Hans Pinckaers — Public notes

About me

Hi! Welcome to my small website. I'm currently a PhD candidate working on deep learning in medical imaging (specifically histopathology), at an university hospital in The Netherlands.

Public notes

Coffee stuff

• Decaffeinated coffee beans — sources in The Netherlands
• My coffee gear
• Notes on dialing in an espresso

Deep learning stuff

• Diagnosing learning curve problems
• Papers on Concept Activation Vectors
• Regression with fast.ai
• Use fast.ai transforms with regular PyTorch

Programming stuff

• Fixing coc-pyright and anaconda import errors
• Make conda fast — how I use conda

Other stuff

• Host tiddlywiki for free with authentication on alwaysdata
• Using a Sony alpha camera as a webcam in Ubuntu
• Using a Sony alpha camera as a webcam in macOS Catalina

My research

My PhD, so far, has been about training a network memory efficient, resulting in a novel method which gives the possibility to train a ResNet34 with gigantic images (up to 268 megapixels) end-to-end with a single label per image.

• Training convolutional neural networks with megapixel images
  • First proof of concept of my method to train memory efficiently.
  • Published at the MIDL 2018 conference.
• Streaming convolutional neural networks for end-to-end learning with multi-megapixel images
  • Full paper about the method.
  • Read this if you want to know how it works.
  • Published in IEEE Transactions on Pattern Analysis and Machine Intelligence. Read older preprint at arXiv
• High resolution whole prostate biopsy classification using streaming stochastic gradient descent
  • First real application of the method.
  • Published at SPIE Medical Imaging 2019.
• Detection of prostate cancer in whole-slide images through end-to-end training with image-level labels
  • Full paper about the application on prostate biopsies
  • Currently in review.
• Neural Ordinary Differential Equations for Semantic Segmentation of Individual Colon Glands
  • Fun little sideproject
  • Published in Medical Imaging meets NeurIPS - 2019