Grad-CAM Guided U-Net for MRI-based Pseudo-CT Synthesis.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Accurate CT imaging is crucial for radiotherapy planning and PET/MRI attenuation correction.
• Synthesizing CT images from MRI (pseudo-CT) offers a radiation-free alternative.

Purpose of the Study:

• To develop an advanced deep learning model for MRI-to-CT image translation.
• To enhance the accuracy of pseudo-CT synthesis, particularly for bone structures.

Main Methods:

• A 2D U-Net architecture with an integrated Grad-CAM attention mechanism was employed.
• A classifier was trained to differentiate MR and CT images, guiding the translation network.
• Image registration was performed on the RIRE dataset due to misalignment.

Main Results:

• The proposed method demonstrated qualitative and quantitative improvements over the baseline U-Net.
• Significant performance gains were observed in the accurate translation of bone regions.
• Generated CT-class-specific localization maps confirmed improved focus on relevant attributes.

Conclusions:

• The Grad-CAM guided attention mechanism enhances pseudo-CT synthesis accuracy from MRI.
• This technique is vital for radiation therapy planning and combined PET/MRI applications.
• Accurate pseudo-CTs can reduce patient exposure to ionizing radiation.