Fredrik Andreas Dahl

People have different views on the current level of intelligence in AI systems, the likely speed of further progress, and the societal impact AI may have in the future. This has contributed to a polarized debate, with narratives ranging from “AI will give us some useful tools”, via “Many jobs will disappear” and “We will all be rich”, to “AI will take over and kill us all”. In this presentation, I will map the main narratives about AI risk and benefit using a railway metaphor in which human extinction is the final destination. The different stops along the way represent arguments for more favourable outcomes. The goal is not to argue for the likelihood of any particular scenario, but to structure the discussion of these narratives in a systematic way.

Abstract Objectives To investigate mammographic features associated with high artificial intelligence (AI) risk scores as provided by two AI models applied to screening mammograms. Materials and methods This retrospective study included 130,031 screening mammograms from 42,371 women attending BreastScreen Norway, 2008–2018. Two AI models (A and B) developed for cancer detection on screening mammograms were applied. An informed radiological review was conducted for mammograms within the highest 5% of AI risk scores by both models in two study samples: (1) High AI risk score, but no breast cancer detected within 6 years ( n  = 120), and (2) High AI risk score in mammograms with screen-detected cancers ( n  = 120). Mammographic density (BI-RADS a–d), features (mass, spiculated mass, asymmetry, architectural distortion, calcification alone, and density with calcification), and radiologists’ interpretation scores (1–5) were analyzed descriptively. Results Mammographic density was higher in sample 1 compared to sample 2 (BI-RADS d: 11% vs 3%, respectively). In sample 1, calcifications alone were the most frequent AI-marked feature (model A: 72%; model B: 68%), predominantly with amorphous morphology and a cluster distribution, and 76% were interpreted as benign by the radiologists (interpretation score 1). In sample 2, a spiculated mass was the most frequent mammographic feature among the screen-detected cancers (29%). Conclusion Mammograms assigned high AI risk scores exhibit distinct features depending on screening outcome. Systematic characterization of these features may help refine AI thresholds, improve specificity, reduce AI false-positive findings, and decrease the recall rate in breast cancer screening. Key Points Question Knowledge about mammographic features associated with high AI risk scores is essential for distinguishing cancer from non-cancer cases. Findings Calcifications were the dominant feature in non-cancers in screening mammograms with high AI risk score, whereas spiculated mass was the most frequent feature among cancers. Clinical relevance Calcifications in non-cancer screening mammograms with a high AI risk score were frequently interpreted as benign or probably benign by radiologists. This knowledge may help refine AI thresholds and thereby improve specificity and reduce false-positive results in mammographic screening. Graphical Abstract

I denne rapporten beskrives arbeidet og resultater fra en utvidelse av validering og evaluering av produkter for snødekningsgrad (FSC) fra 2024. Hensikten har vært å undersøke hvordan avvikene (feil) i FSC-verdier, relativt til «fasiter» fra bilder med høyere oppløsning, varierer i tid og rom med aggregeringsnivå, arealdekke og terreng. Analysene er gjort for Østlandet og tilsigsområder valgt ut av NVE med satellittdata over flere år. Referansedata («fasiter») er basert på Sentinel-2 MSIprodukter i 10 m oppløsning, mens FSC-produktene som ble analysert, er basert på 0,5 km data fra Sentinel-3 SLSTR. Resultatene viser at avvikene har tydelig romlig og tidsmessig struktur. Romlig aggregering reduserer MAE og RMSE, mens bias i hovedsak bevares. En betydelig del av feilen midles likevel ikke effektivt ut ved aggregering opp til de største testede skalaene, noe som tyder på systematiske avvik over større områder. I høydesoneanalysene fremkommer et robust mønster med økte avvik rundt 400-500 m. Når FSC aggregeres til sone-middelverdier per dato reduseres avvikene sammenliknet med pikselbasert stratifisering, men mønsteret i høyde avtar ikke fullt ut. Avvikene er generelt større i skog enn i områder med bart fjell og sparsom vegetasjon, med en negativ bias i skog, som er konsistent med utfordringer knyttet til snø under trekroner. Samlet viser analysen hvordan avvik endrer karakter ved aggregering til modelleringsrelevante enheter, og peker på forhold som bør tas hensyn til ved bruk av FSC som areal- og høydesoneaggregert modellinput

In the AIforScreening project we have tested different ways of utilizing time sequence information in mammography screening, including published methods and home-grown ones. The simplest ones include regression modelling, where we apply a single-image breast cancer risk model at a sequence of images of a given breast and use linear regression to onstruct a modified score. This method givs a very modest improvement of the order 0.001 on the AUC scale, which was statistically significant only for the inferior holistic model. The more advanced methods try include co-registration of the current and previous images and various ways of merging the model’s features to produce improved risk scores, utilizing various so-called Siamese net models. Over-all, the results were negative, as none of the advanced methods gave improvements above the linear model. This is contrary to published results, and we speculate that this may be due to the fact that our model has a high performance to begin with, leaving less room for improvement. The linear model places positive weight on the previous risk scores, which go against the intuition that an increase in risk score over time should increase the likelihood of cancer. Apparently, the ’direct effect’ that an elevated risk score is associated with future cancer is stronger.

