A NOVEL IMAGE-SPECIFIC TRANSFER APPROACH FOR PROSTATE SEGMENTATION IN MR IMAGES

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Prostate segmentation in Magnetic Resonance (MR) images is crucial for prostate cancer treatment.
• Existing methods often fail due to variations in patient images, necessitating improved approaches.

Purpose of the Study:

• To develop a novel transfer learning approach for accurate prostate segmentation in MR images.
• To address the limitations of global classifiers by accounting for inter-patient image discrepancies.

Main Methods:

• An image-specific classifier is developed for each training MR image.
• A novel Semi-Coupled Dictionary Transfer Learning (SCDTL) method is employed to obtain dictionaries and a mapping matrix.
• Classifiers are selectively transferred from source to target domains using learned dictionaries and mapping matrix.

Main Results:

• The proposed transfer approach demonstrates competitive performance against state-of-the-art transfer learning methods.
• The SCDTL-based method outperforms conventional deep neural network approaches for prostate segmentation.
• Achieved improved accuracy in segmenting prostates within MR images.

Conclusions:

• The novel transfer learning strategy effectively improves prostate segmentation accuracy in MR images.
• SCDTL offers a promising direction for enhancing medical image analysis and prostate cancer treatment.
• This method provides a more robust solution compared to existing deep learning techniques.