A deep learning approach for objective evaluation of microscopic neuro-drilling craniotomy skills.

Background: Minimally invasive microscopic and endoscopic neurosurgery demands precise use of high-speed micro-drilling tools to prevent potential complications. Present-day neuro-drilling training methods include cadaveric specimens and surgical simulators. However, skills assessment is mostly manual, and there is a pressing need for automation and personalized feedback for trainee surgeons. The lack of well-annotated datasets limits deep learning (DL)-based automation.

Methods: The study poses microscopic neuro-drilling skill evaluation as a rank estimation problem. It presents a geometric-order-learning based framework to effectively train transformer-based DL models in ultra-low-data settings. The study demonstrates that the proposed framework enhances feature separability in embedding space. Furthermore, it suggests a framework for automatic detection of the drilling regions. Additionally, the study contributes a comprehensive dataset of 435 images encompassing the micro-drilling on various specimens, including deceased-sheep heads and scapulae. The enhancement in the performance and practical utility of the proposed system is illustrated using various qualitative and quantitative methods.

Results: The proposed model exhibits a mean squared error (MSE) of 0.77 and ±1 accuracy of 95.17%. Utilization of the proposed framework results in an average improvement of 24.71%, in ± 1 accuracy, across five state-of-the-art (SOTA) transformer-based architectures. Additionally, significant enhancement is observed in feature separability in the embedding space for both Convolution Neural Network (CNN) and Transformer-based architectures. Furthermore, the proposed model outperforms the independent expert evaluator by 12.96% in MSE and 8.47% in ± 1 accuracy.

Conclusion: This study introduces the first-ever well-annotated unbiased microscopic neuro-drilling effectiveness dataset and automated skill evaluation system, which surpasses the performance of an independent expert evaluator. It can be used as an unbiased automated evaluation tool for neurosurgical training worldwide. Source Code and Dataset.