SHERLOCK

Abstract

Whole Slide Images (WSI) are gigantic images (e.g. 100k x 100k pixels) of tissue samples. The goal of SHERLOCK is to detect cancer cells in those tissue samples. We do this by using a Masked Autoencoder (MAE) that we train on patches of inflammatory tissues samples from the PanNuke dataset. The benefit of using an MAE is that unlike supervised learning the WSI’s don’t need to be labeled. This is important because it will save a lot of time and money that would be spent on labeling WSI’s. Also it’s easy to miss things in WSI’s because they are so large. SHERLOCK aims to be like the famous detective and not miss any detail no matter how small.

Methods

SHERLOCK uses PyTorch as its machine learning framework. We created custom PanNuke Datasets that we trained a Masked AutoEncoder (MAE) on. This uses Vision Transformers (ViT’s) to do binary classification on our dataset.

Results

We finetune our model and do binary classification which detects if the patch from the WSI is inflammatory or not.