Summary
Graft-versus-host disease (GVHD) is a complication that can occur after a stem cell or bone marrow transplant, where the donor’s immune cells attack the recipient’s body, particularly affecting the skin, liver, and gastrointestinal tract.
In this study, regions of interest (ROIs) were extracted from pathological images using the QuPath application. The ROIs were selected based on specific blue-marked areas in the images, and each ROI was labeled with varying sizes and shapes to enhance data diversity. Approximately 40-50 ROIs were extracted per image, with focus on 128×128 and 256×256 pixel areas. These ROIs were labeled into three categories: Tubular Adenoma, GVHD, and None.
Feature extraction was carried out using the ‘Prov-Gigapath’ foundational model, resulting in a dataset with 728 ROIs and 1536 feature columns. The classification task aimed to determine whether the extracted features could effectively distinguish between the three classes.
Dataset
The ROIs are labelled in three class identifiers; 0. Tubular Adenoma, 1. GVHD, 2. None
The ROIs with ‘None’ labels are selected randomly based on the non-overlapping coordinates that exclude GVHD and Tubular Adenoma coordinates.
Feature extraction was performed using the ‘Prov-Gigapath’ foundational model. This model is capable of generating a feature vector with a shape of 1×1536 floats for each ROI. A total of 728 different ROIs were extracted, resulting in a CSV file containing 728 rows and 1536 columns of features.
Performance
The classification task utilized in our study aims to see that whether generated features are distinguishable among the other classes. As it is mentioned that there are 3 different classes which are labelled regarding the coordinates of ROIs. This label information also added as a separate column in our CSV file as ‘label’ column.
The classification experiment conducted on the random forest algorithm shows that the AUC (area under the curve), which represents the model’s ability to distinguish between classes, demonstrates the model’s performance. An AUC value closer to 1 indicates a better performing model, while a value closer to 0.5 suggests a model with no discriminative ability.
| dataset | model | test_auc | test_acc | test_precision | test_recall | test_kappa | cvt_auc | cvt_acc | cvt_precision | cvt_recall |
| gvhd_128.csv | randomforest | 0.982 | 0.877 | 0.9 | 0.854 | 0.796 | 0.944 | 0.86 | 0.874 | 0.849 |
You can also see the additional classification results with the image below:
