MAET-SAM: Magneto-Acousto-Electrical Tomography segmentation network based on the segment anything model

Area of Science:

• Biomedical Imaging
• Medical Image Analysis
• Computational Imaging

Background:

• Magneto-Acousto-Electrical Tomography (MAET) is a hybrid imaging technique combining ultrasound and electrical impedance tomography for biological tissue conductivity mapping.
• Current MAET reconstruction suffers from noise, leading to blurred boundaries, low contrast, and artifacts, hindering accurate diagnosis.
• Existing methods struggle with the complex conductivity variations in different tissue and disease organizations.

Purpose of the Study:

• To develop an advanced segmentation method for MAET images to improve resolution and reduce noise.
• To introduce a novel deep learning model, MAET-SAM, for enhanced MAET image analysis.
• To establish a specialized dataset, MAET-IMAGE, for training and validating MAET image segmentation models.

Main Methods:

• A dataset of MAET-reconstructed conductivity maps, MAET-IMAGE, was created.
• A MAET tomography segmentation network, MAET-SAM, was designed based on the Segment Anything Model (SAM).
• The SAM encoder weights were frozen, and an adaptive, prompt-free decoder was developed for end-to-end MAET image segmentation.

Main Results:

• MAET-SAM demonstrated superior performance in segmenting MAET images compared to traditional methods.
• The proposed model significantly improved image resolution and reduced noise in MAET reconstructions.
• Qualitative and quantitative experiments confirmed the effectiveness of MAET-SAM over models with initial weights.

Conclusions:

• MAET-SAM offers a significant advancement in MAET image segmentation, outperforming existing techniques.
• The developed model enhances the diagnostic potential of MAET in medical imaging analysis.
• This work paves the way for improved clinical diagnosis through more accurate MAET-based imaging.