Deep learning-based auto-segmentation of paraganglioma for growth monitoring

TL;DR

Paragangliomas in the head and neck require longitudinal, precise tumor-volume measurements to guide treatment decisions, but existing methods are slow and variable. The authors train a 3D nnU‑Net to automatically segment paragangliomas on 3D TOF gadolinium-enhanced MRI and comprehensively evaluate qualitative and quantitative performance against manual delineations, including observer variability. They extend the analysis to growth-monitoring by linking volumes over time in 208 patients (311 tumors) and fitting seven growth models with constrained optimization, comparing results to prior work. The results show high concordance with manual delineations and reveal both the promise and limits of universal growth functions, providing a foundation for large‑scale, automated growth studies and potential clinical utility in radiology and oncology.

Abstract

Volume measurement of a paraganglioma (a rare neuroendocrine tumor that typically forms along major blood vessels and nerve pathways in the head and neck region) is crucial for monitoring and modeling tumor growth in the long term. However, in clinical practice, using available tools to do these measurements is time-consuming and suffers from tumor-shape assumptions and observer-to-observer variation. Growth modeling could play a significant role in solving a decades-old dilemma (stemming from uncertainty regarding how the tumor will develop over time). By giving paraganglioma patients treatment, severe symptoms can be prevented. However, treating patients who do not actually need it, comes at the cost of unnecessary possible side effects and complications. Improved measurement techniques could enable growth model studies with a large amount of tumor volume data, possibly giving valuable insights into how these tumors develop over time. Therefore, we propose an automated tumor volume measurement method based on a deep learning segmentation model using no-new-UNnet (nnUNet). We assess the performance of the model based on visual inspection by a senior otorhinolaryngologist and several quantitative metrics by comparing model outputs with manual delineations, including a comparison with variation in manual delineation by multiple observers. Our findings indicate that the automatic method performs (at least) equal to manual delineation. Finally, using the created model, and a linking procedure that we propose to track the tumor over time, we show how additional volume measurements affect the fit of known growth functions.

Figures (6)

• Figure 1: Example of delineation of the vessels. The left image shows the transversal view, the middle image shows the coronal view and the right image shows the sagittal view.
• Figure 2: Example of tumor detected over multiple locations by the model with two separate delineations in the reference (coronal view, blue: human/reference, pink: model). It shows an example of when the model delineates the tumors as one tumor while the human/reference delineation consists of two tumors.
• Figure 3: Examples of patients shown during the blinded study (blue: human, pink: model). In the first column, the delineations were judged equal. In the second column, the automatic delineation (pink) was preferred. In this case it seems as if a vessel was wrongly included in the human delineation. In the third column, the human (blue) delineation was preferred. In this case it seems as if the model tries to exclude the black hollow space.
• Figure 4: Examples of different delineations of the different observers (pink = model, red = h.a.1, yellow = h.a.2 and blue = h.a.3.), with A in the transversal plane and B in the sagittal plane. The delineations of the first tumor (carotid tumor) are more similar than the delineations of the second tumor (jugulotympanic).
• Figure 5: Box plots of the RMSE (in cc) of fitted functions on the three data sets. Numbers above the boxplots indicate the number of tumors for which RMSE $> 5$.