Learning to Exploit Temporal Structure for Biomedical Vision-Language Processing

TL;DR

BioViL-T tackles the limitation of conventional vision-language pre-training that ignores longitudinal information in biomedical data. It introduces a multi-image encoder and a text model trained with image-guided MLM and contrastive objectives to exploit prior images and prior reports, enabling robust static and temporal VLP. The approach achieves state-of-the-art results on pneumonia detection, phrase grounding, temporal image classification, and report generation, and it introduces the MS-CXR-T benchmark to quantify temporal semantics. The work demonstrates data-efficient learning and broader applicability of temporal context in clinical VLP, with public release of model weights and the MS-CXR-T dataset.

Abstract

Self-supervised learning in vision-language processing exploits semantic alignment between imaging and text modalities. Prior work in biomedical VLP has mostly relied on the alignment of single image and report pairs even though clinical notes commonly refer to prior images. This does not only introduce poor alignment between the modalities but also a missed opportunity to exploit rich self-supervision through existing temporal content in the data. In this work, we explicitly account for prior images and reports when available during both training and fine-tuning. Our approach, named BioViL-T, uses a CNN-Transformer hybrid multi-image encoder trained jointly with a text model. It is designed to be versatile to arising challenges such as pose variations and missing input images across time. The resulting model excels on downstream tasks both in single- and multi-image setups, achieving state-of-the-art performance on (I) progression classification, (II) phrase grounding, and (III) report generation, whilst offering consistent improvements on disease classification and sentence-similarity tasks. We release a novel multi-modal temporal benchmark dataset, MS-CXR-T, to quantify the quality of vision-language representations in terms of temporal semantics. Our experimental results show the advantages of incorporating prior images and reports to make most use of the data.

Figures (10)

• Figure 1: (a) Existing visual--language pre-training approaches boecking2022makinghuang2021gloriazhang2020contrastive often use only a single image for contrastive learning (e.g., InfoNCE oord2018representation). (b) In such settings, discarding the temporal connectivity of images limits the alignment of image--text pairs as shown with the affinity matrix, leading to suboptimal pre-training and missed opportunity to create additional model supervision for free. (c, d) Our approach exploits this domain knowledge by learning to incorporate a series of images and correlate them to reports, leading to pre-trained models that can generalise to a wider range of downstream tasks whilst achieving SOTA performance.
• Figure 2: The proposed self-supervised VLP training framework BioViL-T: Image representations $\mathbf{V}$ are extracted from single and multiple input scans (whenever available) using a hybrid CNN and transformer encoder park2022howd2021convit. This design choice is to increase the data-efficiency and enable the fusion of temporal content without requiring image registration. They are later matched with their corresponding text representations obtained with CXR-BERT boecking2022making using local huang2021gloria and global InfoNCE oord2018representation training objectives. As an additional model supervision, multi-modal fused representations, obtained with cross-attention, are used for image-guided masked language modelling.
• Figure 3: Attention rollout maps abnar-zuidema-2020-quantifying from the reference patch (marked with $\filledstar$) to the current and prior images. The bounding boxes, annotated by a radiologist, show the extent of consolidation. Note that the reference patch attends to its anatomical neighbourhood in the prior image despite the misalignment between prior and current images. The grid ($14 \times 14$) represents the patch tokens processed in the transformer encoder blocks.
• Figure 4: Mean token-level increase in image-guided MLM loss when prior image is discarded, grouped by token category. The prior image is excluded during inference to measure its impact on masked token predictions. Progression tokens are significantly better predicted when prior images are incorporated into image embeddings. The top five Progression tokens are 'persist', 'improving', 'remains', 'unchanged', and 'residual'.
• Figure A.1: Cross entropy between model predictions and MS-CXR-T temporal classification labels. 'Disagreement' indicates cases for which annotations differed amongst radiologists. Model performance is higher for cases with with low ambiguity ('Agreement').