PET-CT image fusion using random forest and à-trous wavelet transform.

Area of Science:

• Medical Imaging
• Image Processing
• Machine Learning

Background:

• Multimodal medical imaging requires effective fusion techniques to combine complementary information from different sources.
• Existing image fusion methods may not fully leverage the potential of advanced algorithms and transforms for optimal results.

Purpose of the Study:

• To propose new image fusion rules for multimodal medical images.
• To evaluate the proposed method against traditional techniques using established and novel performance metrics.

Main Methods:

• Decomposition of source images into approximation and detail coefficients using translation-invariant à-trous wavelet transform (AWT).
• Application of a random forest learning algorithm to select pixels for constructing fused image coefficients.
• Reconstruction of the fused image via inverse AWT.

Main Results:

• The proposed random forest and AWT-based fusion method demonstrated superior performance over a traditional Mallat wavelet transform method.
• Experimental results showed improvements in both visual and quantitative qualities of the fused images.
• A newly introduced image fusion performance measure proved meaningful in comparing fusion methods.

Conclusions:

• The developed image fusion rules offer a significant advancement over traditional methods for multimodal medical images.
• The combination of random forest and à-trous wavelet transform provides an effective approach for pixel-level image fusion.
• The proposed performance measure is valuable for assessing the efficacy of medical image fusion techniques.