Attention Consistent Longitudinal Medical Visual Question Answering Guided by Vision Foundation Models

A new machine-learning architecture designed to answer questions about chest X-rays taken at different points in time has been proposed, using a combination of image alignment and attention-guided masks to track anatomical changes, according to a preprint submitted on 3 June 2026 [1]. The model targets longitudinal medical visual question answering (VQA), a task that requires reasoning about differences between a current medical image and an earlier reference image [2]. The authors introduce a lightweight affine registration module to co-register the two images, aiming to reduce nuisance motion before analysis [2]. The registered pair is then processed by an image encoder, followed by a frozen DINO-based mask generator and a trainable adaptive mask generator that produce saliency masks applied to the original images [2]. The masked image pairs are re-encoded and combined with text features before being passed to a multimodal transformer-based decoder that generates the final answer [2]. To stabilize training and sharpen the change signal, the framework incorporates auxiliary objectives inspired by DINO-v3: a mask rebuilding loss, a pairwise Gram-style consistency loss, and a KoLeo uniformity loss that shapes the geometry of the learned representation [2]. The work was posted on arXiv, an open-access repository that hosts preprints across physics, computer science, and related fields without peer review [9]. The repository, launched in 1991, now receives roughly 24,000 submissions per month [9]. On the Medical-Diff-VQA benchmark, the model delivered strong BLEU, ROUGE-L, CIDEr, and METEOR scores while offering intrinsic interpretability through the shared saliency mask [2]. The authors argue that saliency-conditioned generation with mild pre-alignment constitutes a principled framework for longitudinal reasoning in medical imaging [2]. The training strategy also illustrates a broader paradigm for using image foundation models in biomedicine, optimizing supervised and unsupervised learning objectives simultaneously [2]. The preprint appears within the Electrical Engineering and Systems Science section of arXiv, under the Image and Video Processing sub-category [1]. The work is associated with arXivLabs, a framework that lets collaborators develop and share experimental features on the platform under community values of openness and user data privacy [1].