A Clinical Guideline-Driven Automated Linear Feature Extraction for Vestibular Schwannoma

Vestibular schwannoma (VS) is the most common brain tumor that grows in the cerebellopontine angle, accounting for 8% of all the intracranial adult brain tumours. Patients may be treated by surgery, radiosurgery or with a conservative “wait-and-scan” strategy. Irrespective of the type of treatment received, most patients require prolonged radiological follow-up and regular surveillance imaging. The choice and timing of treatment depends on the patient's symptoms and the behavior of the tumor. For this, a consistent and standardized measurement of the tumor should be extracted and reported. The Consensus Meeting on Systems for Reporting Results in Acoustic Neuroma (2001) advised that (1) the intrameatal and extrameatal portion of the VS should be clearly distinguished and (2) the largest extrameatal diameter should be measured. If the tumor only has an intrameatal portion, the maximum tumor diameter of the whole tumor region should be reported.

In current clinical settings, these features are extracted manually by a neuroradiologist, prior to a multidisciplinary meeting where treatment decisions are made. This manual linear feature extraction task is typically performed by an experienced neuroradiologist but is time-consuming, labor-intensive and prone to variability and subjectivity.

Despite the fact that linear measurements are typically used in current practice, volumetric tumor measurements have been shown to be the most accurate and sensitive measure of detecting VS growth. Implementing routine volumetric measurements would enable clinicians to offer earlier appropriate intervention. Artificial intelligence (AI)-driven clinical support tools have the potential to improve patient outcomes and experience by the standardization and personalization of VS treatment. In our previous research work, we have indeed demonstrated that AI tools are technically capable of automatically detecting and segmenting VS and even delineating intra-/extra-meatal components. Such tools can serve as a foundation for the automated feature extraction task.

In this work, we propose an automated clinical feature extraction framework for VS. First, we developed deep learning based segmentation models for intra-/extra-meatal segmentation for different MRI modalities. Using the tumor contours, we then developed feature extraction algorithms to measure tumor size while presenting the most appropriate linear measurement following the standardized clinical guidelines as implemented at a large specialist tertiary referral center in the United Kingdom. We analyzed the relationship between our extracted features and the manually extracted features by an expert neuroradiology consultant (S.C.) for 50 patients. The results indicated a significant correlation between the manual and automated measurements (p < 0.0001).

Our framework can generate clinically relevant features to aid the management of patients with vestibular schwannoma. Our results indicate a promising expansion of this work. These features could be incorporated into a formal report that may be used to support clinical decision making by the specialist multidisciplinary team meetings. This work could also be adapted to other benign brain tumours to output tumor-specific features.

Fig. 1 The proposed framework—segmentation module—consists of a two-stage approach for preoperative scans, whereas the fine-tuned stage 1 model is used for the postoperative scans. This is integrated with the feature extraction module to provide clinical features.

Publikationsverlauf

Artikel online veröffentlicht:
05. Februar 2024