💻 computer science

Seeing What Shouldn't Be There: Counterfactual GANs for Medical Image Attribution

This paper proposes a counterfactual GAN-based method for medical image attribution that generates plausible counterfactual instances to provide self-explanatory, analogy-based insights into classification decisions, addressing the limitations of existing discriminative visualization techniques.

Original authors: Shakeeb Murtaza

Published 2026-05-08

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Shakeeb Murtaza

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a doctor looking at an X-ray of a patient's lungs. A super-smart computer program (an AI) looks at the same X-ray and says, "This patient has Tuberculosis."

Here is the problem: The computer is a "black box." It gives you the answer, but it won't tell you why. It's like a magician pulling a rabbit out of a hat but refusing to show you the trick. In the medical world, doctors can't just trust a magic answer; they need to know what in the image made the computer say "disease."

This paper proposes a new way to open that black box using a concept called Counterfactual Explanation.

The "What If" Game

Think of a counterfactual explanation as a "What If" game.

The Question: "Why did the computer say this patient is sick?"
The Answer: "Because of these specific spots in the image."
The Counterfactual: "If we magically removed those specific spots, the computer would say the patient is healthy."

The paper argues that simply pointing a red marker at the "sick" spots (which older methods do) isn't enough. It's better to show the doctor a side-by-side comparison:

The Real Image: The patient's actual X-ray with the disease.
The "Healthy" Twin: A generated image that looks exactly like the patient's, but without the disease.

By comparing the two, the doctor can see exactly what changed. The difference between the "Sick" image and the "Healthy" twin is the Explanation.

The Problem with Old Methods

The author explains that previous attempts to create these "Healthy Twins" were like a clumsy painter.

The Old Way: Imagine trying to fix a photo of a broken vase by just pasting a square of a "perfect vase" over the broken part. The result looks fake, weird, and doesn't match the rest of the photo. In medical terms, these old methods created "implausible" images that looked like noise or glitches, confusing the doctor rather than helping them.
The Issue: They didn't have a perfect pair of "Sick" and "Healthy" photos to learn from (you can't take an X-ray of the same patient twice, once sick and once healthy).

The New Solution: The "Magic Mirror" (CX-GAN)

The author built a new system called CX-GAN. Think of this system as a highly skilled, magical mirror with two special functions:

The Healer (Generator 1): This part looks at a "Sick" image and tries to paint a "Healthy" version of it. It learns to remove the disease while keeping the rest of the anatomy (the ribs, the lungs' shape) looking natural and realistic.
The Detective (Generator 2): This part creates a "Change Map." It's like a stencil that highlights only the pixels that were different between the Sick and Healthy versions.

How they work together:
The system uses a "Cycle" trick. It takes a Sick image, turns it into a Healthy one, and then tries to turn that Healthy one back into the original Sick image. If it can't get back to the original perfectly, it knows it made a mistake and fixes its painting technique. This ensures the "Healthy Twin" looks believable and real, not like a glitchy mess.

The Results: A Better Explanation

The author tested this "Magic Mirror" on three things:

Fake Data: Simple computer-generated circles to see if the logic worked.
Tuberculosis X-rays: Real chest scans from patients.
Brain Tumors: MRI scans of brains.

The Findings:

Old Methods (like CAM/Grad-CAM): These just highlighted a blurry blob. Sometimes they pointed at the wrong area, or they highlighted too much of the image, making it hard to tell exactly what the disease looked like.
The New Method (CX-GAN): It produced a clean, realistic "Healthy Twin." When you subtract the Healthy Twin from the Sick image, the result is a clear, sharp map showing exactly where the disease is. It didn't add random noise or fake details.

The Bottom Line

The paper claims that by generating a believable "Healthy Twin" and showing the difference map, doctors get a much clearer, self-explanatory reason for why the AI made its diagnosis. It moves beyond just "pointing" at a problem to showing "what the world would look like if the problem didn't exist," making the AI's decision much easier for humans to trust and understand.

