Deep learning-based classification of kidney transplant pathology: a retrospective, multicentre, proof-of-concept...

Area of Science:

• Computational pathology
• Artificial intelligence in medicine
• Transplant nephrology

Background:

• Histopathological assessment of kidney allograft biopsies is the standard for diagnosing rejection but is complex and expertise-intensive.
• Current methods require significant time and effort, highlighting a need for efficient diagnostic support systems.
• Developing automated systems can potentially streamline the diagnostic workflow for transplant biopsies.

Purpose of the Study:

• To analyze the utility of deep learning for preclassifying kidney allograft biopsy histology.
• To develop a potential biopsy triage system focused on identifying transplant rejection.
• To categorize biopsies into normal, rejection, or other disease categories using artificial intelligence.

Main Methods:

• A retrospective, multicenter study utilized 5844 digital whole slide images from 1948 kidney transplant recipients.
• Convolutional neural networks (CNNs) were trained to classify biopsies into normal, rejection, or other disease categories.
• Performance was validated using cross-validation and an external real-world cohort, with Area Under the Receiver Operating Characteristic Curve (AUROC) as the primary metric.

Main Results:

• Serial CNNs achieved high AUROC values for distinguishing normal from disease and for classifying disease subtypes (rejection vs. other diseases).
• A single CNN demonstrated good performance in cross-validation and generalized well for normal and rejection classes on real-world data.
• Visualisation techniques successfully highlighted rejection-relevant areas within the tubulointerstitium of biopsies.

Conclusions:

• Deep learning-based classification of transplant biopsies can effectively support pathological diagnostics.
• AI tools show potential for improving the efficiency and accuracy of identifying kidney allograft rejection.
• This approach could serve as a valuable biopsy triage system, assisting in patient management.