LA-Seg: Disentangled sinogram pattern-guided transformer for lesion segmentation in limited-angle computed...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Image Reconstruction

Background:

• Limited-angle computed tomography (LACT) provides patient benefits like reduced radiation but suffers from artifacts due to incomplete data.
• Existing reconstruction methods often neglect lesion-specific details by optimizing for overall image quality separately from segmentation.
• This non-end-to-end approach limits the effectiveness of lesion segmentation in LACT.

Purpose of the Study:

• To introduce LA-Seg, a novel transformer-based segmentation model for LACT data.
• To enhance lesion segmentation robustness by integrating an auxiliary reconstruction task within the sinogram domain.
• To improve the identification and interpretation of lesions in LACT scans.

Main Methods:

• LA-Seg utilizes a transformer architecture adapted from video prediction models to process sinogram data.
• An auxiliary reconstruction task estimates incomplete sinogram regions, improving feature reconstruction.
• Contrastive abnormal feature loss is employed to differentiate between normal and abnormal tissue regions.

Main Results:

• LA-Seg demonstrates superior performance compared to existing medical segmentation methods across various LACT conditions.
• The model effectively captures spatial structures and sequential patterns in sinograms.
• Reconstruction of incomplete regions is guided by distinctive patterns, enhancing segmentation accuracy.

Conclusions:

• LA-Seg offers a robust and effective solution for lesion segmentation in limited-angle computed tomography.
• The integrated reconstruction and segmentation approach addresses limitations of previous methods.
• This work advances the potential of LACT for precise medical interpretation and diagnosis.