Maintaining the performance of a deep learning model trained for breast cancer detection on a specific scanner type is challenging in a multi-scanner setting due to domain shifts caused by variations in imaging data. This often results in a performance drop when models trained on one scanner are tested on another. While re-training with labeled data from the new scanner is an option, delays in obtaining ground-truth labels make this approach impractical. To overcome this limitation, Unsupervised Domain Adaptation (UDA) offers a promising alternative by enabling models to adapt across scanners without requiring labeled target data. In this study, we investigated Conditional Domain-Adversarial Network (CDAN), an adversarial UDA approach, to adapt a classifier trained on Siemens scanner data using nearly 3 million mammograms from the Norwegian breast cancer screening program. We compared it to Maximum Mean Discrepancy (MMD), a simpler statistical feature alignment method, and evaluated histogram matching, which required no additional training. Our findings showed that the AUC drop on the target GE data (0.96 to 0.62) without adaptation was mitigated by histogram matching (AUC 0.84), but that was less effective than MMD (AUC 0.87), which performed competitively with CDAN. Further ablation with Domain-Adversarial Neural Network (DANN), the foundation of CDAN, suggested limitations in the domain discriminator. Unlike prior work focusing solely on performance, we paired UDA with explainability. This revealed how feature relevance shifted across scanner domains, offering novel insights into model generalizability in cancer detection.

Delineation of structures and estimation of landmarks in mammograms is a critical step in the evaluation of image quality in breast cancer screening, but requires the estimation of the uncertainty of the predicted landmarks to refer uncertain cases to clinicians. Of particular importance – and the focus of this work – is on the pectoral muscle, where the variability in muscle visibility across images introduces significant uncertainty. While graph convolutional networks (GCN) have been demonstrated to accurately predict landmarks by explicitly leveraging structural relationships between landmarks, they typically lack the ability to provide accurate uncertainty estimates for the landmarks. To address this shortcoming, in this work we propose a novel GCN-based approach that not only locates key points along the muscle boundary but also provides accurate uncertainty estimates, capturing both the aleatoric and epistemic uncertainties. Our method was evaluated on in-house annotated mammograms demonstrating comparable accuracy to human annotators, while at the same time providing highly accurate uncertainty estimates, confirming its potential for identifying cases that require human review. We further validate our proposed approach on the publicly available CSAW-S and INBreast datasets, demonstrating its robustness to domain shift, as well as its potential to detect incorrect or untypical annotations.

Two deep learning–based artificial intelligence (AI) models, one commercially available and one in-house, showed good performance for stand-alone cancer detection on retrospective mammography screening data. AI markings on the mammograms corresponded well to the true cancer location. Purpose - To evaluate cancer detection and marker placement accuracy of two artificial intelligence (AI) models developed for interpretation of screening mammograms. Materials and Methods - This retrospective study included data from 129 434 screening examinations (all female patients; mean age, 59.2 years ± 5.8 [SD]) performed between January 2008 and December 2018 in BreastScreen Norway. Model A was commercially available and model B was an in-house model. Area under the receiver operating characteristic curve (AUC) with 95% CIs were calculated. The study defined 3.2% and 11.1% of the examinations with the highest AI scores as positive, threshold 1 and 2, respectively. A radiologic review assessed location of AI markings and classified interval cancers as true or false negative. Results - The AUC value was 0.93 (95% CI: 0.92, 0.94) for model A and B when including screen-detected and interval cancers. Model A identified 82.5% (611 of 741) of the screen-detected cancers at threshold 1 and 92.4% (685 of 741) at threshold 2. Model B identified 81.8% (606 of 741) at threshold 1 and 93.7% (694 of 741) at threshold 2. The AI markings were correctly localized for all screen-detected cancers identified by both models and 82% (56 of 68) of the interval cancers for model A and 79% (54 of 68) for model B. At the review, 21.6% (45 of 208) of the interval cancers were identified at the preceding screening by either or both models, correctly localized and classified as false negative (n = 17) or with minimal signs of malignancy (n = 28). Conclusion - Both AI models showed promising performance for cancer detection on screening mammograms. The AI markings corresponded well to the true cancer locations.

