Alba Ordoñez

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.

This study investigates the applicability of the Mammo-CLIP Dissect framework from Salahuddin et al. for concept-based explainability in deep learning models for mammography. We first reproduced key results from the original paper using the Mammo-CLIP image encoder and a curated mammography concept vocabulary, confirming the expected layer-wise emergence of clinically meaningful concepts. We then applied the framework to an in-house ResNet101 classifier developed within the AIforscreening project. Compared with Mammo-CLIP, the ResNet101 model exhibited lower semantic alignment and a higher prevalence of non-mammography concepts, reflecting differences in the absence of multimodal training. These findings suggest that models trained solely on images may provide less interpretable explanations for clinicians than multimodal vision-language models. We highlight the importance of jointly considering accuracy and interpretability, noting that model performance was not evaluated on the probe set in this study. Future work includes applying the framework to Cancer Registry data and exploring multimodal training for improved clinical relevance.

Chain-of-Thought (CoT) prompting has emerged as a simple yet powerful strategy to elicit structured reasoning in large language models. This study investigates how CoT prompting influences reasoning behavior and task performance of large/vision–language models (LLMs/VLMs) applied to Earth Observation (EO). We compare an EO-specialized model (Falcon) with three general-purpose models (LLaVA, LLaVA-CoT, and o3) across two datasets: RSVQAxBEN, a large open EO benchmark, and a proprietary aerial dataset from Narvik. Experiments contrast baseline and CoT-style prompts to assess both factual accuracy and reasoning quality, complemented by an LLM-as-judge evaluation. Results show that CoT prompting benefits only the large-scale o3 model, while smaller or mid-scale models experience degraded accuracy—confirming that effective reasoning is an emergent property of scale. CoT adds transparency by revealing how models reason, though its outputs can still be partly opaque due to safety or internal constraints. On Narvik, o3 generalizes well to unseen EO data, but CoT prompting does not improve quantitative accuracy. These findings suggest that CoT currently offers greater value for interpretability than for performance. Future work should explore inference-time perception–reasoning strategies—where an EO model like Falcon provides scene-level facts that guide o3’s reasoning—to improve both trustworthiness and accuracy without retraining.

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.

The age determination of fish is fundamental to marine resource management. This task is commonly done by analysis of otoliths performed manually by human experts. Otolith images from Greenland halibut acquired by the Institute of Marine Research (Norway) were recently used to train a convolutional neural network (CNN) for automatically predicting fish age, opening the way for requiring less human effort and availability of expertise by means of deep learning (DL). In this study, we demonstrate that applying a CNN model trained on images from one lab (in Norway) does not lead to a suitable performance when predicting fish ages from otolith images from another lab (in Iceland) for the same species. This is due to a problem known as dataset shift, where the source data, i.e., the dataset the model was trained on have different characteristics from the dataset at test stage, here denoted as target data. We further demonstrate that we can handle this problem by using domain adaptation, such that an existing model trained in the source domain is adapted to perform well in the target domain, without requiring extra annotation effort. We investigate four different approaches: (i) simple adaptation via image standardization, (ii) adversarial generative adaptation, (iii) adversarial discriminative adaptation and (iv) self-supervised adaptation. The results show that the performance varies substantially between the methods, with adversarial discriminative and self-supervised adaptations being the best approaches. Without using a domain adaptation approach, the root mean squared error (RMSE) and coefficient of variation (CV) on the Icelandic dataset are as high as 5.12 years and 28.6%, respectively, whereas by using the self-supervised domain adaptation, the RMSE and CV are reduced to 1.94 years and 11.1%. We conclude that careful consideration must be given before DL-based predictors are applied to perform large scale inference. Despite that, domain adaptation is a promising solution for handling problems of dataset shift across image labs.

