Segment Anything Model for Instance Segmentation in Kidney Histopathology Images

Jurre Weijer, Sanne Truijen, Tri Nguyen, Mitko Veta, Nikolas Stathonikos

Keywords: Multi-Class Segmentation, Segment Anything Model, Nephrology, Kidney Pathology, Kidney Transplantation

Abstract:

Kidney transplantation offers a significant improvement in the quality of life for individuals
with irreversible kidney failure. Early detection of rejection through pathologists’ assessment
of kidney biopsies is critical to ensure long-term graft survival. Traditional assessment
methods rely on semi-quantitative estimations from a pathologist while implementing deep
learning models holds promise for providing more accurate measurements. Large annotated
datasets, required to train such models, are challenging to obtain for kidney tissue. In this
study, we fine-tune and modify the Segment Anything Model (SAM) to facilitate instance
segmentation on whole-slide imaging (WSI) data. Leveraging SAM’s zero-shot capability,
we accelerate dataset creation by automatically obtaining annotations which we refine and
label. We demonstrate promising results with limited annotated slides for training. Additionally,
our approach allows for iterative dataset expansion to enhance model performance
over time. Code is available at: https://github.com/JurreWeijer/SAM-Nephro.

Submission Number: 72