An Explainable Deep Learning Model for Brain Imaging-based Surgical Skill Assessment

Introduction: Objective metrics derived from the non-invasive brain imaging has shown to be promising in correlating brain activations to the behavioral measures of surgical skill. However, the existing models are limited to the prediction of surgical task performance and lacks the ability to reveal brain activation sequences that are associated with the task specific proficiency. In this study, we present an explainable deep learning model to simultaneously classify the skill level from the brain imaging data and identify the segment of the task that contributes the most to the prediction of surgical expertise.

Materials and

Methods: A continuous-wave near-infrared spectrometer was deployed to record signals, while participants performed fundamental of laparoscopic surgery intracorporeal suturing task following IRB/HRPO approved protocol. Both the experts and novice subjects performed five trials each. Each trial lasts a maximum of ten minutes, with two minutes of rest in between trials. The deep learning model was developed using the dataset comprised of HbO2 signals obtained from pre-frontal cortex from 29 trials performed by 7 experts and 26 trials by 9 novice subjects. The model employs an attention-based Long Short-Term Memory recurrent autoencoder to encode the context of a multivariate brain imaging signals for skill classification. One of the novel aspects of this model is the attention mechanism that not only enhances the generalizability of the prediction, but also highlights the most discriminative part of the surgical task. 10-fold cross-validation is performed for the independent assessment of the classifier.

Results, Conclusions, and Discussions: The performance metrics given in Table 1 indicates that the classifier misclassifies only one expert trials as novice trial. All novice trials are accurately identified. The area under the curve (AUC) value of 1.00 for both the receiver operating characteristic (ROC) and precision-recall (PR) curves indicates the robustness of the classifier. The high value of Matthew’s correlation coefficient (MCC) and F1-score further underscores the networks’ ability to classify classes with high accuracy.

The LSTM recurrent autoencoder proposed herein demonstrate high performances in classifying subjects based on their skill level. This architecture leverages the whole spatiotemporal features of the functional near-infrared spectroscopy (fNIRS) data set and as expected outperforms AI model using data reduced input. The heatmaps obtained from the multivariate analysis using PFC channels (Figure 1) has revealed that early PFC activation ( < 30s) contributes the most to the classification of skill level, independent of FLS scores. While the heatmaps of the experts highlights the importance of initial suturing subtasks, those for the novice also imphasize the importance of other subtasks. These findings demonstrate the unique potential of the proposed deep leaning model in not only classifying the skill level based on neuroimaging, but also identifying the subtasks where novice learners may benefit from interventions.

Acknowledgements (Optional): The authors gratefully acknowledge the support of this work through the Medical Technology Enterprise Consortium (MTEC) award #W81XWH2090019 (2020-628), and the U.S. Army Futures Command, Combat Capabilities Development Command Soldier Center STTC cooperative research agreement # W912CG2120001.