Two Visual Intelligence PhD defenses within two days

Congratulations to Iver Martinsen and Durgesh Kumar Singh, who successfully defended their PhD theses at UiT The Arctic University of Norway on August 20th and 21st respectively.

By Petter Bjørklund, Communications Officer at SFI Visual Intelligence

This represents the first time where two PhD Candidates successfully defended their theses and received their PhD degrees one day apart from each other.

Martinsen's thesis, titled "Uncertainty and Representation Learning in Image Recognition: Advancing Deep Learning for Microfossil Analysis", contributes to advancing deep learning methodologies while demonstrating their potential for impactful applications in microfossil analysis. The trial lecture title was "Foundation Models for Structured Data: Progress, Challenges, and Opportunities".

Singh's thesis, titled "Towards more accurate and label-efficient Left Ventricle Automatic Measurements", presents a comprehensive framework that advances echocardiography analysis by combining anatomical constraints and weak supervision to improve the accuracy and reliability of LV linear measurements. Singh's trial lecture title was "Probabilistic deep learning and uncertainty in medical image analysis".

About Martinsen's thesis and defense

Summary of the thesis

Alongside the rapid development of deep learning, image recognition has achieved significant advances, with particularly promise within the supervised learning paradigm. While current models demonstrate strong predictive performance on benchmark datasets, deep neural networks often lack robustness, exhibiting unpredictable failures in operation. Challenges such as out-of-distribution data from unseen groups and domain-shift between training and test data highlight the need for accurate uncertainty estimates. Although many methods have been proposed, current approaches fail to address all sources of uncertainty comprehensively. Additionally, existing uncertainty measures are insufficient for practical evaluation, as they do not adequately assess the real-world usability. Microfossil analysis is one of many applications that stands to benefit from deep learning advancements. Microfossils are abundant worldwide and serve as indicators of past environments and subsurface structures, making them invaluable for both academic research and the energy sector. Recent digitization efforts have also produced large volumes of unlabeled data, underscoring the importance of methodologies that can exploit this. This thesis aims to address three core challenges in deep learning methodology: uncertainty, self-supervised learning, and interpretability. First, this work presents advances in uncertainty estimation and microfossil analysis. From a methodological perspective, uncertainty is advanced in two directions: by comparing deep learning uncertainty to human participants, leading to the proposal of new uncertainty measures, and by introducing a novel method for estimating uncertainty in out-of-distribution scenarios. Second, this thesis establishes a new state-of-the-art in microfossil analysis by using self-supervised learning. These advancements enable the development of an automated pipeline for microfossil analysis and the automatic generation of biostratigraphy reports. Finally, this work investigates interpretability methods, with a particular focus on analyzing a state-of-the-art model for sea ice prediction. By addressing these challenges, this thesis contributes to advancing deep learning methodologies while demonstrating their potential for impactful applications in microfossil analysis.

Supervisory Committee

• Professor Fred Godtliebsen, IMS, UiT (Main Supervisor)
• Professor Benjamin Ricaud, IFT, UiT
• Researcher Thomas Haugland Johansen, Forskningsinstitutt NORCE
• Researcher Steffen Aagaard Sørensen, IG, UiT
• Researcher David Wade, Equinor
• Professor Geir Olve Storvik, Institutt for matematikk, UiO
• Researcher Alba Ordoñez, Norsk Regnesentral, Oslo
• Researcher Anders Ueland Waldeland, Norsk Regnesentral, Oslo

Evaluation Committee

• 1st Opponent: Professor Anders Bjorholm Dahl, Denmark Technical University
• 2nd Opponent: Professor Arnoldo Frigessi, UiO
• Internal member and leader of the committee: Associate professor Elisabeth Wetzer, IFT, UiT

‍

About Singh's thesis and defense

Summary of the thesis

Accurate left ventricular (LV) linear measurements are critical for cardiac function assessment in echocardiography. However, manual measurements are time-consuming and prone to variability due to anatomical complexity and operator dependency. While deep learning (DL) models have automated landmark detection, existing fully automated B-mode approaches often produce inaccurate predictions caused by shifted landmarks and lack the flexibility to handle clinically challenging cases such as septal bulge or mid-ventricular measurements. This thesis addresses these challenges by developing methods to improve automation, accuracy, and generalization in echocardiographic LV linear measurements.

The first contribution presents a semi-automatic framework, Enhanced LVAM (EnLVAM), that leverages Anatomical Motion Mode (AMM) imaging to constrain landmark predictions along a user-defined virtual scanline. By aligning predictions with clinically relevant orientations, EnLVAM reduces variability and overcomes the limitations of B-mode models, particularly in complex anatomical scenarios. The semi-automatic design further allows human feedback, providing flexibility and robustness, and demonstrates improved measurement accuracy over fully automated baselines.

Building on this, the second contribution proposes WiseLVAM, which advances the framework toward full automation by eliminating the need for manual scanline selection. WiseLVAM learns to predict the optimal scanline position directly from B-mode images using weak supervision generated by the AMM-based approach. By estimating the LV contour and long axis, WiseLVAM performs shape-aware scanline placement aligned with clinical guidelines. The model then utilizes the trained AMM-based detector to execute precise LV linear measurements, offering a fully automated and clinically feasible solution.

Finally, the thesis explores learning under limited supervision by proposing a clustering-based regularization framework (SuperCM) designed for semi-supervised learning (SSL) and unsupervised domain adaptation (UDA). SuperCM explicitly enforces clustering assumptions during training, promoting compact and class-consistent feature representations. This approach improves model generalization when labeled data is scarce—a common challenge in medical imaging—and demonstrates effectiveness on standard SSL and UDA benchmarks.

Overall, this thesis presents a comprehensive framework that advances echocardiography analysis by combining anatomical constraints and weak supervision to improve the accuracy and reliability of LV linear measurements. Additionally, the thesis provides a detailed discussion highlighting the contributions within the broader research context of echocardiography analysis and outlines future research directions for developing robust and clinically deployable solutions.

Supervisory Committee

• Associate Professor Michael Kampffmeyer, IFT, UiT (Main Supervisor)
• Professor Robert Jenssen, IFT, UiT
• Postdoktor Ahcene Boubekki, IFT, UiT

Evaluation Committee

• 1st Opponent: Dr. Alberto Gomez, forsker ved Ultromics LTD, Stor Britannia
• 2nd Opponent: Professor Lasse Løvstakken, Institutt for sirkulasjon og bildediagnistikk, NTNU
• Internal member and leader of the committee: Associate profesor Elisabeth Wetzer, IFT, UiT