Medical Image Translation using CycleGAN with Brain Tumor Detection

Introduction

Medical Image Translation using CycleGAN with Brain Tumor Detection is a project aimed at translating CT scan images into corresponding MRI-like images while also incorporating brain tumor detection. This project leverages the power of CycleGAN, a deep learning architecture, to bridge the gap between different imaging modalities and enhance diagnostic accessibility in medical imaging.

Methodology

The methodology involves the following steps:

1. Data Preprocessing:

   • CT and MRI images are preprocessed to ensure uniformity in size and format.
   • Data augmentation techniques may be applied to increase dataset diversity.

2. Model Architecture:

   • CycleGAN architecture is utilized for unpaired image translation.
   • Conditional GANs are employed with cyclic connections between generators.
   • Generators are conditioned with concatenation of alternate unpaired patches from input and target images.
   • Discriminators validate the translated imagery using minimax function.
   • Adaptive dictionaries are incorporated into generators to reduce possible degradation.

3. Training:

   • The model is trained on a dataset consisting of CT and MRI image pairs.
   • Loss functions, including adversarial, non-adversarial, forward-backward cyclic, and identity losses, are utilized to minimize variance.

4. Evaluation:

   • Translated MRI-like images are evaluated for anatomical fidelity and diagnostic accuracy.
   • Brain tumor detection is performed on generated MRI images.

Results

The results of the project demonstrate:

• High fidelity translation of CT images into MRI-like images.
• Successful brain tumor detection on translated MRI images.

Further Research Directions

1. Data Diversity:

   • Expand the training dataset to include more diverse anatomical regions and pathologies.
   • Explore translation capabilities across other imaging modalities such as PET, X-ray, etc.

2. Model Optimization:

   • Investigate techniques to improve model performance and reduce computational requirements.

References

[1] Isola, P. et al. (2018) Image-to-image translation with conditional adversarial networks, arXiv.org. Available at: https://arxiv.org/abs/1611.07004

[2] Rai, S., Bhatt, J. S., & Patra, S. K. (2023, November 4). A strictly bounded deep network for unpaired cyclic translation of medical images. arXiv.org. https://arxiv.org/abs/2311.02480

[3] Dataset: DARREN2020. CT and MRI brain scans, https://www.kaggle.com/datasets/darren2020/ct-to-mri-cgan

Outputs

For detailed explanation about the project, refer this report.