Synergizing Convolutional Backbones with Inductive kNN-Residual Graph Neural Networks for Preoperative Diagnosis of MTM-HCC

Abstract

AbstractPurpose: Macrotrabecular-massive hepatocellular carcinoma (MTM-HCC) is an aggressive subtype of liver cancer with limited imaging signatures, making accurate noninvasive diagnosis challenging. This study introduces a novel residual network 101 based k-nearest neighbors (kNN) graph assisted Residual Graph Convolution Network (ResNet101-kNN-ResGCN) designed for the preoperative classification of macrotrabecular-massive hepatocellular carcinoma (MTM-HCC) using multicenter CT data, with the goal of investigating whether leveraging local structural relationships among image features could improve diagnostic accuracy. Materials and Methods: This retrospective study included 503 patients from two institutions, comprising an internal cohort of 372 patients (87 MTM-HCC and 285 non-MTM-HCC) and an external cohort of 131 patients (34 MTM-HCC and 97 non-MTM-HCC). Two-dimensional slices from portal venous phase contrast-enhanced CT scans were used. Images were manually segmented by experienced radiologists to define regions of interest. The internal cohort was split into training and validation sets of 85% and 15%, with the external cohort held as an independent test set. An ResNet101-kNN-ResGCN architecture that combines convolutional feature extraction with graph-based relational modeling was designed. Deep feature are extracted using a ResNet101 backbone, followed by the dynamic construction of an inductive kNN graph within each batch in the feature space. This batch-wise graph encodes local feature similarity among samples within each mini-batch and serves as a structural prior for subsequent residual GCN layers. It should be noted that this graph reflects local relationships within the batch rather than global relationships across the entire dataset. Two residual GCN blocks based on GCN progressively transform the backbone features into representation, which is finally fed into a fully connected layer for binary classification. A class-weighted cross-entropy loss to mitigate class imbalance. The area under the receiver operating characteristic curve (AUC), Accuracy (ACC), sensitivity (SEN) and specificity (SPE), and were used to evaluate diagnostic performance. Results: The ResNet101-kNN-ResGCN architecture demonstrated improved diagnostic performance compared to baseline and alternative GCN configurations. The model achieved a peak AUC of 0.867 (95% CI: 0.803-0.919). In contrast, the standalone ResNet101 baseline yielded a lower AUC of 0.653 (95% CI: 0.544-0.765), while the chain-structured GCN and ResGCN reached an AUC of 0.815 (95% CI: 0.738-0.886) and 0.836 (95% CI: 0.762-0.897), respectively. Conclusions: This study demonstrates that a kNN-based residual GCN can capture discriminative patterns in the feature space associated with MTM-HCC from CT scans, offering a new approach for the deep learning based non-invasive preoperative risk stratification in clinical practice. The results indicate that incorporating residual connections within a kNN-based graph convolution framework may provide performance gain over both sequential GCN and standalone ResNet101 structures.