Local Label-Informed Feature Transfer for Generating Ground-Truth Medical Images: A Comparison of GAN- and Diffusion-Based Approaches

2026-07-21

AI Summary

This work addresses a critical gap in explainable artificial intelligence (XAI) for medical imaging: the absence of clinically realistic ground-truth datasets with known lesion locations. To this end, the authors propose LLIFT, a novel framework that generates spatially controllable and structurally realistic semi-synthetic brain MRI lesion images using only image-level labels and user-specified regions, eliminating the need for pixel-level annotations. LLIFT is instantiated in two variants: LLIFT-GAN, based on a custom generative adversarial network, and LLIFT-DM, leveraging a ControlNet-augmented diffusion model. Experiments on Human Connectome Project data demonstrate that the generated images achieve Fréchet Inception Distance (FID) scores comparable to the inter-class reference values between healthy and pathological scans, while qualitative assessments confirm their clinical plausibility. This provides the first spatially precise ground-truth benchmark for evaluating XAI methods in neuroimaging.

Abstract

Validating Explainable Artificial Intelligence (XAI) methods in medical imaging requires ground-truth data with known locations of informative features. However, current approaches rely on expert annotations, which are prone to labeling errors, or on hand-crafted artificial perturbations superimposed onto healthy images to mimic lesions or malignant features, which lack clinical realism. We present Local Label-Informed Feature Transfer (LLIFT), a framework for generating semi-synthetic brain magnetic resonance images with realistic lesions placed in user-controlled regions, which does not require pixel-level lesion annotations during training. We implement LLIFT with two generative paradigms: LLIFT-GAN, a custom GAN that learns pathological features from binary class labels alone, and LLIFT-DM, a diffusion-based inpainting pipeline conditioned on bounding-box masks via ControlNet. Both approaches are evaluated on brain magnetic resonance imaging data derived from the Human Connectome Project. In evaluations, both achieve Fréchet Inception Distance scores, with respect to the real pathological distribution, that are comparable to the inter-class reference between healthy and pathological images in the given dataset. Furthermore, qualitative inspection confirms the realism of lesion structures. The resulting benchmark datasets provide spatially controlled ground truth data for evaluating XAI methods in medical imaging.

Problem

• Explainable Artificial Intelligence
• Medical Imaging
• Ground-Truth Data
• Lesion Realism
• Label Errors

Innovation

• Local Label-Informed Feature Transfer
• semi-synthetic medical image generation
• label-only supervised generation
• diffusion-based inpainting
• ControlNet