TY - JOUR T1 - Automatic Sample Segmentation & Detection of Parkinson’s Disease Using Synthetic Staining & Deep Learning JF - bioRxiv DO - 10.1101/2022.08.30.505459 SP - 2022.08.30.505459 AU - Bradley Pearce AU - Peter Coetzee AU - Duncan Rowland AU - David T Dexter AU - Djordje Gveric AU - Stephen Gentleman Y1 - 2022/01/01 UR - http://biorxiv.org/content/early/2022/09/01/2022.08.30.505459.abstract N2 - The identification of Parkinson’s Disease (PD) from post-mortem brain slices is time consuming for highly trained neuropathologists, often taking many hours per case. In this study, we demonstrate fully automatic PD detection, from single 1000um regions, from sections spanning from the dorsal motor nucleus of the vagus nerve to the frontal cortex. This is achieved via image processing and statistical methods, with improved accuracy demonstrated when using machine learning. Digitised stained brain sections were processed via a deep neural network to produce re-coloured, or ‘synthetically stained’, images which were then filtered and passed to a secondary network for classification. We demonstrate state-of-the-art PD detection (>90% accuracy on single 1000um regions), with the ability to perform binary classification on high resolution sections within minutes, in addition to demarcating regions of interest to the pathologist for manual visual verification.Executive Summary The identification of Parkinson’s Disease (PD) from post-mortem brain slices is time consuming for highly trained neuropathologists, often taking many hours per case. Accurate classification and stratification of PD is critical for the confirmation that the brain donor suffered from PD and to maximise the potential usefulness of the brain in research studies to better understand the causes of PD and foster drug development.Parkinson’s UK Brain Bank, at Imperial College London, has produced a dataset containing digitised images of brain sections immunostained for the protein alpha-synuclein (α-syn), the pathological marker of PD; along with control cases from healthy donors. This dataset is much larger (over 400 cases), more consistent, and of higher quality (all have been stained with the same protocol and imaged within the same laboratory) than has been documented elsewhere in the literature; including those found in a meta-analysis study on detection of neurological disorders containing over 200 papers (Lima et al., 2022).The project team, consisting of neuroscientists and subject matter experts from: Imperial, NHS AI Lab Skunkworks, Parkinson’s UK, and Polygeist have undertaken a 12 week project to examine the possibility of producing a Proof-of-Concept (PoC) tool to automatically load, enhance and ultimately classify those brain sections containing α-syn. The initial focus of the project was to make a tool that could make a biomarker of PD, α-syn, more visible to the pathologist; saving time in searching for the protein manually. This goal was quickly reached, producing a tool that could ‘synthetically stain’ the α-syn, marking regions of interest in a high-contrast bright green, making them quickly identifiable for the pathologist. Statistical analysis of the synthetically stained images showed that very few regions in the control group were stained compared to the PD group, raising the possibility that an automatic classifier could be developed, which became a stretch goal for the project.A bespoke neural network model was designed that processed the synthetically stained segments of each immunostained section and produced a binary judgement (whether a segment contains PD pathology or not). The model achieved >90% sensitivity for PD detection, much higher than is reported for neuropathologists (~60% sensitivity when searching for α-syn patches across all stages, Signaevsky et al., 2022). While expert raters are still more precise (~6% better than the model), the model performed ~20% better than expert raters when considering precision and recall.The key output of the project is an open-source PoC tool that can automatically classify PD from digitised images of brain sections with accuracy that is approaching viability for real world applications. An MIT Licensed code repository has been released, containing all of the model development code, along with associated documentation, to allow others to build on the project team’s work. This report summarises the scientific and engineering process undertaken through the development of the PoC tool.Competing Interest StatementThe authors have declared no competing interest. ER -