Rescind: Countering Image Misconduct in Biomedical Publications with Vision-Language and State-Space Modeling

TL;DR

This work tackles the problem of image misconduct in biomedical publications by introducing Rescind, a large-scale, semantically controlled forgery dataset generated via vision-language prompts and diffusion-based synthesis, and Integscan, a structured state-space model with prompt-conditioned fusion for precise forgery localization. A VLM-based verification loop ensures semantic fidelity of generated forgeries, enhancing dataset quality and realism. Empirical results show state-of-the-art performance in both detection and localization across diverse biomedical modalities, supported by robustness analyses and comprehensive ablations. Collectively, the approach provides a scalable, interpretable foundation for automated scientific integrity analysis and practical tools for journals and researchers to assess and mitigate image manipulation in biomedical literature.

Abstract

Scientific image manipulation in biomedical publications poses a growing threat to research integrity and reproducibility. Unlike natural image forensics, biomedical forgery detection is uniquely challenging due to domain-specific artifacts, complex textures, and unstructured figure layouts. We present the first vision-language guided framework for both generating and detecting biomedical image forgeries. By combining diffusion-based synthesis with vision-language prompting, our method enables realistic and semantically controlled manipulations, including duplication, splicing, and region removal, across diverse biomedical modalities. We introduce Rescind, a large-scale benchmark featuring fine-grained annotations and modality-specific splits, and propose Integscan, a structured state space modeling framework that integrates attention-enhanced visual encoding with prompt-conditioned semantic alignment for precise forgery localization. To ensure semantic fidelity, we incorporate a vision-language model based verification loop that filters generated forgeries based on consistency with intended prompts. Extensive experiments on Rescind and existing benchmarks demonstrate that Integscan achieves state of the art performance in both detection and localization, establishing a strong foundation for automated scientific integrity analysis.

Figures (6)

• Figure 1: Forgery synthesis in BioFors vs. Rescind across biomedical modalities: macroscopy, microscopy, FACS, and blot/gel. BioFors uses rectangular insertions lacking realism; Rescind leverages MedSAM ma2024segment for irregular mask generation and prompt-guided diffusion for semantically consistent edits. Red = forged; blue = pristine. Eg. Microscopy Prompt: "Duplicate the fluorescent cell cluster from the highlighted region to simulate manipulation."
• Figure 2: Rescind-V Curation Pipeline. Given a pristine biomedical image, BiomedCLIP zhang2024biomedclip generates a modality-aware caption (e.g., macroscopy image). MedSAM ma2024segment produces a binary segmentation mask that identifies regions of interest. A Forgery Generator creates a Rescind-I forged image and corresponding forgery mask by duplicating or manipulating content within the masked region. A Prompt Generator formulates a semantic forgery prompt based on the predicted modality, explicitly referencing the MedSAM-highlighted region. Finally, a Stable Diffusion inpainting model Rombach_2022_CVPR synthesizes the Rescind-V forged image using the forgery mask and forgery prompt as conditional inputs.
• Figure 3: Integscan Model Architecture Overview
• Figure 4: (a) IntegSSM with state-space modeling; (b) INTEG_ATTN with rotary encoding, channel-aware attention.
• Figure 5: Semantic Consistency Score across forgery types, using LLaVA-Med on Rescind-I, Rescind-G, Rescind-V.