Meningiomas are the most common primary brain tumors; while usually benign, these tumors can lead to significant morbidity (1–3). Given their origin from meningothelial cells, meningiomas can have significant skull-brain interface characteristics with extracranial extension and spread along the dura (1, 2). Treatment is often individualized and planned based on anatomical location as well as World Health Organization (WHO) classification (1–3). Treatment typically involves a combination of observation, surgical resection, and radiation therapy. Radiation therapy plays a large role in mitigating recurrence after resection (4–8) for WHO grade 1 and grade 3 tumors as well as being the primary treatment modality for unresectable meningiomas (1, 9, 10). Auto-segmentation tools, i.e. deep neural network (DNN) models, have promising implications for use in radiation therapy planning which require a gross tumor volume (GTV) to be contoured (11–14). These DNNs have the potential to reduce contour variability, leading to improved clinical outcomes while increasing workflow efficiency (11–13). Most automated brain tumor segmentation efforts to date have been focused on gliomas (15–17), despite meningiomas being the most common type of primary brain tumor. The few segmentation studies on meningiomas are commonly related to general volume segmentation, for clinical use in analyzing tumor growth (18–21). These meningioma segmentation studies are based on contrast-enhancing tissue observed in magnetic resonance imaging (MRI) (18–21). Auto-segmentation of meningioma for radiotherapy targeting is even less studied. Meningioma patients that are treated with radiation are either post-resection or have unresectable tumors where delineation of meningioma GTVs incorporates an additional layer of complexity. GTV segmentation proves to be a challenge due to meningioma’s extra-axial characteristic locations and up to 72% having an associated dural tail (22). Meningioma GTV segmentation may include the dural enhancement and invaded bone in addition to the main contrast-enhancing mass (23, 24). The extension of the dural tail segmentation can be subjective and difficult to define.

High-performing segmentation DNNs can serve as a valuable tool; however, model performance is heavily reliant on the robustness of the training data. When used for complex structures and at higher risk clinical scenarios, the accuracy and accountability of these models are an area of concern (25). Predictions made without uncertainty quantification may not be trustworthy causing DNN models to create a false sense of certainty (25, 26). DNNs sometimes make unexpected incorrect predictions and the need for uncertainty quantification is apparent. There are numerous studies based on analyzing model uncertainty through two approaches: model-based uncertainty (epistemic) and data-based uncertainty (aleatoric) (25, 26). Epistemic uncertainty quantification methods include Bayesian neural networks (27, 28) and Monte Carlo dropout (29, 30), while aleatoric uncertainty methods include test-time augmentation (31, 32). However, there is not a standard method to quantify uncertainty. Specifically, a pixel-wise uncertainty quantification method remains unavailable. A pixel-wise uncertainty metric would quantify the likelihood that each pixel belongs in the segmentation region, providing a corresponding uncertainty quantification with each segmentation prediction (33).

This innovative work investigated the use of spherical image projection to quantify DNN segmentation aleatoric uncertainty for meningioma radiotherapy target delineation. Inspired by spherical camera image processing (34), our group developed a novel spherical projection-based method to enable image segmentation uncertainty estimation of DNN models. Such design emphasizes locoregional image texture details near the projection center while maintaining the global anatomy information (35). Previously, our group has demonstrated that our spherical projection design can successfully quantify uncertainty in multi-parametric MRI-based glioma segmentation (33). This work is the first of its kind in deep learning-based meningioma segmentation with uncertainty analysis. This feature serves as an invaluable tool for clinicians, offering visual cues that facilitate efficient correction of deep learning segmentation outputs. We hypothesize that the implementation of spherical projection-based segmentation and uncertainty quantification can hold great value in meningioma radiotherapy target delineation.

Figure 5 shows three examples of outputs from the U-Net and SPU-Net models. In all three examples, SPU-Net’s segmentation displays more visual similarity to the GT compared to U-Net’s segmentation. The most obvious visual discrepancy is observed around the tumor edges and near dural vessels, for SPU-Net and more so for U-Net predictions. In the bottom examples, with 3 tumors, SPU-Net over-segmented the dural tail region of the left-most lesion with U-Net not able to segment the middle lesion. Similarly observed in other test cases, U-Net was observed to have difficulty delineating small lesions, especially those in proximity with vasculature. Generally, SPU-Net is observed to have better segmentations when meningiomas are near dural vessels and generally have better dural tail segmentations. SPU-Net can better delineate and differentiate between meningioma tissue and neighbouring blood vessels.

The SPU-Net’s segmentation mask and uncertainty map showcase a correlation. Segmentation errors are reflected in the uncertainty map with the region having high uncertainty. Accurate regions in the segmentation mask are reflected to yield low uncertainty. This observation meets the expectation of a favored U visualization. SPU-Net achieved a U-score of 0.718 ± 0.183, which will be used as benchmark results for future meningioma segmentation uncertainty studies.

Table 1 summarizes the segmentation results evaluated between the SPU-Net and traditional U-Net models. While both models achieved acceptable results, SPU-Net achieved slightly higher sensitivity and DSC values. SPU-Net achieved comparable sensitivity (0.758 vs. 0.746) and Dice similarity coefficient (0.760 vs. 0.742) results from U-Net. SPU-Net also had smaller mHD (0.612 cm vs 0.744 cm) and smaller HD95 (2.682 cm vs 2.912 cm). Additionally, smaller variances are observed in SPU-Net results. These numerical result improvements, though promising, did not achieve statistical significance in Wilcoxon signed-rank tests.

