Dr Daniel Christiadi1,2,3, Dr Benjamin Mashford1, Dr Giles Walters1,2,3,5, Dr Krishna Karpe2,5, Dr Alice Kennard2,5, Dr Richard Singer2,5, A/Prof Girish Talaulikar2,5, A/Prof Mitali Fadia4,5, A/Prof Simon Jiang1,2,3
1Department of Immunology and Infectious Disease, John Curtin School of Medical Research, Australian National University, Acton, Australia, 2Department of Renal Medicine, Canberra Hospital, Garran, Australia, 3Centre for Personalised Medicine, Australian National University and Canberra Health Services, Canberra, Australia, 4Department of Anatomical Pathology, Canberra Health Services, Canberra, Australia, 5School of Medicine and Psychology, Australian National University, Canberra, Australia
Biography:
Daniel Christiadi is a nephrologist at Canberra Hospital. He is completing his PhD at John Curtin School of Medical Research in the application of Machine Learning and Deep Learning for management of patients with kidney disease
Background:
Glomerulonephritis (GN) is the second most common cause of end-stage kidney disease (ESKD). Diagnosis primarily relies on the histopathological examination of biopsy samples by trained anatomical pathologists. However, this method is labour-intensive, susceptible to interobserver variability, and highly qualitative in nature. Quantitative measurement of kidney histopathology will enable greater insight into the underlying disease.
Methods:
Kidney biopsy slides were digitised to create whole-slide images (WSIs) using an Axio Scan at 40x magnification with a resolution of 0.11 micrometer/pixel. A pipeline was created using several trained DL models to extract morphological information from WSIs at different magnifications.
Results:
A total of 99 normal kidney biopsies and 335 GN slides (IgA nephropathy, ANCA-associated vasculitis, minimal change disease, membranous nephropathy, focal segmental glomerulosclerosis, and class 3/4 lupus nephritis) were digitised. Utilising 120 normal and GN slides, we trained several DL models to construct KidneyVision. KidneyVision can segment the kidney cortex with an intersection-over-union (IoU) of 94% with an accuracy of 99%, sensitivity of 97%, and positive predictive value of 96%, identifying arteries, tubules, and glomerular structures with high fidelity. KidneyVision extracted morphometric data, such as glomerular size and tubular thickness which correlated with kidney function and demonstrated diagnostic capability.
Conclusions:
We developed KidneyVision, an automated DL pipeline that can extract quantitative morphological information from digitised GN kidney biopsy slides, assisting with histopathologic diagnosis.”
Presentation Slides PDF – Click Here