Accurate uncertainty information associated with essential climate variables (ECVs) is crucial for reliable climate modeling and understanding the spatiotemporal evolution of the Earth system. In recent years, geoscience and climate scientists have benefited from rapid progress in deep learning to advance the estimation of ECV products with improved accuracy. However, the quantification of uncertainties associated with the output of such deep learning models has yet to be thoroughly adopted. This survey explores the types of uncertainties associated with ECVs estimated from deep learning and the techniques to quantify them. The focus is on highlighting the importance of quantifying uncertainties inherent in ECV estimates, considering the dynamic and multifaceted nature of climate data. The survey starts by clarifying the definition of aleatoric and epistemic uncertainties and their roles in a typical satellite observation processing workflow, followed by bridging the gap between conventional statistical and deep learning views on uncertainties. Then, we comprehensively review the existing techniques for quantifying uncertainties associated with deep learning algorithms, focusing on their application in ECV studies. The specific need for modification to fit the requirements from both the Earth observation side and the deep learning side in such interdisciplinary tasks is discussed. Finally, we demonstrate our findings with two ECV examples, snow cover and terrestrial water storage, and provide our perspectives for future research.

In this study, we apply the recent explainability methodologies Concept Relevance Propagation (CRP) and Relevance Maximization (RelMax), for understanding the decision-making processes of a deep learning classifier trained for mammogram-based breast cancer detection. The primary goal was to leverage CRP and RelMax to identify learned concepts that significantly influence the model's classifications. We aimed to present these in a manner intelligible to humans, with the intention of using this insight for quality assurance of the model. The study found that, although most high relevance concepts were shared across cancer subclasses, the model had formed a small number of concepts exclusive to our selected subclasses. These identified patterns seemed to capture well our intuitive understanding of what the given cancer subclass should look like. In addition, they also aligned well with the metadata information provided by radiologists and what they are typically familiar with. We believe this could further enhance their trust in the cancer classifier, since what the model has learned becomes more comprehensible and visually accessible. Our results refine a previously used explainability method, which was limited to localizing where the model directed its attention, but lacked in providing what was its understanding or offering additional clarification on the identified cancer subclass.

This study introduces an enhanced version of a two-stage modelling approach using artificial intelligence (AI) for breast cancer detection in mammography screening. Leveraging a large dataset of 2,863,175 mammograms from the BreastScreen Norway, the approach uses two convolutional neural networks. The first one is trained to classify whole images, and an explainable-AI method is applied to this network to identify a region of interest (ROI). The second neural network subsequently classifies the ROI for malignancy. While a prior method used simple gradient saliency maps to identify ROIs, a key enhancement of the present methodology is the application of Layered GradCam, which identifies cancerous areas more consistently and allows smaller ROIs. Layered GradCam is also used to display identified cancers to the user. By the AUC criterion, our model performs well, 0.974 for screen-detected and 0.931 for all cancers (screen-detected and interval), compared to a commercial program; 0.959 and 0.918, respectively. Comparisons with the radiologist scores indicate that the model has equal performance with two radiologists, and superior performance to one, for the detection of all cancers (screening- and interval type). Our tests indicate that our model generalizes well for different breast centers, but so far only images from a single manufacturer have been tested.

Identifying alternatives to acute hospital admission is a priority for many countries. Over 200 decentralised municipal acute units (MAUs) were established in Norway to divert low-acuity patients away from hospitals. MAUs have faced criticism for low mean occupancy and not relieving pressures on hospitals. We developed a discrete time simulation model of admissions and discharges to MAUs to test scenarios for increasing absolute mean occupancy. We also used the model to estimate the number of patients turned away as historical data was unavailable. Our experiments suggest that mergers alone are unlikely to substantially increase MAU absolute mean occupancy as unmet demand is generally low. However, merging MAUs offers scope for up to 20% reduction in bed capacity, without affecting service provision. Our work has relevance for other admissions avoidance units and provides a method for estimating unconstrained demand for beds in the absence of historical data.

