🧠 Corrective Machine Unlearning for MRI Reconstruction

“AI that learns to forget — ensuring safety, trust, and adaptability in clinical diagnostics.”

🔬 Overview

This project explores the corrective machine unlearning paradigm to enhance the reliability and robustness of MRI reconstruction systems. Our methods target both privacy-preserving unlearning (e.g., “right to be forgotten”) and corrective forgetting of poisoned or faulty training data — without retraining from scratch.

Our pipeline includes:

• ⚙️ End-to-End Variational Network (E2E VarNet) for baseline MRI reconstruction
• 🧽 Selective Synaptic Dampening (SSD) as our primary unlearning strategy
• 🌀 CycleGAN for synthesizing MRI images with cysts from healthy images
• ⚔️ Novel adversarial training objective to perturb healthy MRIs and introduce synthetic tumor-like features for robustness and evaluation

---

🧩 Key Components

1. 🏗️ Corrective Unlearning

We implement several unlearning strategies:

• Selective Synaptic Dampening (SSD): Leverages Fisher Information to selectively dampen parameters influenced by corrupted data
• Bad Teacher Distillation
• Noisy Labeling
• Gradient Ascent-based Unlearning

The unlearnt model learns to discard adversarial patterns while retaining useful medical features.

2. 🌀 CycleGAN for Pathological Translation

We train a CycleGAN to:

• Convert healthy brain MRIs to cyst-injected versions
• Provide a data augmentation path and simulate real-world pathological variation
• Evaluate whether reconstruction models hallucinate structures when faced with such transformations

3. ⚔️ Adversarial Objective for Tumor Induction

A novel adversarial objective generates perturbed versions of healthy MRIs:

• Introduces tumor-like patterns while preserving anatomical plausibility
• Acts as an evaluation and stress-testing mechanism for reconstruction models
• Assesses robustness and hallucination resistance

---

📊 Datasets

We utilize:

• M4RAW — High-resolution anatomical k-space data for training E2E VarNet
• SKM-TEA — Multi-contrast T2 mapping MRI data for cross-modal learning
• CMRxRecon — Cardiac MR dataset for generalization and benchmark testing -- BraTS - Brain and Tumor Segmentation dataset. -- ExBox - MRI Motion and metal object artefact dataset.

---

📈 Results

Metric	Original Model	SSD (Forget Set)	SSD (Retain Set)
SSIM	0.44	0.0967	0.3824
PSNR	26.92 dB	13.97 dB	22.16 dB
NMSE	0.0687	1.6210	0.1807

🔍 Qualitative results show that reconstructions from the unlearnt model resemble clean data, effectively reversing poisoning effects.

---

🧪 Getting Started

git clone https://github.com/<your-org>/CorrectiveUnlearningMRI.git
cd CorrectiveUnlearningMRI

# Setup environment
conda create -n unlearning_env python=3.8
conda activate unlearning_env
pip install -r requirements.txt

To train the E2E VarNet:

python train_varnet.py --dataset M4RAW --epochs 100

To perform corrective unlearning with SSD:

python unlearning_.py --alpha 0.2 --lambda 0.1

To run the CycleGAN:

python cyclegan_train.py --dataset_path ./data/healthy_to_cyst/

---

🎥 Visual Summary

• 🧠 Baseline and Unlearnt Reconstructions
• 📊 SSD Performance Landscape
• 🔄 CycleGAN Transformations
• ⚔️ Tumor Perturbation Examples

---

✍️ Authors

• Aryaman Bahl
• Chinmay Sharma
• Sairam Babu

---