Learning from Many Eyes: Uncertainty-Aware and Calibrated AI for Trustworthy Pancreatic Cancer Analysis in Clinical Contexts

By: Meritxell Riera i Marín
Supervisors: Dr. Miguel Ángel González Ballester, Dr. Javier García López & Dr. Adrian Galdrán Cabello

Date: September 14, 2026 - 11:00 h
Room: 55.309

Abstract

This dissertation addresses the fundamental ”Ground Truth Problem” in medical imaging Artificial Intelligence (AI), where the conventional practice of collapsing divergent expert opinions into a single consensus mask creates overconfident and untrustworthy systems. In the context of high-stakes clinical decision-making, such as the analysis of Pancreatic Ductal Adenocarcinoma (PDAC), this deterministic approach is particularly hazardous. PDAC is characterized by ill-defined infiltrative margins and a complex relationship with adjacent vascular structures, leading to profound inter-rater variability that reflects inherent biological ambiguity rather than mere observer error. This research advocates for a paradigm shift, transitioning from rigid deterministic models toward probabilistic, multi-rater aware frameworks that treat expert disagreement as a valuable diagnostic signal.

The work is structured around three primary pillarss: evaluation, benchmarking, and training. First, we address the instability of current reliability assessments by introducing the Multi-Rater Expected Calibration Error (MR-ECE). This metric virtually expands test sets by considering each expert’s perspective independently, providing a stable measure of model calibration even in data-limited clinical scenarios. We further provide a thorough study on the complementarity of multi-rater metrics, demonstrating that a multi-faceted evaluation suite is essential to capture the nuances of model behavior.

Second, to enable the systematic study of diagnostic uncertainty, this thesis led the creation and public release of the CURVAS and CURVAS-PDACVI datasets through a MICCAI challenge series. These FAIR-compliant benchmarks provide the community with independent annotations from multiple experts, serving as a gold standard for evaluating algorithms under real-world human variability.

Third, we propose ordinal consensus learning utilizing a Ranked Probability Score (RPS) loss. By interpreting annotator consensus as an ordered hierarchy of confidence levels, our models produce inherently calibrated segmentations that preserve the ”gray zones” of clinical interpretation. Clinical validation in PDAC staging reveals that macroscopic volumetric overlap (Dice Score) is a poor proxy for the high-resolution precision required at the tumor-vessel interface.

Ultimately, this dissertation demonstrates that uncertainty-aware models provide a vital clinical ”safety net.” By flagging ambiguous cases for multidisciplinary review, these frameworks can help to bridge the trust gap between AI developers and clinical specialists, facilitating the integration of transparent and reliable AI tools into the clinical decision-making process of pancreatic cancer.

Computational Analysis of Piano Practice: Towards Modeling Practice Mistakes and Repetitions for Understanding Learning Behaviours

By: Alia Ahmed Morsi Moustafa
Supervisor: Dr. Xavier Serra

Date: September 29, 2026 - 14:00 h
Room: 55.309

Abstract

The main goal of this research is to understand the process of musical instrument learning in the Western Classical tradition, through developing computational methods for the analysis of piano practice recordings. We focus on mistakes and the organizational structure of practice sessions, as both can reveal information about one’s musical expertise. Specifically, our research (i) models mistakes in a manner that reflects educational understanding beyond deviations from a reference, (ii) trains classifiers that can automatically detect common piano learning mistakes, and (iii) develops methods to understand the organization of musical content in practice sessions based on similarity matrix analysis.

We propose treating piano mistakes as sequences that include both initial error and subsequent recovery phases, with each composed of low-level operations (note insertions, deletions, and time shifts) and release a toolkit that generates synthetic labelled data according to the proposed framework. Considering mistakes as behavioral phenomena with a contextual and qualitative aspect, we investigate whether the locations of salient errors in a performance can be detected through neighbouring context without music score comparisons, through supervised learning experiments combining synthetic and real data. Our results show feasibility while revealing generalization challenges across mistake types and performance contexts.

As for the computational analysis of practice organization, we develop a similaritybased segmentation method that identifies repeated attempts at passages, with emphasis on grouping together attempts of the same passage despite the existence of mistakes. We discuss how self-similarity analysis enables close examination of mistake patterns across repetitions to provide insights into planning and learning strategies, and to enable future comparative studies that track the evolution of learning.

This dissertation operates in a domain where computational measurement and educational meaning are still being aligned, and takes that gap seriously as a methodological starting point. The contributions above are components of a bigger picture towards formalizing realistic and useful computational analysis targets for piano learning. The choices of approach in mistake modelling, simulation, detection, and practice structure analysis collectively illustrate how methodological transparency can deepen the analysis itself.