In this work, a novel method to quantify deep learning segmentation uncertainty was investigated for use in meningioma radiotherapy target delineation. This novel method of spherical projection allows for projecting planar MR images onto a spherical surface, which could be considered as a form of nonlinear image transformation. Utilizing multiple projection centers across the FOV, structures are magnified at varying scales. The spherical projection-based method obtains multiple independent segmentation predictions from a single mpMRI input and allows for improved segmentation and most importantly, uncertainty quantification.

Meningioma segmentation presents a significant challenge even for experienced radiation oncologists due to the inherent subjectivity surrounding the delineation of GTV boundaries. GTV boundaries often overlap with blood vessels specifically in the dural tail segmentation region, where the extent of the dural tail segmentation is based on judgement and experience. Some physicians may be more cautious and extend the GTV edges and dural tails further than others. SPU-Net addresses the challenge of meningioma segmentation and the major concern regarding DNN models of not having a confidence metric. SPU-Net provides a pixel-wise uncertainty map to accompany the binary segmentation prediction, to provide indication of prediction confidence. The pixel-wise uncertainty map in the same FOV as the segmentation mask substantially enhances the interpretability of a binarized segmentation mask. Figure 5 illustrates the value of the uncertainty map. With meningioma being a more challenging segmentation task due to interobserver variability, there is a higher likelihood of segmentation errors where the uncertainty map offers visual cues on the anatomical locations of these errors. The uncertainty map highlights the areas of high uncertainty, predominantly observed around the GTV periphery, regions overlapping with dural vessels, and the dural tail region. This observation corresponds to known areas of clinical uncertainty for oncologists. This alignment between SPU-Net’s uncertainty map and the clinical challenges faced by radiation oncologists demonstrates its potential for clinical adoption, providing more interpretable and clinically relevant results that reflect oncologists’ real-world experiences.

The results indicate comparable performance between SPU-Net and U-Net in quantitative segmentation metrics of sensitivity, DSC, mHD, and HD95. However, SPU-Net was observed to handle complex meningioma shapes and dural tail segmentations more effectively compared to U-Net. By incorporating spherical magnification, the SPU-Net GTV segmentations had better dural tail segmentations compared to the traditional U-Net model ( Figure 5 ). U-Net sometimes struggled to segment areas where the meningioma overlapped with dural vessels. In certain instances, U-Net failed to generate a prediction altogether in such complex regions, highlighting the limitations of the traditional model in handling these challenging anatomical features. SPU-Net’s ability to handle the challenging dural tail region more effectively highlights its potential. Given the extra-axial nature of meningiomas, segmentation difficulties were expected for both models, but any improvement, especially in the dural tail region, is notable.

This study highlights the importance of incorporating uncertainty quantification into deep learning models, specifically through the use of spherical projection-based DNNs for meningioma segmentation, a method that has also shown promise in glioblastoma segmentation. Though the results are not statistically significant, the technical innovation justifies further exploration. A key advantage of spherical projection is its ability to generate an uncertainty map that is spatially aligned with the segmentation prediction. This uncertainty map provides valuable insights into the underlying mechanisms driving the DNN’s decisions, enhancing interpretability. For clinical applications, the uncertainty map offers clinicians a clearer understanding of the model’s predictions, helping them make more informed decisions and increasing their confidence in adopting the model. Future studies with larger cohorts and in different clinical applications could provide a deeper understanding of SPU-Net’s full potential and its broader applicability across different tumor types and segmentation challenges. The use of spherical projection is a model-agnostic method, it can be applied to other DNN architectures and for other segmentation applications (35).

This study can be further analyzed and improved by being more inclusive to inter-personal variations in GTV segmentation and incorporating these into the study on a quantitative level. This study utilized a GTV contoured by one singular physician, which introduces bias to the study. A more accurate representation of the ground truth may be an average of multiple physicians’ GTV contours. The variability between multiple physicians’ GTV contours could also represent a clinically acceptable uncertainty range.

The challenge of meningioma GTV subjectivity in segmentation can also be addressed by incorporating a larger dataset; to support a publicly available large dataset, the image data included in this work will be submitted to the 2024 BraTS meningioma radiotherapy segmentation challenge. The results in this study showed promising improvements through the incorporation of spherical projections. This issue could be potentially overcome by utilizing a larger cohort and increasing the sample size. With a larger cohort, the study can also incorporate infratentorial located tumors. Infratentorial tumors are particularly challenging to segment due to the complex anatomy of the posterior fossa (a vast number of blood vessels) and frequent imaging artifacts observed in this region, making it particularly tricky for an MRI only based segmentation task. The inclusion of CT scans is being considered to facilitate infratentorial tumor segmentation. Expansion upon this current study will require re-optimizing the SPU-Net workflow to be more robust to the variation in anatomically located meningiomas.

A spherical projection-based U-Net (SPU-Net) was successfully developed for meningioma segmentation using multi-parametric MRI. In a comparison study against a classic U-Net, SPU-Net achieved comparable results of sensitivity, DSC, mHD, and HD95. While comparable to U-Net, SPU-Net has the added benefit of providing uncertainty quantification in the form of an uncertainty map which is spatially aligned with the segmentation prediction. This uncertainty map closely matches clinical expectations, accurately reflecting areas where oncologists typically face difficulty in delineating meningiomas, such as regions overlapping with the skull, blood vessels, and the dural tail. These regions are known to have higher uncertainty in manual delineation, and SPU-Net’s results mirror this clinical challenge. By providing an interpretable outcome that highlights these uncertainties associated with segmentation, SPU-Net significantly enhances the potential for clinical adoption after future preclinical validation studies using a larger dataset, bridging the gap between automated systems and human expertise in the pursuit of precision medicine.