Bridging Ex-Vivo Training and Intra-operative Deployment for Surgical Margin Assessment with Evidential Graph Transformer

PURPOSE: The use of intra-operative mass spectrometry along with Graph Transformer models showed promising results for margin detection on ex-vivo data. Although highly interpretable, these methods lack the ability to handle the uncertainty associated with intra-operative decision making. In this paper for the first time, we propose Evidential Graph Transformer network, a combination of attention mapping and uncertainty estimation to increase the performance and interpretability of surgical margin assessment. METHODS: The Evidential Graph Transformer was formulated to output the uncertainty estimation along with intermediate attentions. The performance of the model was compared with different baselines in an ex-vivo cross-validation scheme, with extensive ablation study. The association of the model with clinical features were explored. The model was further validated for a prospective ex-vivo data, as well as a breast conserving surgery intra-operative data. RESULTS: The purposed model outperformed all baselines, statistically significantly, with average balanced accuracy of 91.6%. When applied to intra-operative data, the purposed model improved the false positive rate of the baselines. The estimated attention distribution for status of different hormone receptors agreed with reported metabolic findings in the literature. CONCLUSION: Deployment of ex-vivo models is challenging due to the tissue heterogeneity of intra-operative data. The proposed Evidential Graph Transformer is a powerful tool that while providing the attention distribution of biochemical subbands, improve the surgical deployment power by providing decision confidence.