Dr Daniel Christiadi1, Dr Giles Walters1,2,4, Dr Krishna Karpe2,4, Dr Alice Kennard2,4, Dr Richard Singer2,4, Associate Professor Girish Talaulikar2,4, Dr Mitali Fadia3,4, Dr Simon Jiang1,2,5
1Department of Immunology and Infectious Disease, John Curtin School of Medical Research, The Australian National University, Acton, Australia, 2Department of Renal Medicine, The Canberra Hospital, Garran, Australia, 3Department of Anatomical Pathology, ACT Pathology, The Canberra Hospital, Garran, Australia, 4Medical School, Australian National University College of Health and Medicine, Canberra, Australia, 5Centre of Personalised Medicine, The Australian National University, Acton, Australia
Aim
To develop a deep learning (DL) algorithm capable of identifying structures on kidney biopsy whole slide images.
Introduction
Kidney biopsy remains an essential tool in kidney disease diagnosis. However, biopsy interpretation is limited by inter-pathologist reliability and the inability to accurately quantify structures and pathology across a whole kidney sample.
Methods
Healthy native kidney biopsy slides were scanned with Zeiss AxioScan. Randomly extracted image patches were then annotated by six nephrologists and cross-examined by a nephropathologist using Automated Slide Analysis Platform annotation software. A Detectron2 model was trained using the annotated patches, and its performance was analysed on testing image patches.
Results
A total of 152 image patches from 11 patients were annotated. The proximal tubule (PT), distal tubule (DT), and glomeruli (G) identification accuracy were 95.46%, 16.4%, and 95%. The positive predictive value were 98.87%, 75%, and 95%, respectively. Due to the limited number in the training set, the algorithm failed to identify 2 of 2 arterioles (A).
The mean average precision (AP) was 20.99, driven by high AP in PT and G segmentation (47.26 and 34.33).
The platform had been designed to allow a human re-annotation in the DL predicted annotation, allowing iterative, improved annotation speed.
Conclusions
DL segmentation of kidney structures using a deep learning algorithm provides the foundation of automatic feature extraction from kidney biopsy images. The process will potentially create input data for predictive models that enhance patient management.
Biography:
Daniel Christiadi completed his medical degree from Airlangga University, Indonesia. In 2011, Daniel started the Basic Physician Training in Northwest Regional Hospital, Burnie. He then completed specialist training in Nephrology at The Canberra, Royal Darwin and Imperial College (UK) Hospitals.
Following Acute Kidney Injury Fellow at Prince of Wales Hospital, Sydney, in February 2021, Daniel started PhD at John Curtin School of Medical Research, Australian National University, studying the artificial intelligence application to improve the diagnosis and prediction of patients with kidney disease.