Project
New Frontiers in Kidney Digital Pathology Diagnostics – Developing Foundation Models for Kidney Histopathology
Recent technological developments enable digitization of pathology by scanning tissue samples with specialized high-resolution scanners. This results in so-called Whole Slide Images (WSI), currently by far the largest data generated in medicine. This digital workflow enables the effective use of AI and particularly deep learning (DL) approaches to automate and improve pathology diagnostic. The most recent development in the DL domain is the training of foundation models. These offer significant advancement in performance across various tasks due to their high generalizability. However, the existing foundation models for digital pathology either completely ignored or used only rudimentary data from kidney pathology. Therefore, the applicability of current foundation models on kidney diseases is limited.
We aim to develop, test and validate the first foundation models for comprehensive assessment of kidney pathology. Therefore, we collected a large, multicenter, international dataset of kidney histology of more than 200,000 WSIs. Due to the huge amount of training data and the computational complexity of the DL algorithms, advanced, high performance computational resources are required.
Project Details
Project term
April 1, 2025–July 1, 2026
Affiliations
Uniklinik RWTH Aachen
Institute
Institut für Pathologie
Principal Investigator
Methods
We use the popular DINOv2 framework, a state of the art framework for self-supervised model training, for our deep learning networks. We train several variants on different versions of our dataset. Therefore, we organize the data in a modular fashion using HDF5 containers to reduce the number of files and maintain clarity. We train Vision Transformer (ViT) models of different sizes and a FlexiViT. The computational complexity of the FlexiViT can be adjusted for the task to ensure efficiency for simple tasks and better performance for complex tasks. We evaluate the models on several medically challenging private and public testsets including publicly available datasets.
Results
So far, we established the workflow and confirmed that the training process is stable for all model configurations. We collected and prepared all data, resulting in more than 700.000.000 images for model training. The data includes images from multiple medical centers, different histological staining protocols and various scanners and magnifications. A data curation pipeline was estabilished to automatically filter blurry and redundant patches. We implemented and tested the training pipeline for large and giant ViT models. Then, we trained a preliminary model on a subset of the total data and evaluated it on several testsets covering multiple categories, including classification and segmentation tasks.
Discussion
Collecting, transfering and cleaning the huge amounts of data was highly complex and a major challenge of this work. Balancing the dataset between different histological strainings, medical centers, scanner protocolls was challenging. Preparing appropriate testsets to evaluate the model and compare to existing models was requiring lots of effort. While we already observe good performance with the preliminary model, we expect it to further increase with adding more heterogenous training data. However, a lot of work still remains, including training and evaluation of the final models. Additionally, applying the models in an efficient way for daily use requires some future work.
Additional Project Information
DFG classification: 205-06 Pathology
Software: Kidney Foundation Model
Cluster: CLAIX