Scribd
Upload
10 views1 page

Synopsis

Uploaded by

aninditasinha23
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as DOCX, PDF, TXT or read online on Scribd
10 views1 page

Synopsis

Uploaded by

aninditasinha23
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as DOCX, PDF, TXT or read online on Scribd
You are on page 1/ 1

Proposed Methodology for Anomaly Detection in Medical Images using

GANs
This paper highlights the limitations of current unsupervised Deep Learning (DL) based Anomaly Detection (AD)
techniques for medical imaging. It emphasizes the need for more robust and generalizable models. Here's a proposed
methodology based on the paper's findings:

Data Collection and Preprocessing:

1. Select diverse medical image datasets: Choose datasets with various modalities (CT, MRI, Mammography),
organs (brain, breast), and anomaly types. This helps assess model performance across different scenarios.
2. Data Augmentation: Augment the training data with techniques like flipping, rotation, and noise addition. This
helps improve model generalizability and prevent overfitting.
3. Patch Extraction (Optional): For high-resolution images, consider extracting patches around regions of interest
(ROIs) identified by experts. This might improve model performance for localized anomalies.

Model Selection and Training:

1. Choose a GAN-based AD model: Consider models like f-AnoGAN or GANomaly, which utilize both a
generator and discriminator network.

Explore other recent advancements in GAN-based AD as well.

2. Train the model on normal data: Train the model using only images with normal anatomy. The goal is for the
model to learn the "normal" distribution of the data.
3. Loss Function: Utilize a combination of loss functions like adversarial loss, contextual loss, and encoder loss
(depending on the chosen model) to optimize training.

Evaluation and Refinement:

1. Evaluate on unseen data: Test the model on a separate dataset containing both normal and abnormal images.
Analyze the model's ability to detect anomalies across different categories.
2. Analyze results from data and model perspective:
o Data: Investigate if the model struggles with specific anomaly types, subtleties, or modalities.
o Model: Consider limitations arising from model architecture, training parameters, or loss functions.
3. Refine based on insights: Based on the analysis, refine the model architecture, training strategy, or data
preprocessing techniques to improve performance.

You might also like