Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Predicting Knee Osteoarthritis Progression from Structural MRI using Deep Learning

Egor Panfilov, Simo Saarakkala, Miika T. Nieminen, Aleksei Tiulpin

TL;DR

This work tackles predicting knee OA progression from structural MRI using an end-to-end deep learning approach. It proposes a hybrid CNN-Transformer architecture that processes slice-wise DESS MRI with a 2D CNN encoder and aggregates across slices via a Transformer to predict progression within 96 months. On the OAI dataset, the 2D CNN + Transformer configuration achieves the best performance (AP 0.58 and ROC AUC 0.79), outperforming 2D and 3D baselines and establishing a baseline for MRI-based KOA progression prediction. The method advances disease understanding and trial design potential, with code and data sources described to enable translation and further research.

Abstract

Accurate prediction of knee osteoarthritis (KOA) progression from structural MRI has a potential to enhance disease understanding and support clinical trials. Prior art focused on manually designed imaging biomarkers, which may not fully exploit all disease-related information present in MRI scan. In contrast, our method learns relevant representations from raw data end-to-end using Deep Learning, and uses them for progression prediction. The method employs a 2D CNN to process the data slice-wise and aggregate the extracted features using a Transformer. Evaluated on a large cohort (n=4,866), the proposed method outperforms conventional 2D and 3D CNN-based models and achieves average precision of $0.58\pm0.03$ and ROC AUC of $0.78\pm0.01$. This paper sets a baseline on end-to-end KOA progression prediction from structural MRI. Our code is publicly available at https://github.com/MIPT-Oulu/OAProgressionMR.

Predicting Knee Osteoarthritis Progression from Structural MRI using Deep Learning

TL;DR

This work tackles predicting knee OA progression from structural MRI using an end-to-end deep learning approach. It proposes a hybrid CNN-Transformer architecture that processes slice-wise DESS MRI with a 2D CNN encoder and aggregates across slices via a Transformer to predict progression within 96 months. On the OAI dataset, the 2D CNN + Transformer configuration achieves the best performance (AP 0.58 and ROC AUC 0.79), outperforming 2D and 3D baselines and establishing a baseline for MRI-based KOA progression prediction. The method advances disease understanding and trial design potential, with code and data sources described to enable translation and further research.

Abstract

Accurate prediction of knee osteoarthritis (KOA) progression from structural MRI has a potential to enhance disease understanding and support clinical trials. Prior art focused on manually designed imaging biomarkers, which may not fully exploit all disease-related information present in MRI scan. In contrast, our method learns relevant representations from raw data end-to-end using Deep Learning, and uses them for progression prediction. The method employs a 2D CNN to process the data slice-wise and aggregate the extracted features using a Transformer. Evaluated on a large cohort (n=4,866), the proposed method outperforms conventional 2D and 3D CNN-based models and achieves average precision of and ROC AUC of . This paper sets a baseline on end-to-end KOA progression prediction from structural MRI. Our code is publicly available at https://github.com/MIPT-Oulu/OAProgressionMR.
Paper Structure (14 sections, 3 figures, 2 tables)

This paper contains 14 sections, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Method
  3. Overview
  4. Definition of Osteoarthritis Progression
  5. Architecture
  6. Reference Methods
  7. Experiments
  8. Data
  9. Implementation Details
  10. Metrics
  11. Results
  12. Conclusion
  13. Compliance with ethical standards
  14. Acknowledgments

Figures (3)

  • Figure 1: Overview of the proposed end-to-end system for OA progression prediction from volumetric knee MRI scan. The scan is shown by evenly sampled sagittal slices
  • Figure 2: Schematic representation of (a) the proposed model based on 2D CNN and Transformer, (b, c) the 2D models with conventional feature aggregation via FC and Bi-LSTM, respectively, and (d) the volumetric models. In (a), slice-wise features $s_1, ..., s_k$ are projected into embeddings $e_1, ... e_k$, combined with positional $p_1, ..., p_k$ and class $c$ tokens, and processed in Transformer. $o_{f,k}$ and $o_{b,0}$ are the terminal states of Bi-LSTM output, in forward and backward passes, respectively. FE - feature extractor, GAP - global average pooling, FC - fully connected layer
  • Figure 3: (a) Confusion matrix for the best model (before pooling the "slow" and "fast" classes). (b) ROC and (c) Precision-Recall curves for the top performing 2D and 3D models.