DeepSPV: An Interpretable Deep Learning Pipeline for 3D Spleen Volume Estimation from 2D Ultrasound Images

Zhen Yuan, David Stojanovski, Lei Li, Alberto Gómez, Haran Jogeesvaran, Esther Puyol-Antón, Baba Inusa, Andrew P. King

Introduction

Splenomegaly, or the enlargement of the spleen, serves as a vital clinical indicator for various systemic conditions, most notably sickle cell disease (SCD). While 3D volumetric analysis via CT or MRI is considered the "gold standard" for assessment, these modalities are often inaccessible in resource-limited settings due to high costs, availability, and radiation exposure. Consequently, clinicians frequently rely on 2D ultrasound (US) length measurements. However, 2D linear dimensions are poor proxies for actual 3D volume. DeepSPV addresses this gap by providing a deep learning-based pipeline to estimate 3D spleen volume directly from standard 2D ultrasound images.

The DeepSPV Pipeline

The DeepSPV architecture is a multi-stage pipeline designed to overcome the lack of paired 2D-3D datasets and provide interpretable results for clinical use.

Overview of the DeepSPV pipeline, including synthetic image generation, segmentation, and volume estimation.

1. Segmentation and Volume Estimation

The pipeline first utilizes an 8-layer U-Net to extract the spleen boundary from 2D ultrasound images. The resulting segmented mask is then processed by a Variational Autoencoder (VAE). The VAE maps the mask into a low-dimensional latent space, which serves as the basis for volume regression. Among several regression strategies tested—including Nearest Neighbor and Post Linear Regression—the Regression VAE (RVAE) emerged as the superior method. The RVAE incorporates a fully connected layer trained to predict volume directly from the latent representation.

2. Synthetic Data Generation

To mitigate the scarcity of paired 2D US and 3D CT data, the authors developed the Ultrasound Semantic Diffusion Model (USDM). This generative model translates 3D CT spleen segmentations into realistic synthetic 2D ultrasound images. This approach allows the model to be trained on a diverse range of spleen morphologies with known ground-truth volumes derived from CT data, effectively bridging the domain gap between 3D imaging and 2D ultrasound.

Experimental Setup and Datasets

The study leveraged three primary datasets to train and validate the pipeline:

  • CT Dataset: 149 manual segmentations from public sources (Medical Segmentation Decathlon and Gibson et al.) used for ground-truth volume.
  • Real 2D US Dataset: 363 images from pediatric SCD patients at Evelina London Children’s Hospital.
  • Synthetic Dataset: 149 USDM-generated images, which have been made publicly available to facilitate further research.

Results and Clinical Interpretability

DeepSPV demonstrates significant improvements over traditional clinical methods. The RVAE model achieved a Mean Relative Volume Accuracy (MRVA) of 92.58% in dual-view settings (coronal and transverse) and 86.62% in single-view settings. In contrast, human experts using standard clinical formulas achieved only 68.54% accuracy from single-view measurements.

Beyond raw accuracy, the pipeline emphasizes interpretability. The VAE's latent space allows clinicians to visualize the "progression" of spleen sizes, providing a qualitative justification for the model's quantitative output. Furthermore, the "RVAE-CI" variant generates 95% confidence intervals, offering a measure of uncertainty essential for clinical decision-making. The model also proved robust to probe rotation errors of up to 15 degrees, a common occurrence in real-world clinical acquisitions.

Conclusion and Future Work

DeepSPV represents a significant step toward automated, accurate, and interpretable spleen volume estimation. By bridging the data gap with diffusion models and utilizing a VAE for interpretable regression, the pipeline significantly outperforms traditional human-led assessments. Future work will focus on expanding the USDM to generate transverse views, refining models for age-stratified pediatric growth, and conducting large-scale clinical trials to establish definitive volumetric thresholds for diagnosing splenomegaly.

Summary updated

BibTeX
@article{yuan2025deepspv,
  author = {Zhen Yuan and David Stojanovski and Lei Li and Alberto Gómez and Haran Jogeesvaran and Esther Puyol-Antón and Baba Inusa and Andrew P. King},
  title = {{DeepSPV: An Interpretable Deep Learning Pipeline for 3D Spleen Volume Estimation from 2D Ultrasound Images}},
  journal = {Medical Image Analysis},
  year = {2025},
  doi = {10.1016/j.media.2025.103671},
}
Figure