**Introduction **¶
In this challenge, we publish a H&E-stained whole slide image dataset of prostatectomy and biopsy specimens with annotations performed by experienced pathologists. Additionally, we provide a set of images scanned by multiple scanners to assess the algorithm performance of handling variations caused by image digitalization. The submitted algorithm should not only be accurate enough to detect different Gleason Patterns but also generalized to process images scanned by different scanners. To the best of our knowledge, this is the first challenge in the field of digital pathology that investigate the variations caused by image scanning.
The dataset for this challenge is obtained from National University Hospital, Singapore. The challenge data is released under the creative commons license (CC BY-NC-SA 4.0). The release date of the training data is Apr 1, 2022. Only registered participants can download the data.
21/6/2022 - Please note that publicly available data is allowed to train the model, including (open) pre-trained nets (This information can be found here. This link is posted on MICCAI Registered Challenge). Private data is prohibited.
About A!HistoNotes¶
A!HistoNotes is a annotation platform built on AWS cloud. In this study, all annotations are performed by pathologists using A!HistoNotes.
Cohorts¶
Subset 1: Whole mount images scanned by a Akoya Biosciences scanner¶
Training set:105 cases
Test set: 45 cases
Subset 2: Biopsy images scanned by a Akoya Biosciences scanner
Training set: 37 cases
Test set: 16 cases
Subset 3: Whole mount images scanned by multiple scanners
Training set: each specimen is scanned by multiple scanners (Akoya Biosciences, Olympus, Zeiss, Leica, KFBio, Philips). Each scanner scanned 26 cases except 25 for Olympus and 15 for Zeiss. In total, 26*4+25*1 +15*1 = 144 cases.
Test set: each specimen is scanned by multiple scanners (Akoya Biosciences , Olympus, Zeiss, Leica, KFBio, Philips). Each scanner scanned 12 cases except 7 for Zeiss. In total, 12*5+7*1 = 67 cases.
The name of the scanner is specified in the filename of the image. For example, "Subset3_Train_20_Akoya.tiff" and "Subset3_Train_20_Leica.tiff" refer to glass slide "Subset3_Train_20" scanned by Akoya scanner and Leica scanner respectively, but there is no "Subset3_Train_20_Olympus.tiff" because glass slide "Subset3_Train_20" was not scanned using Olympus scanner.
Important information¶
1. In training set, each image comes with a set of binary masks of annotations ("Gleason Pattern3", "Gleason Pattern4", "Gleason Pattern5", "Normal", "Stroma") performed by pathologists. The number of binary masks varies from case to case and maximum is 5. For example, pathologists only annotated "Gleason Pattern3", "Gleason Pattern4" and "Normal" in some cases, so there are only 4 masks for those cases.
2. In some cases, there are overlapping region between masks of different classes, i.e. some pixels are assigned with more than one label. For example, as shown in the figure below, the annotations of Gleason Pattern 3 (G3) and Gleason Pattern 4 (G4) are partially overlapped. This happened when pathologists annotated regions where more than one patterns are mixed together. Participants shall process the binary masks according to their understanding.
3. The size of the pixel is 0.5 µm/pixel (20x) for images and annotations. All H&E images are converted from its original formats to " .tiff" using MATLAB.
4. Manual annotations are performed on images scanned by Akoya Biosciences only. For subset 3, images scanned by different scanners are not aligned with ones scanned by Akoya Biosciences. Therefore, we performed the image registration to transform the original masks to match with images scanned by other scanners.
**Target **¶
Detect different patterns in the images and generate 5 binary masks to indicate Stroma, Normal, Gleason Pattern 3, Gleason Pattern 4, Gleason Pattern 5, respectively. Mask can be empty if no tissue of that class is detected. Only fully automated methods are allowed.