Beyond Strong labels: Weakly-supervised Learning Based on Gaussian Pseudo Labels for The Segmentation of Ellipse-like Vascular Structures in Non-contrast CTs

TL;DR

A novel weakly-supervised framework using the elliptical topology nature of vascular structures in CT slices using 2D Gaussian heatmaps serving as pseudo labels is introduced, which outperforms strong-label-based fully-supervised learning in non-contrast CTs.

Abstract

Deep-learning-based automated segmentation of vascular structures in preoperative CT scans contributes to computer-assisted diagnosis and intervention procedure in vascular diseases. While CT angiography (CTA) is the common standard, non-contrast CT imaging is significant as a contrast-risk-free alternative, avoiding complications associated with contrast agents. However, the challenges of labor-intensive labeling and high labeling variability due to the ambiguity of vascular boundaries hinder conventional strong-label-based, fully-supervised learning in non-contrast CTs. This paper introduces a weakly-supervised framework using ellipses' topology in slices, including 1) an efficient annotation process based on predefined standards, 2) ellipse-fitting processing, 3) the generation of 2D Gaussian heatmaps serving as pseudo labels, 4) a training process through a combination of voxel reconstruction loss and distribution loss with the pseudo labels. We assess the effectiveness of the proposed method on one local and two public datasets comprising non-contrast CT scans, particularly focusing on the abdominal aorta. On the local dataset, our weakly-supervised learning approach based on pseudo labels outperforms strong-label-based fully-supervised learning (1.54\% of Dice score on average), reducing labeling time by around 82.0\%. The efficiency in generating pseudo labels allows the inclusion of label-agnostic external data in the training set, leading to an additional improvement in performance (2.74\% of Dice score on average) with a reduction of 66.3\% labeling time, where the labeling time remains considerably less than that of strong labels. On the public dataset, the pseudo labels achieve an overall improvement of 1.95\% in Dice score for 2D models while a reduction of 11.65 voxel spacing in Hausdorff distance for 3D model.

Figures (13)

• Figure 1: Conventional and elliptical boundaries ((b) and (c)) outlining an aorta derived from a non-contrast CT scan (a). Both enclosed regions share fundamental attributes such as the central point $(x,y)$ and $(x',y')$, rotation angles $\theta$ and $\theta'$, semi-major axes $w$ and $w'$, and semi-minor axes $h$ and $h'$. The computed Dice coefficient for the two enclosed regions is 98.1%$\pm$0.5 across the entire local dataset. A 2D Gaussian distribution generated from (c) is depicted in (d), containing pixel intensities within the range of $[0.0,1.0]$. This Gaussian distribution functions as a pseudo label.
• Figure 2: Comparison of (a) strong-label-based training and (b)-(c) proposed pseudo-label-based weakly-supervised training approaches. In (a), conventional strong-label-based training requires time-intensive, expert-elaborated labeling and heavy supervision of vascular surgeons for strong labels, employing Dice and BCE loss for model optimization. For the proposed method, (b) shows the generation of pseudo label. Based on the proposed annotation standards, elliptical structures are efficiently annotated by the experts. The elliptical structures are then processed via an ellipse-fitting algorithm to establish five foundational parameters: the location of the central point $(x,y)$, the semi-major and semi-minor axes $w$ and $h$, and the rotation angle $\theta$. These parameters create 2D Gaussian heatmaps with pixel intensities in $[0,1]$, employing a constant to restrict intensities exceeding $0.5$ within the ellipse boundary. The Gaussian heatmaps serve as pseudo labels, generated through a weak but efficient process. With the annotation standards and efficiency, external public data can be incorporated to enrich the training set without exhaustive annotator efforts or heavy supervision of surgeons. Consequently, we regard the subsequent training in (c) as a weakly-supervised training strategy. It adopts a novel combination of a voxel reconstruction loss and a distribution loss to adapt the pseudo labels for model optimization.
• Figure 3: Annotations by a single expert using ImageJ schneider2012nih displayed in boundary and binary masks. The samples are randomly selected from three types of abdominal aortas in non-contrast CT slices: regular-shaped (circular), irregular-shaped (elliptical), and large-sized (aneurysm-contained) aortas. The annotations are obtained from three mechanisms of ImageJ: 1) Brush Tool uses a tiny draggable circular brush for target region filling, and 2) Free Hand delineates along boundaries, the red dotted boxed indicates intra-observer variability of the two approaches, and 3) Elliptical Tool, employed in our approach, selects elliptical regions. The series of intra-observer variability between Brush Tool and Elliptical Tool are marked by the red dotted box. Annotations via Elliptical Tool adhere to the proposed annotation standards, depicting stable topologies. The average labeling time per slice of each tool is showed in last row.
• Figure 4: Process of pseudo label generation, including (a) the efficient labeling of ellipse-like structures based on the proposed annotation standards, (b) ellipse-fitting to obtain the five parameters numerically defining the ellipse, and (c) 2D Gaussian heatmap generation based on the elliptical parameters. The generated Gaussian heatmap contains the pixel intensities of $[0,1]$, used as the pseudo label for the weakly-supervised learning. Note that only step (a) is manually performed, while steps (b) and (c) are fully automatic processes.
• Figure 5: The division of dataset for 3-fold cross-validation. In each fold, the local data are separated to 15 samples (volumes) for training, 5 for validation and the rest 10 for testing. There are 30 samples of external data MSD, including 10 Lungs and 20 Livers, serving as additional training set. Note that all the training and additional training sets are trained with pseudo labels while the inference in validation and testing sets are evaluated by strong labels.