Technical Summary: Counterfactual GANs for Medical Image Attribution

Problem Statement
The paper addresses the "black box" nature of deep learning models, particularly in safety-critical medical image diagnostics (MID). While existing visualization techniques (e.g., CAM, Grad-CAM) highlight discriminative regions used by classifiers, they often fail to account for all noticeable objects and rely on minimal feature sets, potentially leading to misleading explanations. Furthermore, existing Counterfactual Explanation (CX) methods often generate implausible counterfactual instances (CIs). These methods typically replace image regions with arbitrary values or require paired datasets (input and counterfactual pairs), which are rarely available in medical domains (e.g., a healthy scan of the exact same patient). The lack of semantically plausible CIs hinders the ability to provide self-explanatory, analogy-based reasoning (e.g., "if this tumor were not present, the scan would look like X").

Methodology
The authors propose CX-GAN, an integrated model based on Generative Adversarial Networks (GANs) designed to jointly learn the generation of plausible CIs and the corresponding Counterfactual Explanations (CXs) without requiring paired training data.

Architecture: The model utilizes a Cycle-Consistent GAN (CycleGAN) framework extended with two generators and two discriminators:
- Generator $G$ : Translates an input image (e.g., anomalous/tumorous) to a counterfactual domain (e.g., normal/healthy), producing a plausible CI.
- Generator $G_M$ : Generates a Change Map (the CX). This map represents the specific differences between the input and the generated CI. When added to the CI, it reconstructs the original input.
- Discriminators ( $D_x, D_y$ ): Ensure the generated images and maps belong to their respective real data distributions.
Loss Functions: The model is optimized using a composite loss function:
- Adversarial Loss ( $L_{GAN}$ ): Ensures the generated CIs and change maps are indistinguishable from real data in their respective domains.
- Cycle-Consistency Loss ( $L_{cyc}$ ): Enforces that translating an input to a CI and back (or adding the change map to the CI) reconstructs the original image, ensuring semantic alignment and preventing irrelevant transformations.
- Integrated Objective: The total loss combines adversarial, change map, and cycle-consistency terms, allowing the network to learn the translation and the explanation simultaneously.
Network Design: The generators employ a shallow U-Net architecture with skip connections to preserve localization and contextual information, while the discriminators use standard CNN structures.

Key Contributions

Plausible CI Generation: The paper poses and solves the problem of generating semantically plausible counterfactual instances for medical images without requiring paired input-output datasets.
Integrated CX Method: It proposes a unified framework that generates both the counterfactual image and the specific "change map" (attribution) explaining the difference, avoiding the cascaded errors of two-step approaches.
Evaluation Metric: A novel technique is introduced to evaluate the quality of generated CIs by measuring "non-resemblance" scores between the original and counterfactual regions, ensuring the changes are meaningful.
Benchmarking: The method is evaluated on synthetic data, a Tuberculosis chest X-ray dataset, and the BraTS (Brain Tumor) dataset, providing baseline results against existing methods like CAM, Grad-CAM, and VA-GAN.

Results
The proposed CX-GAN was evaluated quantitatively using Dice scores, Intersection over Union (IoU), and Normalized Cross Correlation (NCC), as well as qualitatively through visual inspection.

Synthetic Data: CX-GAN achieved superior performance with an IoU of 91.4 and a Dice score of 95.5, significantly outperforming VA-GAN (IoU 87.2) and discriminative methods (CAM/Grad-CAM).
Medical Datasets: On the BraTS dataset, CX-GAN achieved an IoU of 91.4 and a Dice score of 94.0, outperforming VA-GAN (IoU 89.5, Dice 93.2). On the Tuberculosis dataset, it achieved an IoU of 74.73 and Dice of 81.47.
Qualitative Findings: Visual results indicate that discriminative methods (CAM/Grad-CAM) often highlight minimal or overly broad regions, while VA-GAN introduces noise and false positives due to unaligned mappings. In contrast, CX-GAN produces fine-grained, noise-free change maps that accurately localize abnormalities (e.g., tumors) while generating plausible healthy counterparts.

Significance and Claims
The paper claims that accompanying CXs with believable CIs provides self-explanatory, analogy-based explanations crucial for medical professionals. By generating a "what it should look like" (CI) alongside the "what it looks like" (input) and the "difference" (CX), the method offers a causal reasoning process ("if X had not happened, Y would not have happened"). The authors emphasize that this approach overcomes the limitations of existing CX techniques that produce implausible instances or require unavailable paired data, thereby enhancing the interpretability and trustworthiness of deep learning models in medical diagnostics. The study also highlights the limitations of current state-of-the-art CX methods in producing plausible instances, positioning the proposed integrated GAN approach as a necessary advancement for safety-critical applications.