Graph Echo State Network for MRI-based Tendon Pathology Classification

Background and Objective:

Diagnosing tendon-related pathologies is crucial for early diagnosis and effective treatment planning, enabling timely interventions that can significantly improve patient outcomes.

Methods:

This study presents an end-to-end tendon pathology detection (classification) module utilizing a custom ankle MRI dataset comprising 76 subjects (45 healthy, 31 pathological). We propose a graph-based module that converts images into graph representations for classification. Superpixels are first generated by grouping pixels with similar intensity values, serving as the graph’s nodes, while edges connect neighboring superpixels to establish the graph structure. Next, a Graph Echo State Network (GESN) is employed for classification, leveraging its echo state property that eliminates the need for iterative backpropagation. This property produces rich graph embeddings which are fed into a linear readout layer, where weights and biases are learned using ridge regression with regularization. For baseline comparison, a non-graph-based module extracts radiomic features from both the entire image and individual superpixels, employing four traditional machine learning classifiers. We analyze and compare the performance of the graph-based and non-graph-based modules, using majority voting on slice-level predictions to generate subject-wise predictions.

Results:

Our baseline non-graph-based module achieved relatively better performance by extracting global radiomic features from entire images compared to local features derived from superpixels. In contrast, our graph-based module effectively integrates both local and global perspectives. The graph embeddings produced by the Graph Echo State Network (GESN) enhance data representation, resulting in a mean accuracy of 0.953 ± 0.013 and a mean sensitivity of 0.943 ± 0.035, both significantly surpassing the baseline performance. Additionally, hyperparameters such as the reservoir spectral radius and scaling factors had a notable impact on the outcomes of the graph-based classification.

Conclusions:

Our findings highlight the effectiveness of graph-based models, particularly GESN, in capturing meaningful representations and improving classification performance in tendon pathology detection.