MD Mizanur Rahman

Work · Medical vision

Attribution sanity lab

  • Demo ready
  • Synthetic demonstration
  • Independent implementation

Grad-CAM, occlusion and parameter-randomisation checks that show when an attribution map changes and when its interpretation is not justified.

Repository on GitHub

Problem

An attribution map that stays the same after the model's weights are randomised is not explaining the model. The lab measures that.

Contribution

Relationship to research: Independent implementation. Related to the coauthored fingerprint-classification article only as computer-vision context; that work is biometrics, not diagnostic imaging.

Account owner’s role: Commissioned, reviewed and published the implementation. The code was produced with AI-assisted generation on commission and reviewed before release; it is not the original code of any cited publication.

What was built

  • Hook-based Grad-CAM
  • Occlusion, blur and brightness perturbations with recorded output deltas
  • Cascading parameter randomisation with map correlations

Method

Maps are always generated from the selected model and class; normalisation and colour scale are documented.

Data and access

Synthetic planted-shape images; an optional local MedMNIST loader.

Evaluation protocol

Actual output changes under perturbation and map correlation under randomisation, reported as computed.

Results

No measured result is published for this project. Demonstration outputs from the repository’s own synthetic fixtures are labelled as such in the repository and are not presented here as results.

Limitations

  • Attribution is not a segmentation or a clinical explanation.
  • No localisation labels exist, so no localisation accuracy is claimed.

Reproducibility

Offline; the built-in model is untrained unless a checkpoint is supplied.

Repository metadata: Python, MIT, last push 2026-09-13, 0 stars.

Source publication and links