[Paper Review] Self-supervised Learning for Segmentation and Quantification of Dopamine Neurons in Parkinson's Disease

Fatemeh Haghighi, Soumitra Ghosh+6 more|arXiv (Cornell University)|Jan 11, 2023

Cell Image Analysis TechniquesBiochemistry, Genetics and Molecular Biology3 citations

TL;DR

This paper proposes a self-supervised deep learning framework for automated segmentation and quantification of dopaminergic neurons in Parkinson’s disease mouse models, leveraging cross-domain pretraining on natural and pathology images to overcome limited annotated data. The method achieves high-precision neuron counting and individual cell characterization, including tyrosine hydroxylase (TH) intensity, enabling unbiased, high-throughput analysis critical for preclinical drug efficacy studies.

ABSTRACT

Parkinson's Disease (PD) is the second most common neurodegenerative disease in humans. PD is characterized by the gradual loss of dopaminergic neurons in the Substantia Nigra (SN). Counting the number of dopaminergic neurons in the SN is one of the most important indexes in evaluating drug efficacy in PD animal models. Currently, analyzing and quantifying dopaminergic neurons is conducted manually by experts through analysis of digital pathology images which is laborious, time-consuming, and highly subjective. As such, a reliable and unbiased automated system is demanded for the quantification of dopaminergic neurons in digital pathology images. Recent years have seen a surge in adopting deep learning solutions in medical image processing. However, developing high-performing deep learning models hinges on the availability of large-scale, high-quality annotated data, which can be expensive to acquire, especially in applications like digital pathology image analysis. To this end, we propose an end-to-end deep learning framework based on self-supervised learning for the segmentation and quantification of dopaminergic neurons in PD animal models. To the best of our knowledge, this is the first deep learning model that detects the cell body of dopaminergic neurons, counts the number of dopaminergic neurons, and provides characteristics of individual dopaminergic neurons as a numerical output. Extensive experiments demonstrate the effectiveness of our model in quantifying neurons with high precision, which can provide a faster turnaround for drug efficacy studies, better understanding of dopaminergic neuronal health status, and unbiased results in PD pre-clinical research. As part of our contributions, we also provide the first publicly available dataset of histology digital images along with expert annotations for the segmentation of TH-positive DA neuronal soma.

Motivation & Objective

• To address the lack of reliable, automated methods for quantifying dopaminergic neurons in Parkinson’s disease preclinical research.
• To overcome the challenge of limited annotated histology data for training deep learning models in digital pathology.
• To develop a model that not only counts dopaminergic neurons but also extracts individual cell characteristics such as TH intensity for assessing neuronal health.
• To enable faster, more objective, and scalable analysis of drug efficacy in animal models of Parkinson’s disease.
• To establish the first publicly available dataset of TH-positive dopaminergic neuron soma annotations in mouse brain sections.

Proposed method

• Utilizes cross-domain self-supervised learning pretraining on large-scale unlabeled natural images and in-domain pathology images to learn robust feature representations.
• Fine-tunes the self-supervised backbone using only limited annotated data for the specific task of dopaminergic neuron segmentation.
• Employs a U-Net architecture as the segmentation head due to its superior performance on medical image segmentation tasks.
• Applies a comprehensive data augmentation strategy including Flip, Rotation, RGBShift, Blur, GaussianNoise, and RandomResizedCrop to improve generalization.
• Introduces a practical counting pipeline that identifies and counts individual neuron cell bodies from the segmentation masks.
• Uses TH intensity within segmented cell bodies as a quantitative biomarker for assessing dopaminergic neuronal health status.

Experimental results

Research questions

• RQ1 Can self-supervised pretraining on diverse image domains improve segmentation accuracy for dopaminergic neurons with limited labeled data?
• RQ2 How does the performance of the proposed self-supervised framework compare to random initialization, ImageNet-supervised fine-tuning, and domain-specific self-supervised pretraining?
• RQ3 To what extent can the model extract biologically meaningful features such as TH intensity from individual neurons?
• RQ4 Can the model achieve high-precision neuron counting and segmentation in complex histological images with overlapping cells and variable staining?
• RQ5 Is the model generalizable across different animal study cohorts and independent validation sets?

Key findings

• The proposed self-supervised framework outperformed training from scratch, fine-tuning ImageNet-supervised models, and domain-specific self-supervised pretraining in segmentation accuracy.
• The combination of data augmentation techniques including RGBShift, Blur, and GaussianNoise significantly improved model generalization and performance.
• RandomResizedCrop contributed positively to performance, while elastic transformations had a negative impact, indicating sensitivity to augmentation design.
• U-Net architecture demonstrated superior segmentation performance compared to DeepLabV3+ on the neuron segmentation task.
• The model achieved high-precision counting of dopaminergic neurons, enabling faster and more objective analysis than manual methods.
• The model successfully extracted TH intensity as a quantitative measure of neuronal health, providing insights into disease progression and drug response.