Project for NecstCamp NL2: Automatic Gleason Grading
Credits to:
- Luca Pagano
- Ernesto Natuzzi
- Daniele Kota Russica
For a more documentated explanation, check report.pdf
Prostate Cancer (PCa) is the sixth most common and second deadliest cancer among men worldwide. The aggressiveness of prostate cancer is measured by Gleason grading, a system based on the appearance of cancer cells. It is usually performed via visual inspection (with a microscope) of the prostate tissue by expert pathologists. However, this is a time-consuming task and suffers from very high inter-observer variability. Automatic computer-aided methods have the potential for improving the speed, accuracy, and reproducibility of the results. This challenge aims at the automatic Gleason grading of prostate cancer from H&E-stained histopathology images through the use of a neural network, performing multiclass semantic segmentation.
- 244 patients with six respective labels
- 96 patients without labels
For the training we split the 244 patients in a training dataset and a validation dataset, following the 80-20 convention
Each patient has from 3 to 6 label, each annotated by an expert pathologist, with 6 pixel values (0 for background, 1-5 for Gleason Grading).
Firstly, we decided to combine them through the STAPLE algorithm, in order to have a single label per image that should best represent the tumour grading.
We decided for hardware reason to resize the image from 5120x5120 (on average) to 1024x1024 and divide them into patches of 512x512 with 50% overlapping, for a total of 9 patches per image.
For the training we used a UNet with an EfficientNetB4 as the backbone; we used data augmentation for our 244 patients, in order to improve the accuracy of the model and avoid overfitting
Below some predictions from the validation dataset
