DeepSPV: A 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

TL;DR

This paper introduces DeepSPV, a novel deep learning pipeline that accurately estimates 3D spleen volume from standard 2D ultrasound images. This advancement is crucial for diagnosing conditions like sickle cell disease, especially in resource-limited settings where advanced imaging is scarce. The pipeline achieves state-of-the-art accuracy, outperforming human experts and existing methods.

What Problem Does This Paper Address?

Splenomegaly, or an enlarged spleen, is a critical indicator for various medical conditions, most notably sickle cell disease (SCD). While spleen length measured from 2D ultrasound is commonly used, spleen volume is the gold standard for assessment. However, obtaining accurate spleen volume typically requires 3D imaging modalities like CT or MRI, which are expensive, not widely available, and can involve ionizing radiation. 3D ultrasound is a potential alternative but faces challenges with image acquisition quality and completeness due to anatomical obstructions and limited acoustic windows. Current methods for estimating spleen volume from 2D ultrasound often rely on manual measurements and linear regression, which are operator-dependent and less accurate. This paper aims to bridge this gap by developing a deep learning pipeline that can precisely estimate 3D spleen volume directly from readily available 2D ultrasound images, aligning with current clinical workflows.

Methods

The DeepSPV pipeline consists of two main components:

  1. 2D Spleen Segmentation: The pipeline utilizes a U-Net architecture, a well-established convolutional neural network for image segmentation, to automatically segment the spleen from 2D ultrasound images. This is a refinement of the authors' previous work.
  2. 3D Spleen Volume Estimation: Following segmentation, a Variational Autoencoder (VAE) based framework is employed to estimate the 3D spleen volume from the 2D segmentations. Three methods are proposed within this framework:
  3. Nearest Neighbour (NN) Search: Maps the input segmentation to the VAE's latent space and finds the closest training sample's corresponding ground truth volume.
  4. Post Linear Regression (PLR): Trains a linear regression model on the latent space representations of the VAE's encoder to predict spleen volume.
  5. End-to-End Regression VAE (RVAE): Integrates regression layers directly into the VAE architecture, allowing it to learn potentially non-linear relationships between latent embeddings and spleen volume. This method also incorporates an additional regression loss term during training.

    The RVAE method, due to its superior performance, is further extended to provide 95% confidence intervals for volume estimates (RVAE-CI).

To address the scarcity of paired 2D US images and ground truth volumes, the authors also developed a US Semantic Diffusion Model (USDM). This model generates realistic synthetic 2D ultrasound images from 2D CT spleen segmentations, enabling end-to-end pipeline evaluation.

Datasets

  • CT Segmentation Dataset: 149 manual 3D spleen segmentations derived from CT volumes. This dataset was sourced from two publicly available collections and includes cases with and without splenomegaly.
  • Real 2D Ultrasound (US) Dataset: 363 anonymized 2D US images of pediatric patients with SCD, collected from Evelina London Children's Hospital. These images were used for segmentation and qualitative assessment of synthetic data.
  • Synthetic US Dataset: Generated using the USDM, this dataset comprises synthetic 2D US images and corresponding ground truth spleen volumes, derived from CT segmentations. This synthetic dataset was crucial for evaluating the full DeepSPV pipeline.

Key Insights and Ideas

  • Bridging the Modality Gap: The core innovation is accurately estimating 3D volume from 2D images, a critical step for making advanced diagnostics accessible in resource-constrained settings.
  • VAE for Interpretability and Volume Estimation: The VAE framework not only facilitates volume estimation but also provides a latent space that shows clear clustering of normal and splenomegalic spleens, offering a degree of interpretability.
  • Synthetic Data Generation for Training: The USDM is a significant contribution, enabling robust evaluation of the pipeline in the absence of large, paired real-world datasets. This approach demonstrates the potential of diffusion models for medical image synthesis.
  • RVAE Outperforms Other Methods: The end-to-end regression VAE (RVAE) consistently outperformed simpler regression methods and nearest neighbor approaches, highlighting the benefit of learning non-linear relationships.
  • Confidence Intervals Enhance Clinical Utility: The RVAE-CI method provides confidence intervals, which are vital for clinical decision-making and building trust in AI-driven diagnostic tools.

Key Results

  • Spleen Volume Estimation Accuracy: The RVAE method achieved a mean relative volume accuracy (MRVA) of 86.62% for single-view and 92.58% for dual-view experiments on CT-derived segmentations. This significantly outperformed human experts (MRVA of 68.54% and 81.60%) and other state-of-the-art methods like Pix2Vox++ and regression-based networks.
  • Full Pipeline Evaluation: On the synthetic US dataset, the DeepSPV pipeline (using RVAE) achieved an MRVA of 83.0% for single-view estimation, demonstrating its effectiveness even with predicted segmentations.
  • Robustness: The RVAE method showed stable performance against moderate in-plane rotational errors in probe positioning.
  • Synthetic Image Quality: The USDM generated synthetic US images that were highly similar to real US images, with a radiologist identifying 13 out of 25 synthetic images as real. The Dice score for segmentation on synthetic images was comparable to or better than on real images.
  • Interpretability: Latent space visualizations showed clear separation between normal and splenomegalic cases, with RVAE demonstrating more pronounced progression in reconstructed spleen size.

Limitations and Future Work

  • Synthetic Data for Single-View: The current evaluation of the full pipeline was limited to single-view (coronal) synthetic US images due to the lack of transverse 2D US images for USDM training. Future work will involve generating synthetic transverse images to enable dual-view pipeline evaluation.
  • Domain Gap: While USDM generates realistic images, a slight domain gap between US and CT-derived segmentations may persist.
  • Paired Real-World Data: The authors emphasize the need for future collection of paired real-world US and CT datasets for more comprehensive validation.
  • Clinical Thresholds: Establishing clinically acceptable thresholds for volumetric measurement errors remains an open challenge requiring further large-scale validation.
  • Age Stratification: The CT data was from adults, while US data was from pediatric SCD patients. While the authors believe this difference has minimal impact, future work may benefit from more age-stratified datasets for broader generalizability.
  • Clinical Integration: The authors suggest that while the method shows promise, it would initially be used alongside current clinical standards, with gradual adoption as clinician trust grows.

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: A 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