Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Evaluating unsupervised contrastive learning framework for MRI sequences classification

Yuli Wang, Kritika Iyer, Sep Farhand, Yoshihisa Shinagawa

TL;DR

The paper tackles the challenge of automatically identifying MRI sequence types across nonstandardized protocols to support clinical workflows. It proposes an unsupervised contrastive pre-training pipeline using SimCLR or SimSiam with a ResNet-18 backbone on 2D MRI slices, followed by supervised fine-tuning for nine sequences. On a large in-house brain MRI dataset and multiple public datasets (BraTS, ADNI, Organs), the method achieves an accuracy exceeding $0.95$ on the nine-type task and demonstrates strong generalization across domains, though external datasets show some drop when transferring. The work suggests that unsupervised pre-training reduces labeling burden and enhances robustness to protocol variability, with practical implications for scalable MRI sequence identification in clinical practice.

Abstract

The automatic identification of Magnetic Resonance Imaging (MRI) sequences can streamline clinical workflows by reducing the time radiologists spend manually sorting and identifying sequences, thereby enabling faster diagnosis and treatment planning for patients. However, the lack of standardization in the parameters of MRI scans poses challenges for automated systems and complicates the generation and utilization of datasets for machine learning research. To address this issue, we propose a system for MRI sequence identification using an unsupervised contrastive deep learning framework. By training a convolutional neural network based on the ResNet-18 architecture, our system classifies nine common MRI sequence types as a 9-class classification problem. The network was trained using an in-house internal dataset and validated on several public datasets, including BraTS, ADNI, Fused Radiology-Pathology Prostate Dataset, the Breast Cancer Dataset (ACRIN), among others, encompassing diverse acquisition protocols and requiring only 2D slices for training. Our system achieves a classification accuracy of over 0.95 across the nine most common MRI sequence types.

Evaluating unsupervised contrastive learning framework for MRI sequences classification

TL;DR

The paper tackles the challenge of automatically identifying MRI sequence types across nonstandardized protocols to support clinical workflows. It proposes an unsupervised contrastive pre-training pipeline using SimCLR or SimSiam with a ResNet-18 backbone on 2D MRI slices, followed by supervised fine-tuning for nine sequences. On a large in-house brain MRI dataset and multiple public datasets (BraTS, ADNI, Organs), the method achieves an accuracy exceeding on the nine-type task and demonstrates strong generalization across domains, though external datasets show some drop when transferring. The work suggests that unsupervised pre-training reduces labeling burden and enhances robustness to protocol variability, with practical implications for scalable MRI sequence identification in clinical practice.

Abstract

The automatic identification of Magnetic Resonance Imaging (MRI) sequences can streamline clinical workflows by reducing the time radiologists spend manually sorting and identifying sequences, thereby enabling faster diagnosis and treatment planning for patients. However, the lack of standardization in the parameters of MRI scans poses challenges for automated systems and complicates the generation and utilization of datasets for machine learning research. To address this issue, we propose a system for MRI sequence identification using an unsupervised contrastive deep learning framework. By training a convolutional neural network based on the ResNet-18 architecture, our system classifies nine common MRI sequence types as a 9-class classification problem. The network was trained using an in-house internal dataset and validated on several public datasets, including BraTS, ADNI, Fused Radiology-Pathology Prostate Dataset, the Breast Cancer Dataset (ACRIN), among others, encompassing diverse acquisition protocols and requiring only 2D slices for training. Our system achieves a classification accuracy of over 0.95 across the nine most common MRI sequence types.
Paper Structure (9 sections, 3 figures, 3 tables)

This paper contains 9 sections, 3 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Dataset and Methods
  3. Dataset Curation
  4. Pre-training: Unsupervised Contrastive Learning
  5. Fine-tuning: Supervised Learning
  6. Experiments and Results
  7. Downstream Classification: In-house Dataset
  8. Downstream Classification: External Datasets
  9. Discussion and Conclusion

Figures (3)

  • Figure 1: The pipeline of our framework includes (a) unlabeled data from various MRI sequences; (b) an unsupervised pre-training stage using either the SimCLR or SimSiam framework; and (c) a fine-tuning stage applied to downstream tasks on different datasets, including BraTS, ADNI, and the Multi-Organs dataset.
  • Figure 2: Example of latent space visualizations from SimCLR during pre-training: (a) at epoch 0, (b) at epoch 50. Additionally, visualizations are shown after pre-training and supervised fine-tuning using different percentages of labeled data: (c) 0.5%, (d) 1%, (e) 5%, and (f) 50%.
  • Figure 3: MR sequence classification classification Performance of model on external datasets using varying percentage data for fine-tuning.