Deep Supervision by Gaussian Pseudo-label-based Morphological Attention for Abdominal Aorta Segmentation in Non-Contrast CTs

TL;DR

The paper tackles abdominal aorta segmentation in non-contrast CTs, where boundary ambiguity risks overfitting to ill-defined borders. It introduces Morphological Attention (MA) based on Gaussian pseudo-labels integrated into encoder–decoder models via deep supervision, where pseudo labels encode ellipse-like morphology rather than exact boundaries. The pseudo-labels are generated from ellipse-fitting to strong labels and converted into Gaussian heatmaps, and the training loss combines Dice/BCE on the strong labels with an $L_\infty$-based regularization across decoder layers. On a local dataset of 30 non-contrast CTs (5749 slices), MA improves Dice and Hausdorff distance for both 2D and 3D models, preserves morphology, and reduces overfitting, suggesting plug‑and‑play potential for non-contrast CT guided interventions and extension to other vascular structures.

Abstract

The segmentation of the abdominal aorta in non-contrast CT images is a non-trivial task for computer-assisted endovascular navigation, particularly in scenarios where contrast agents are unsuitable. While state-of-the-art deep learning segmentation models have been proposed recently for this task, they are trained on manually annotated strong labels. However, the inherent ambiguity in the boundary of the aorta in non-contrast CT may undermine the reliability of strong labels, leading to potential overfitting risks. This paper introduces a Gaussian-based pseudo label, integrated into conventional deep learning models through deep supervision, to achieve Morphological Attention (MA) enhancement. As the Gaussian pseudo label retains the morphological features of the aorta without explicitly representing its boundary distribution, we suggest that it preserves aortic morphology during training while mitigating the negative impact of ambiguous boundaries, reducing the risk of overfitting. It is introduced in various 2D/3D deep learning models and validated on our local data set of 30 non-contrast CT volumes comprising 5749 CT slices. The results underscore the effectiveness of MA in preserving the morphological characteristics of the aorta and addressing overfitting concerns, thereby enhancing the performance of the models.

Figures (4)

• Figure 1: (a) The general decoder architecture of DL models involving Morphological Attention. (b) Detailed mechanism of Morphological Attention.
• Figure 2: The produce of generation of pseudo labels $P$ from strong labels $S$.
• Figure 3: Convergence of loss of training and validation data of U-net with and without MA. The asterisk (*) denotes the epochs when convergence approaches saturation.
• Figure 4: 3D/2D results from DL models with and without abdominal aorta (MA) improvement. In 2D, red and blue regions indicate ground truth and predictions, while purple represents overlay areas. Green and yellow dotted boxes are False Positives (FPs) and False Negatives (FNs).