Every year, marine scientists around the world read thousands of otolith or scale images to determine the age structure of commercial fish stocks. This knowledge is important for fisheries and conservation management. However, the age-reading procedure is time-consuming and costly to perform due to the specialized expertise and labor needed to identify annual growth zones in otoliths. Effective automated systems are needed to increase throughput and reduce cost. DeepOtolith is an open-source artificial intelligence (AI) platform that addresses this issue by providing a web system with a simple interface that automatically estimates fish age by combining otolith images with convolutional neural networks (CNNs), a class of deep neural networks that has been a dominant method in computer vision tasks. Users can upload otolith image data for selective fish species, and the platform returns age estimates. The estimates of multiple images can be exported to conduct conclusions or further age-related research. DeepOtolith currently contains classifiers/regressors for three fish species; however, more species will be included as related work on ageing will be tested and published soon. Herein, the architecture and functionality of the platform are presented. Current limitations and future directions are also discussed. Overall, DeepOtolith should be considered as the first step towards building a community of marine ecologists, machine learning experts, and stakeholders that will collaborate to support the conservation of fishery resources.

Fish stock assessment and management requires accurate estimates of fish abundance, which are typically derived from echosounder observations using acoustic target classification (ATC). Skilled operators are regularly assisted in classifying acoustic targets by software and there has been an increasing interest toward using machine learning to create improved tools. Recent studies have applied deep learning approaches to acoustic data, however, algorithm data-preparation strategies (influencing model output) are presently poorly understood and standardization is needed to enable collaborative research and management. For example, a common pre-processing technique is to resample backscatter data coming from echosounder measurements from the original resolution to a coarser resolution in the horizontal (time) and vertical (range) directions. Using data values derived from the volume backscattering coefficient obtained during the Norwegian sandeel survey, we investigate which resampling resolutions are suitable for ATC using a convolutional neural network trained to classify single values of backscatter data. This process is known as pixel-level semantic segmentation. Our results indicate that it is possible to downsample the data if important information related to acoustic characteristics is not smoothed out. We also show that the classification performance is improved when providing the network with contextual information relating to range. These findings will provide input to fisheries acoustic data standards and contribute to the on-going development of automated ATC methods.

This report documents a workshop organised by the COGMAR and CRIMAC projects. The objective of the workshop was twofold. The first objective was to give an overview of ongoing work using machine learning for Acoustic Target Classification (ATC). Machine learning methods, and in particular deep learning models, are currently being used across a range of different fields, including ATC. The objective was to give an overview of the status of the work. The second objective was to familiarise participants with machine learning background to fisheries acoustics and to discuss a way forward towards a standard framework for sharing data and code. This includes data standards, standard processing steps and algorithms for efficient access to data for machine learning frameworks. The results from the discussion contributes to the process in ICES for developing a community standard for fisheries acoustics data.

Recent research published by Li and Abubakar (2020) investigated if the sediment interpretation could be done based on an idea inspired from video analysis, without involving human labelling. The aim of the study was understand how the proposed approach worked and investigate whether we could use it for detecting discontinuities in seismic structures.

The aim of the study was to investigate whether context information related to acoustic data resolution could improve automatic acoustic classification of species using echosounder observations.

During the presentation, we investigated how a model trained for automatic fish age prediction on otolith images from one lab could generalize to novel images from another lab. We identified Unsupervised Domain Adaptation based on coGAN (Liu and Tuzel, 2016) and Generate to Adapt (Sankaranarayanan et al., 2018) frameworks as promising methods.

Age-reading of fish otoliths (ear stones) is important for the sustainable management of fish resources. However, the procedure is challenging and requires experienced readers to carefully examine annual growth zones. In a recent study, convolutional neural networks (CNNs) have been demonstrated to perform reasonably well on automatically predicting fish age from otolith images. In the present study, we carefully investigate the prediction rule learned by such neural networks to provide insight into the features that identify certain fish age ranges. For this purpose, a recent technique for visualizing and analyzing the predictions of the neural networks was applied to different versions of the otolith images. The results indicate that supplementary knowledge about the internal structure improves the results for the youngest age groups, compared to using only the contour shape attribute of the otolith. However, the contour shape and size attributes are, in general, sufficient for older age groups. In addition, within specific age ranges we find that the network tends to focus on particular areas of the otoliths and that the most discriminating factors seem to be related to the central part and the outer edge of the otolith. Explaining age predictions from otolith images as done in this study will hopefully help build confidence in the potential of deep learning algorithms for automatic age prediction, as well as improve the quality of the age estimation.