Adaptive patch sampling and location-aware reasoning for whole body PET-CT multi-organ segmentation.

Area of Science:

• Medical Imaging
• Machine Learning
• Computer Vision

Background:

• Patch-wise learning is standard for dense prediction but assumes fixed sampling, which is inefficient when learning difficulty varies.
• Existing methods struggle with spatially varying and evolving learning difficulties during neural network optimization.

Purpose of the Study:

• To develop a dynamic computation allocation strategy for patch-wise learning.
• To improve the efficiency and performance of neural networks in large-scale dense prediction tasks.

Main Methods:

• Proposed an adaptive patch sampling (APS) algorithm that learns sampling distributions based on voxel-wise uncertainty and prediction error.
• Introduced a patch encoding (PE) block to infer implicit location information and modulate feature representations using context-dependent channel-wise attention.
• Conducted experiments on whole-body multi-organ PET-CT segmentation and validated on the Synapse dataset.

Main Results:

• Demonstrated faster convergence and consistent performance gains in PET-CT segmentation tasks.
• External validation on the Synapse dataset confirmed the robustness of the proposed methods.
• Mechanistic analyses revealed APS-induced sampling behaviors and PE-driven representation modulation.

Conclusions:

• The proposed adaptive patch sampling and patch encoding offer an efficient learning strategy for patch-wise training.
• Dynamic sampling and contextual conditioning significantly influence optimization in large-scale dense prediction tasks.
• This work provides valuable insights into optimizing deep learning models for medical image analysis.