Evaluation of Monte Carlo Dropout for Uncertainty Quantification in Multi-task Deep Learning-Based Glioma Subtyping

Gonzalo Esteban Mosquera Rojas, Sebastian Robert van der Voort, Carolin Pirkl, Sandeep Kaushik, Marion Smits, Stefan Klein

Keywords: Uncertainty quantification · Monte Carlo Dropout · Magnetic Resonance Imaging · Glioma

Abstract:

Uncertainty Quantification (UQ) is essential for enhancing
the trustworthiness of Deep Learning (DL) models in high-stakes medical
imaging applications. Monte Carlo Dropout (MCD) remains one of
the most widely used and foundational approaches for UQ, often serving
as a baseline in comparative studies. In this work, we systematically
evaluate MCD in the context of DL-assisted glioma diagnosis, focusing
on a less-explored yet clinically relevant multi-task setting that combines
glioma subtyping and segmentation. We investigate how key parameters
of MCD, namely the number of MC samples and the dropout rate, may
affect the quality of uncertainty estimates. Additionally, we disentangle
epistemic and aleatoric uncertainty components to gain deeper understanding
of model confidence. The results demonstrate that, when appropriately
tuned, MCD produces well-calibrated uncertainty estimates.
The segmentation task was primarily influenced by epistemic uncertainty,
whereas aleatoric uncertainty constituted the main source of uncertainty
in all classification tasks.

Submission Number: 17