer
Stipend: Paid project (join the community for more details)
Project Duration: 2 months

Tasks/Deliverables:

Phase 1: Develop and customize a CNN-based algorithm for ROI detection in RGB images.
Phase 2: Implement a CNN-based algorithm for point cloud segmentation.
Collaborate with the AlgoSurg team in image data creation, collection, and manual segmentation.
Weekly meetings to track progress, seek guidance, and provide support.

Candidates can apply by filling out the Google Form provided, which includes personal information and your academic background.

Deadline: 11:59 PM, 7th February

Submission Link: https://forms.gle/rPWg5wGGzpWpKDN3A

(Optional) Assignment:
The objective is to implement a basic image segmentation algorithm using convolutional neural networks (CNN). You can use any framework of your choice (PyTorch, TensorFlow, sci-kit learn etc):

1. Dataset Preparation: Create a small synthetic dataset containing images with simple shapes on a contrasting background
2. Masks: Generate corresponding binary masks for each images, where the shapes are segmented
3. Model: Design a simple CNN architecture or use the smaller version of pre-built models for image segmentation
4. I/O: The input should be an RGB image and the output a binary mask indicating the segmented region
5. Training: Aim for a brief iteration on the training set (10-15 epochs)
6. Inference: Report key training metrics, perform segmentations using the model on a few images from the validation set and document your findings

For any general queries, join the ProSpace WhatsApp group- https://chat.whatsapp.com/E09qtrcuShp1uf2w82LCsa

For assignment queries, contact:

Email: satyamm435@gmail.com

Phone: 9324865787