Graph Convolutional Neural Network-based Bone Marrow Classification

Precise delineation and safeguarding of the functional bone marrow (FBM) in the process of radiotherapy planning have the potential to mitigate pelvic hematological toxicity. Radiotherapy is the primary therapeutic option for numerous pelvic cancers; however, there is a possibility that high-intensity radiation might cause harm to the FBM. The traditional approach to delineate the FBM is both time-intensive and demanding in terms of labour, making an efficient and precise bone marrow identification a valuable asset in clinical settings. Graph convolutional neural networks (GNNs) have gained popularity for semi-supervised classification, particularly in bone marrow image analysis. Prior GNN-based approaches primarily utilized a single graph neural network or graph filter to extract bone marrow characteristics, failing to exploit the potential of GNNs. Classical GNNs also face issues of oversmoothing. To address these limitations, we propose a hybrid method, the GNN-GS combination, which leverages both GNN and GraphSAGE. By employing performance criteria such as accuracy (ACC), precision (PRE), and the F1-score, we conducted a comparative evaluation of our suggested technique against other methods deemed to be as advanced as possible. The outcomes reveal a notable enhancement in ACC, PRE, and F1-score by 2.32%, 6.47%, and 2.75%, respectively, for the specified dataset. These comparative results establish the suitability of our proposed method for addressing bone marrow classification challenges.