Universally Slimmable & Dynamically Resizable and DUCKNet for Real-Time Pharynx Segmentation

• Tech Stack: Python, PyTorch, Albumentations, Matplotlib, GitHub, NumPy
• Paper URL: Link
• Slideshow URL: Link
• Github URL: Project Link

Developed a universally slimmable and dynamically resizable DUCKNet and UNet segmentation architectures capable of real-time (17 Hz) inference during endoscopic procedures. This was done by implementing slimmable and universally slimmable architectures with dynamic resizing, knowledge distillation, and transfer learning. The trained models continuously adapted to hardware constraints while maintaining accurate larynx and hypopharynx segmentation. These architectures allow for robust, efficient image analysis directly onsite, facilitating resource-aware medical diagnostics.

Created universally slimmable convolutional, transposed convolutional, residual and batch normalization modules to reconstruct DUCKNet and U-Net architectures for dynamic resizing during training and inference. The universally slimmable modules work by selectively and dynamically dropping weights to shrink the compute required for each layer within the neural network. Achieved 0.95 DICE score with models 20% of the the size of the original DUCKNet. In summary, we achieved state of the art performance with models 5 times smaller than the original DUCKNet model proposed in their paper.

Created training routines for the universally slimmble architectures using the "sandwich technique", knodwledge distillation, and in-place student-teacher learning.