Ensemble Object Detection Methodology for Automated Detection of Inflammatory Cells in Kidney Biopsies

Gunjan Deotale, Abhishek Ambast, Lavish Ramchandani, Dev Kumar Das, Tijo Thomas

Keywords: Kidney Biopsy, Inflammatory Cell, Object Detection, Digital Pathology

TL;DR: This short paper presents an approach for Automated Detection of Inflammatory Cells in Kidney Biopsies which secured 3rd place in the Monkey Grand Challenge.

Abstract:

Automated detection of inflammatory cells in kidney biopsies is essential for kidney disease diagnosis. To address this, we participated in the Machine-learning for Optimal detection of iNflammatory cells in KidnEY (MONKEY) challenge, where the main challenges were
to detect inflammatory cells and further classify them as monocyte and lymphocyte. We employed an ensemble of DETR and YOLOv5-L object detection models, achieving the 3rd place on both leaderboards with Free Response Receiver Operating Characteristic (FROC)
scores of 0.3517 (Task 1) and 0.4471/0.1906 (Task 2). Our approach demonstrated the power of combining transformer-based and convolutional architectures to enhance diagnostic precision in digital pathology, offering a cost-effective alternative to immunohistochemistry (IHC) staining while advancing transplant rejection analysis.

Submission Number: 23