隐空间实验室 LatentLab

← All demos · Treatment Response (pCR) · Identity Recognition · Physiological Age

Risk Assessment: manipulated-image detection

Check whether a mammogram has been altered by a generative model that removed a lesion.

How it works

INPUT One mammogram DICOM / PNG / TIFF 16-bit preferred Prepare mirror: breast on left resize 1408 × 1728 scale to 0–1 Detector (GPU) ResNet-101 trunk + residual head → 1,024 numbers t-test against 0 0 real images centre on 0 OUTPUT p-value p < 0.05 →manipulated p ≥ 0.05 →no sign of editing

Our 2021 study showed that GAN-edited mammograms fool both AI models and radiologists. This detector maps the image (16-bit, 1408 × 1728) through a ResNet-101 trunk and a residual head to 1,024 numbers that are centred on zero for real images. A one-sample t-test against zero gives a p-value; p < 0.05 flags the image as manipulated.

Input

  • One mammogram: DICOM, or PNG/TIFF (16-bit preferred)

Output

  • p-value and verdict (manipulated if p < 0.05)
  • Distribution of the 1,024 detector outputs

Detector by Jintao Wang (M.S. 2024), trained for lesion-removal GAN attacks ("GAN2"). Research prototype: its accuracy on images from other sources has not been established. Typical run time: A few seconds.

Your data is not saved. Uploaded files are written only to a temporary folder on the lab GPU computer and deleted the moment the computation ends (or fails, or you cancel). Results are held in memory for 15 minutes so this page can show them, then discarded. Nothing is copied to our storage or logs.
Remove patient identity information before uploading. Strip names, IDs, birth dates, addresses, hospital and physician names from the files (for DICOM, use your PACS or a de-identification tool). This page checks DICOM headers in your browser before upload and warns you if such fields are still filled in. Research demo only — not for diagnosis or treatment decisions. Upload limit: 400 MB.

Run it

Click or drop one DICOM, PNG or TIFF file