Prostate Gleason Dataset

---

Prostate Gleason Dataset is an image database of prostate cancer biopsies derived from the PANDA dataset. The database consists of about 70K patches of size 256x256 pixels extracted from the PANDA's biopsies with overlap. The patches are grouped in 4 classes depending on their Gleason grade:

• Class 0: Stroma and Gleason 0
• Class 1: Gleason 3
• Class 2: Gleason 4
• Class 3: Gleason 5

resulting in the following dataset distributions:

My hope is to help improving ML-based prostate-cancer assessment by providing a novel patch labeled dataset.

Current computational pathology is usually based on labels of whole slide images of biopsies where the labels are ISUP grades or Gleason composite scores. These labels tend to be noisy and non-informative at the cellular and/or glandular level, for they provide information derived from large, non-contigous and non-omogeneous (w.r.t. the cellular/glandular visual properties of the cancer) areas of the biopsies.

Tipically, modelling this data for predicting the ISUP grade or Gleason score involves 2 steps:

1. finding a relationship between the given (whole-slide) label and the visual properties of the cancer;
2. tranforming this low level properties back into a whole-slide prediction.

The Prostate Gleason Dataset brings mainly two contributions:

1. aims to easy the computational burden on step 1) by providing finer-grained labels;
2. provides information that can be used for transfer learning to other computational pathology problems.

Parent Dataset Overview

---

10616 whole-slide images of H&E-stained biopsies. The images where collected from different scanners at the Radboud University Medical Center and Karolinska Institute. The former institution used a 3DHistech Pannoramic Flash II 250 scanner at 20x magnification and stored the final image at a pixel resolution of 0.48μm. The later used a Hamamatsu C9600-12 scanner and a Aperio ScanScope AT2 scanner, obtaining a pixel resolution of 0.45202μm and 0.5032μm respectively. The smallest image in the dataset has an area of 2304x4352 pixels and the largest is 55296x73728 pixels.

Part of the images were obtained from slices of the same biopsy and therefore they belong to the same patient.

All images were labelled with both an ISUP grade and Gleason score (primary pattern + secondary pattern) accordingly to the ISUP 2014 guidelines.

Furthermore, the biopsies comes with a ML-predicted mask bringing (relatively noisy) information on background vs stroma vs epithelium pixels.

Derivation Process

---

The parent dataset is filtered by Gleason score and just images with equal primary and secondary patterns are kept (Gleason 0+0, 3+3, 4+4 and 5+5). This aims to ensure the final patches don't contain non-omogeneous Gleason grades. The remaining images are divided into 512x512 pixels overlapping patches (stride 256 pixels). Then, 2 different filtering processes were used depending on wether the biopsy's Gleason score was equal to 0+0 or greater:

• Gleason 0+0, two different classes A and B are produced:

  • for class A a patch is kept if the number of stroma pixels is >=75% of the area of the patch, accordingly with the mask;
  • for class B a patch is kept if the number of Gleason 0 pixels is >=75% of the area of the patch, accordingly with the mask;
  • the two subsets are merged to produce class 0

• Gleason 3+3, 4+4, or 5+5: a patch is kept if the number of Gleason X pixels is >=75% of the area of the patch, accordingly with the mask; classes 1, 2 and 3 are produced for Gleason grades 3, 4 and 5 respectively.

In order to remove noisy entries, all the remaining patches went through a manual labelling step and they were finally resized to 256x256 pixels using nearest neighboor interpolation. It follows that the final pixel resolution is in the range from 0.90404μm to 1.0064μm, depending on the scanner used. This pixel resolution and the selected patch size provide a good compromise between the resolution at the cellular/glandular level, the amount of contextual information present in the patch (e.g. glandular arrangment) and the dataset size.

Models

---

For convinience, a collection of PyTorch pre-trained models is provided:

Model	Accuracy	Training Time x Epoch (s)✦
ResNeXt 50 32x4d✶ [PyTorch] [ONNX]	87.4%	251
ResNeXt 101 32x8d✶ [PyTorch] [ONNX]	87.4%	619
WideResNet 50✶ [PyTorch]	87.7%	252
WideResNet 101✶ [PyTorch]	87.7%	414
EfficientNet B0✷ [PyTorch]	87.4%	244
EfficientNet B1✷ [PyTorch]	87.7%	315
EfficientNet B2✷ [PyTorch]	88.2%	329

^✦ training on a NVIDIA® V100 Tensor Core GPU (batch size = 64)

^✶ implemented by pytorch

^✷ implemented by lukemelas

All models were trained using mixed precision and can be loaded with the standard procedure used in PyTorch:

import torch
from torchvision import models

!curl {S3_PATH} -o {LOCAL_PATH}

# Let's say we're loading a ResNeXt model

model = models.resnext_50_32x4d(pretrained=False, num_classes=4)
model.load_state_dict(torch.load(LOCAL_PATH))

Limitations and Future Work

---

Despite the dataset provides a good quality signal for transfer learning to related problems, it still contains noisy labels. My hope is to refine the Prostate Gleason Dataset to the point that it can be used as a benchmark for prostate-cancer-assessment algorithms.