Aims Studies with implantable cardiac monitors (ICMs) show that one-third of patients with cryptogenic stroke/transient ischaemic attack (TIA) have episodes of subclinical atrial fibrillation (SCAF) and benefit switching from antiplatelet- to anticoagulant therapy. However, ICMs are costly and resource demanding. We aimed to build a score based on participant’s baseline characteristics that could assess individual risk of SCAF. Methods and results In a prospective study, 236 eligible patients with a final diagnosis of cryptogenic stroke/TIA had an ICM implantated during the index hospitalization. Pre-specified evaluated variables were: CHA2DS2-VASc, P-wave duration, P-wave morphology, premature atrial beats (PAC)/24 h, supraventricular tachycardia/24 h, left atrial end-systolic volume index (LAVI), Troponin-T, NT-proBNP, and D-dimer. SCAF was detected in 84 patients (36%). All pre-specified variables were significantly associated with SCAF detection in univariate analysis. P-wave duration, followed by PAC/24 h, NT-proBNP, and LAVI, had the largest ratio of SCAF prevalence between its upper and lower quartiles (3.3, vs. 3.2, vs. 3.1 vs. 2.8, respectively). However, in a multivariate analysis, only PAC/24t, P-wave duration, P-wave morphology, and LAVIs remained significant predictors and were included in the PROACTIA score. Subclinical atrial fibrillation prevalence was 75% in the highest vs. 10% in the lowest quartile of the PROACTIA score with a 10-fold higher number of patients with an atrial fibrillation burden >6 h in the highest vs. the lowest quartile. Conclusion The PROACTIA score can identify patients with cryptogenic stroke/TIA at risk of subsequent SCAF detection. The large difference in SCAF prevalence between groups may provide a basis for future tailored therapy. Clinical trial registration Clinical Trial Registration: ClinicalTrials.gov; NCT02725944.

The Cancer Registry of Norway organises a population-based breast cancer screening program, where 250 000 women participate each year. The interpretation of the screening mammograms is a manual process, but deep neural networks are showing potential in mammographic screening. Most methods focus on methods trained from pixel-level annotations, but these require expertise and are time-consuming to produce. Through the screenings, image level annotations are however readily available. In this work we present a few models trained from image level annotations from the Norwegian dataset: a holistic model, an attention model and an ensemble model. We compared their performance with that of pretrained models based on pixel-level annotations, trained on international datasets. From this we found that models trained on our local data with image-level annotation gave considerably better performance than the pretrained models from external data, although based on pixel-level annotations.

Background All stroke patients should receive timely admission to a stroke unit (SU). Consequently, most patients with suspected strokes – including stroke mimics (SM) are admitted. The aim of this study was to estimate the current total demand for SU bed capacity today and give estimates for future (2020–2040) demand. Methods Time trend estimates for stroke incidence and time constant estimates for length of stay (LOS) were estimated from the Norwegian Patient Registry (2010–2015). Incidence and LOS models for SMs were based on data from Haukeland University Hospital (2008–2017) and Akershus University Hospital (2020), respectively. The incidence and LOS models were combined with scenarios from Statistic Norway’s population predictions to estimate SU demands for each health region. A telephone survey collected data on the number of currently available SU beds. Results In 2020, 361 SU beds are available, while demand was estimated to 302. The models predict a reduction in stroke incidence, which offsets projected demographic shifts. Still, the estimated demand for 2040 rose to 316, due to an increase in SMs. A variation of this reference scenario, where stroke incidence was frozen at the 2020-level, gave a 2040-demand of 480 beds. Conclusions While the stroke incidence is likely to continue to fall, this appears to be balanced by an increase in SMs. An important uncertainty is how long the trend of decreasing stroke incidence can be expected to continue. Since the most important uncertainty factors point toward a potential increase, which may be as large as 50%, we would recommend that the health authorities plan for a potential increase in the demand for SU bed capacity.

A flipped classroom, also known as flipped learning, is a teaching method in which students watch online lectures at home, followed by group work in the classroom. This study aimed to evaluate the efficacy of a flipped classroom vs. traditional lectures in a statistics and epidemiology course at Oslo Metropolitan University. The study used a pragmatic randomized controlled trial design in which one group of students received traditional lectures, while another group received flipped classroom teaching. Each participating student had previous experience with both teaching methods. No difference was found in exam grades between the two groups, but the students preferred the flipped classroom significantly (p = .008). Students who received instruction in the flipped classroom preferred this method to a higher degree than those who received traditional lectures (p = .018).

Loss of consciousness, so-called syncope, is a commonly occurring symptom associated with worse prognosis for a number of heart-related diseases. We present a comparison of methods for a diagnosis classification task in Norwegian clinical notes, targeting syncope, i.e. fainting cases. We find that an often neglected baseline with keyword matching constitutes a rather strong basis, but more advanced methods do offer some improvement in classification performance, especially a convolutional neural network model. The developed pipeline is planned to be used for quantifying unregistered syncope cases in Norway.