Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Robust Alzheimer's Progression Modeling using Cross-Domain Self-Supervised Deep Learning

Saba Dadsetan, Mohsen Hejrati, Shandong Wu, Somaye Hashemifar

TL;DR

This work tackles the problem of predicting Alzheimer's disease progression from MRI when labeled data are limited. It introduces a cross-domain self-supervised pretraining framework that combines natural-image SSL with in-domain medical-image SSL to learn transferable representations for regression of the CDR-SB score at 12 months. The results show that cross-domain pretraining, especially Unlabeled_ImageNet -> In-domain, yields the strongest performance and generalizes across multiple cohorts, with SimCLR in-domain pretraining and larger unlabeled MRI datasets providing additional gains. The findings highlight the potential of leveraging unlabeled data from diverse domains to improve prognostic accuracy and robustness in clinical settings, with GradCAM analyses supporting domain-relevant focus in brain regions affected by early AD.

Abstract

Developing successful artificial intelligence systems in practice depends on both robust deep learning models and large, high-quality data. However, acquiring and labeling data can be prohibitively expensive and time-consuming in many real-world applications, such as clinical disease models. Self-supervised learning has demonstrated great potential in increasing model accuracy and robustness in small data regimes. In addition, many clinical imaging and disease modeling applications rely heavily on regression of continuous quantities. However, the applicability of self-supervised learning for these medical-imaging regression tasks has not been extensively studied. In this study, we develop a cross-domain self-supervised learning approach for disease prognostic modeling as a regression problem using medical images as input. We demonstrate that self-supervised pretraining can improve the prediction of Alzheimer's Disease progression from brain MRI. We also show that pretraining on extended (but not labeled) brain MRI data outperforms pretraining on natural images. We further observe that the highest performance is achieved when both natural images and extended brain-MRI data are used for pretraining.

Robust Alzheimer's Progression Modeling using Cross-Domain Self-Supervised Deep Learning

TL;DR

This work tackles the problem of predicting Alzheimer's disease progression from MRI when labeled data are limited. It introduces a cross-domain self-supervised pretraining framework that combines natural-image SSL with in-domain medical-image SSL to learn transferable representations for regression of the CDR-SB score at 12 months. The results show that cross-domain pretraining, especially Unlabeled_ImageNet -> In-domain, yields the strongest performance and generalizes across multiple cohorts, with SimCLR in-domain pretraining and larger unlabeled MRI datasets providing additional gains. The findings highlight the potential of leveraging unlabeled data from diverse domains to improve prognostic accuracy and robustness in clinical settings, with GradCAM analyses supporting domain-relevant focus in brain regions affected by early AD.

Abstract

Developing successful artificial intelligence systems in practice depends on both robust deep learning models and large, high-quality data. However, acquiring and labeling data can be prohibitively expensive and time-consuming in many real-world applications, such as clinical disease models. Self-supervised learning has demonstrated great potential in increasing model accuracy and robustness in small data regimes. In addition, many clinical imaging and disease modeling applications rely heavily on regression of continuous quantities. However, the applicability of self-supervised learning for these medical-imaging regression tasks has not been extensively studied. In this study, we develop a cross-domain self-supervised learning approach for disease prognostic modeling as a regression problem using medical images as input. We demonstrate that self-supervised pretraining can improve the prediction of Alzheimer's Disease progression from brain MRI. We also show that pretraining on extended (but not labeled) brain MRI data outperforms pretraining on natural images. We further observe that the highest performance is achieved when both natural images and extended brain-MRI data are used for pretraining.
Paper Structure (13 sections, 1 equation, 3 figures, 6 tables)

This paper contains 13 sections, 1 equation, 3 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Related works
  3. Methodology
  4. Task and Data
  5. Self-supervised learning platform for progression prediction task
  6. Results
  7. Self-Supervised Pretraining
  8. In- and out-of-Domain Generalization
  9. Visualizing Model Saliency Maps
  10. Discussion
  11. Appendix
  12. Self-supervised Models
  13. MSE values corresponding to Table 4b

Figures (3)

  • Figure 1: Different approaches for self-supervised pretraining on in-domain medical imaging, including (a) random initialization, (b) supervised ImageNet initialization, and (c) self-supervised ImageNet initialization. (d) Performing fine-tuning by transferring the backbone from one of the scenarios a-c. (e) utilization of the trained model on unseen test sets.
  • Figure 2: Average frequency distribution. Each color represents the frequency distribution of residuals for a specific experiment.
  • Figure 3: Illustrating the interpretation of three pretraining models using GradCam technique. The top row showcases the original MRI slices, while the subsequent rows, from top to bottom, illustrate the saliency maps generated by the following models: a randomly-initialized pretrained model, a pretrained model on natural images, and our best model, Barlow Twins $\rightarrow$ SimCLR.