Two Visual Intelligence-authored papers accepted for MICCAI 2026

Visual Intelligence will be well represented at MICCAI 2026, one of the leading AI conferences on medical imaging and computer assisted intervention, with two accepted research papers.

By Petter Bjørklund, Communications Officer at SFI Visual Intelligence

The annual International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI) attracts the world's leading biomedical scientists, engineers, and clinicians from a wide range of disciplines associated with medical imaging and computer assisted intervention. MICCAI has an historical acceptance rate of 30 per cent.

Longitudinal breast cancer risk prediction model

The first paper, titled "Longitudinal Multi-View Modeling for Breast Cancer Risk Prediction", proposes a longitudinal multi-view breast cancer risk prediction model, termed the LMV-Net. The model jointly analyzes anatomically complementary views across multiple screening time points to effectively capture spatial–temporal risk cues.

The authors evaluate this approach on the public EMBED and CSAW-CC datasets, comparing it to state-of-the-art breast cancer risk prediction methods. The model consistently outperforms existing approaches in overall risk prediction performance and across different breast density and cancer subgroups.

Importantly, these improvements highlight the potential of longitudinal multi-view modeling to enhance risk stratification, paving the way for future work on personalized screening, earlier identification of high-risk patients, and more efficient allocation of screening resources.

The paper is authored by centre researchers Solveig Thrun, Zijun Sun, Suaiba Salahuddin, Elisabeth Wetzer, Robert Jenssen, and Michael Kampffmeyer. The paper is also co-authored by Stine Hansen,

Controlled evaluation framework for medical vision-language models

The second paper, titled "Beyond Clean Test Sets: Spurious Correlations in Medical Vision-language Models and the Role of Concept Supervision", proposes a controlled evaluation framework that uses synthetic artifacts modelled after common acquisition confounders to parametrically vary correlation strength. It also pairs two complementary test protocols — artifact removal and artifact inversion — to isolate whether models rely on clinical features or visual shortcuts.

The framework was applied to diabetic retinopathy grading in fundus photography and BI-RADS-based assessment in mammography. Through this, they evaluate five architectures spanning a spectrum from no-concept supervision to full multi-level image–concept alignment.

The findings show that VLM backbones retain clinical signal when shortcuts are absent, yet actively follow spurious associations when they conflict with pathology, degrading faster than standard visual backbones — a dual encoding that is only exposed when evaluation goes beyond clean test sets.

Among concept-based strategies, only architectures that both reshape the feature space toward clinical concepts and shield the classifier from non-clinical signal provide meaningful resilience. The framework is architecture-agnostic and applicable to any vision or multimodal model.

The paper is authored by Valentina Corbetta, PhD Candidate at Maastricht University and a former Guest Researcher at Visual Intelligence, as well as centre researchers Kristoffer Wickstrøm, Elisabeth Wetzer, Veronica Lachi and Robert Jenssen. The paper is also co-authored by Antonio Portaluri (Radboudmc), Michiel Van der Heijden (Netherlands Cancer Institute (NKI)), D. Boeke (NKI), Regina Beets-Tan (NKI), and Wilson Silva (Utrecht University).