PathEdit: Diagnostic-Preserving Counterfactual Medical Image Editing through State-Factorized Generative Intervention
Counterfactual medical images are useful only when a requested clinical change is introduced without silently altering patient characteristics that should remain fixed. Existing diffusion editors can produce convincing images, but realism or target-classifier flips do not establish that an edit is localized, diagnosis-specific, anatomically faithful, or useful on real clinical data. We introduce…
The article introduces PATHEDIT, a new framework for editing medical images that preserves critical diagnostic information while changing only targeted anatomical features. Existing diffusion editors can produce realistic images, but they often alter too much of the data, making the edits unreliable for real-world clinical use. PATHEDIT addresses this by breaking down an abdominal CT scan into its four core components: pathology, anatomy, acquisition, and residual context.
It then applies targeted edits to only the pathology factor, copying the other factors and penalizing any unwanted changes. The edited state is then used to generate a new image, report, and diagnostic readouts. The framework is evaluated using four key properties: whether the desired change is localized, whether other features are preserved, whether the anatomy remains intact, and whether the edited image is useful in real-world clinical settings.
PATHEDIT achieves a 94.2% success rate in targeting the intended change while reducing off-target diagnostic changes to just 3.0%. It also aligns better with observed lesion changes, improves rare-lesion recognition, and is more robust to shortcuts in training data. These results suggest that a more stringent approach to counterfactual medical imaging is needed - one that not only shows what has changed but also what has remained constant.
Written by urgent.news from bioRxiv's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
