Intuitive Axial Augmentation Using Polar-Sine-Based Piecewise Distortion for Medical Slice-Wise Segmentation

TL;DR

This work addresses the need for interpretable data augmentation in medical slice-wise segmentation by introducing a polar-sine-based piecewise affine distortion that simulates radiology posture variability while preserving anatomical relationships. It couples a data preprocessing pipeline with metadata-driven scan table removal and a similarity-guided hyperparameter search to control augmentation intensity, enabling robust training without extra data. Empirical results across gastric cancer CT data and public MRI/CT datasets show consistent Dice gains across multiple segmentation architectures, with the similarity-guided approach providing a principled, computation-efficient way to select augmentation parameters. The approach emphasizes clinical interpretability and practical deployment, offering a plug-and-play augmentation that integrates with standard medical imaging workflows and DICOM metadata.

Abstract

Most data-driven models for medical image analysis rely on universal augmentations to improve accuracy. Experimental evidence has confirmed their effectiveness, but the unclear mechanism underlying them poses a barrier to the widespread acceptance and trust in such methods within the medical community. We revisit and acknowledge the unique characteristics of medical images apart from traditional digital images, and consequently, proposed a medical-specific augmentation algorithm that is more elastic and aligns well with radiology scan procedure. The method performs piecewise affine with sinusoidal distorted ray according to radius on polar coordinates, thus simulating uncertain postures of human lying flat on the scanning table. Our method could generate human visceral distribution without affecting the fundamental relative position on axial plane. Two non-adaptive algorithms, namely Meta-based Scan Table Removal and Similarity-Guided Parameter Search, are introduced to bolster robustness of our augmentation method. In contrast to other methodologies, our method is highlighted for its intuitive design and ease of understanding for medical professionals, thereby enhancing its applicability in clinical scenarios. Experiments show our method improves accuracy with two modality across multiple famous segmentation frameworks without requiring more data samples. Our preview code is available in: https://github.com/MGAMZ/PSBPD.

Figures (12)

• Figure 1: Overview of the proposed method.
• Figure 2: Overview of data preprocess.
• Figure 3: Generated control point map with $A=1$, $\omega=1$. The points of the same color correspond in two sub-figs. The positions of the converted control points have changed, with the points closer to the far end undergoing more drastic changes. However, the relative positional relationships between the control points remain unchanged, and points of similar colors still cluster together.
• Figure 4: Distorted reconstructed slices with different parameters. $\Psi=1$ to better compare the effect. Larger $A$ and $\omega$ will lead to extreme distortion effect, but still remains continuity of adjacent areas. The scanning table is also distorted, this may introduce unnecessary noise.
• Figure 5: Geometric Positional Modeling used in scan table removal. The definition of DICOM symbols are available in \ref{['tab:DICOM_Fields']}.