The imperative need for cooperation: strengthening AI modeling in vestibular schwannoma radiology

In the last decade, artificial intelligence (AI) in medicine has become one of the fastest growing fields of research. Applications of these machine-learning techniques are mostly related to computer-aided detection and diagnosis. Recently, for vestibular schwannoma, several articles on automated tumor segmentation have been published. This technique allows for the automated calculation of tumor volumes, enabling clinicians to accurately track the volume over time, thereby further improving the monitoring of individual patients. Next to this, the possibility to employ AI for various treatment outcome predictions is emerging. For example, enabling the prediction of tumor progression during wait-and-scan, or long-term tumor control following stereotactic radiosurgery, may aid clinicians in selecting the optimal treatment strategy on individual patient basis.
However, one of the major issues concerning AI is the need for large amounts of data. Especially in tumors like vestibular schwannoma, this can become highly troublesome. Its scarce availability can severely limit the amount of data available for medical research centers to develop advanced machine-learning tasks. Furthermore, models trained on single-center data may pose additional limitations caused by several biases occurring in the employed data. Examples of such biases are single-machine MRIs, patient selection, treatment planning, etc. As a result, these aspects make cooperation in AI research concerning VS tumors imperative. Combining data from multiple medical centers increases the heterogeneity in the data, which is essential to generate robust and reliable models. To illustrate the impact in the case of automated tumor segmentation, employing a single-center-data trained model[1] on data from our institution resulted into a significant drop in performance. Updating the model by training it on data from both institutions caused the performance to increase on our data without decreasing the results on the original data. Furthermore, a multi-institution-data trained model[2] presented high performance on data from our institution, which it had not seen during training. These examples clearly highlight the need for sharing the available data in AI research. It increases the heterogeneity in the data and thereby reduces the effect of biases that are present, thereby allowing the creation of robust models that generalize across unseen data.

References:
[1] Shapey, J., Wang, G., Dorent, R., Dimitriadis, A., Li, W., Paddick, I., Kitchen, N., Bisdas, S., Saeed, S. R., Ourselin, S., Bradford, R., & Vercauteren, T. , (2021), An artificial intelligence framework for automatic segmentation and volumetry of vestibular schwannomas from contrast-enhanced T1-weighted and high-resolution T2-weighted MRI, the American Association of Neurological Surgeons, Journal of Neurosurgery, 171-179, 134(1)
[2] Kujawa, A., Dorent, R., Connor, S., Thomson, S., Ivory, M., Vahedi, A., Guilhem, E., Bradford, R., Kitchen, N., Bisdas, S., Ourselin, S., Vercauteren, T., Shapey, J. , (2022), Deep Learning for Automatic Segmentation of Vestibular Schwannoma: A Retrospective Study from Multi-Centre Routine MRI, Cold Spring Harbor Laboratory Press, medRxiv, https://www.medrxiv.org/content/early/2022/08/02/2022.08.01.22278193