Histopathology is the study of diseases using microscopic examination of slides of tissues. These tissues of the patients often come from biopsy or surgery. Much of the machine learning related research in histopathology revolves around cancer, so these tissue slides are inspected for the presence of tumours or other markers that hint toward tumours.
Over the past few years, much progress has been made in using artificial neural networks for histopathology. Many datasets have been created for histopathology, with some of them being made publicly available. But most of the neural network architectures used for these datasets are the ones that have been developed for natural images (think images of animals, boats, cars, people etc.). These include architectures found via neural architecture search (an automated way of designing architectures) done for those natural image datasets. We believe that specialised architectures for histopathology can lead to better results. Specifically, we aim to search for architectures with an inductive bias for histopathological images. We'll use methods developed for neural architecture search to search for architectures. The project would aim to find an architecture that works well across multiple histopathological datasets and has clear advantages over existing architectures.
This project requires you to:
- Familiarise yourself with machine learning (deep learning in particular) and frameworks like PyTorch.
- Go through the neural architecture search literature and understand various approaches' pros, cons, and tradeoffs.
- Design novel objectives for neural architecture search and run experiments.
References:
[1] Cao, Yun-Hao, and Jianxin Wu. "A random cnn sees objects: One inductive bias of cnn and its applications." Proceedings Of The AAAI Conference On Artificial Intelligence. Vol. 36. No. 1. 2022.