Adrenal Volume Quantitative Visualization Tool by Multiple Parameters and an nnU-Net Deep Learning Automatic Segmentation Model

Yi Li, Yingnan Zhao, Ping Yang, Caihong Li, Liu Liu, Xiaofang Zhao, Huali Tang, Yun MaoJ. Imaging Inform. Medicine 2025 · DOI

TL;DR

This paper introduces an automated tool, AVQV, that uses a deep learning model (nnU-Net) to accurately segment and quantify adrenal gland volume from CT scans. This addresses the limitations of manual measurements, offering a more reliable and efficient method for diagnosing and monitoring adrenal-related diseases.

What Problem Does This Paper Address?

Adrenal gland abnormalities, such as hyperplasia and adenomas, are linked to various diseases. Monitoring adrenal gland volume is crucial for diagnosis and treatment, but current methods rely on subjective, manual measurements from 2D CT slices. These manual methods are time-consuming, prone to errors, and suffer from poor interobserver consistency. Furthermore, the adrenal gland's small size and variable shape make accurate segmentation challenging. Existing automated segmentation models often have limitations, such as requiring specific imaging parameters or having poor generalization capabilities, hindering their widespread clinical adoption.

Methods

The core of this work is the development of an automated adrenal gland volume quantification and visualization tool (AVQV) built upon the nnU-Net framework. nnU-Net is a self-configuring deep learning architecture that automatically optimizes hyperparameters and training strategies based on the dataset's characteristics, making it highly adaptable.

The authors employed a U-Net architecture, which consists of an encoder-decoder structure. The encoder progressively downsamples the image to extract features, while the decoder upsamples to restore spatial resolution for precise segmentation. Crucially, nnU-Net incorporates cross-layer connections to integrate features across scales.

Key methodological aspects include:

  • Data Augmentation: Techniques like random rotation, scaling, elastic transformation, gamma correction, and mirroring were used to enhance the model's robustness and generalization.
  • Loss Function: A combination of Dice and cross-entropy loss was used to handle class imbalance.
  • Deep Supervision: Weighted loss was calculated for feature maps at different scales to improve accuracy.
  • Preprocessing: Standardized window width and level settings were applied, and data was resampled to a consistent voxel space.
  • Sliding Window Prediction: Used for inference to handle large images and improve accuracy by assigning higher confidence to central patches.

Datasets

The study utilized a large and diverse dataset comprising 885 CT images for training and internal validation, and an additional 280 CT images for external validation, totaling 1165 cases. This dataset was notable for its heterogeneity, including:

  • Multiple Centers: Data from different hospitals.
  • Multiple Machines: CT scans from five different CT scanner models (SIEMENS HEALTHCARE SOMATOM Definition Flash, TOSHIBA Aquilion ONE, SIEMENS HEALTHCARE SOMATOM Force, SIEMENS HEALTHCARE SOMATOM Definition AS, and GE MEDICAL SYSTEMS Revolution).
  • Multiple Radiation Doses: Including low-dose and routine-dose scans.
  • Multiple Slice Thicknesses: Ranging from 1 mm to 5 mm.
  • Multiple Scanning Modes and Phases: Such as abdominal enhanced scans, abdominal plain scans, and chest low-dose scans.
  • Diverse Adrenal Morphologies: Including normal adrenal glands, adrenal hyperplasia, and adrenal nodules (benign and malignant).

Key Insights and Ideas

  • nnU-Net's Adaptability is Key: The success of the AVQV tool hinges on the inherent adaptability of the nnU-Net framework, which can automatically tune itself to diverse datasets without extensive manual hyperparameter tuning. This is crucial for achieving broad applicability across different imaging protocols and patient populations.
  • Low-Dose CT is Sufficient: The study demonstrates that even low-dose CT scans, commonly used for lung cancer screening, can yield accurate adrenal gland segmentation results, making it feasible to extract valuable adrenal information without additional radiation exposure.
  • Volume is Superior to Linear Measurement: The paper reinforces the idea that adrenal volume is a more comprehensive and reliable indicator of adrenal status than traditional 2D linear measurements, which are prone to significant variability.
  • Comprehensive Dataset is Crucial: The diverse dataset, encompassing various scanners, doses, and pathologies, is a significant strength. This heterogeneity is essential for building a robust model that generalizes well to real-world clinical scenarios, avoiding the overfitting issues seen in previous studies.

Key Results

The AVQV model demonstrated strong performance across both internal and external validation sets:

  • Overall DSC: Achieved an average Dice Similarity Coefficient (DSC) of 0.87±0.07 on the internal validation set and 0.88±0.06 on the external validation set.
  • Low-Dose CT Performance: Specifically for low-dose CT scans, the model achieved a DSC of 0.86 (internal) and 0.87 (external).
  • Comparison to Other Models: The AVQV model outperformed other compared methods, including TotalSegmentator, Swin-UNETR, and V-Net, in terms of DSC and RVE (Relative Volume Error) across various categories. For instance, in external validation, the AVQV achieved a DSC of 0.88±0.06, compared to 0.80±0.19 for TotalSegmentator, 0.87±0.09 for Swin-UNETR, and 0.69±0.14 for V-Net.
  • Consistency with Manual Methods: The tool showed excellent consistency with manual measurements, with an overall intraclass correlation coefficient (ICC) of 0.931, indicating it can effectively replace manual volume determination.
  • 3D Visualization: The tool provides 3D visualization capabilities, aiding in the diagnosis of subtle adrenal abnormalities.

Limitations and Future Work

The authors acknowledge several limitations:

  • Limited Modalities: The current tool is designed for CT scans only and does not incorporate other imaging modalities like MRI, which limits its application in multimodal segmentation.
  • Lack of Longitudinal Comparisons: The study did not include longitudinal model comparisons, which would be valuable for tracking changes over time.

Future work could focus on:

  • Multimodal Integration: Expanding the tool to include MRI data for more comprehensive segmentation.
  • Longitudinal Analysis: Developing methods for tracking adrenal volume changes over time.
  • Standardized Clinical Criteria: Addressing the lack of standardized clinical criteria for adrenal diseases, which can improve the applicability of quantitative indicators.

Summary updated

BibTeX
@article{li2025adrenal,
  author = {Yi Li and Yingnan Zhao and Ping Yang and Caihong Li and Liu Liu and Xiaofang Zhao and Huali Tang and Yun Mao},
  title = {{Adrenal Volume Quantitative Visualization Tool by Multiple Parameters and an nnU-Net Deep Learning Automatic Segmentation Model}},
  journal = {J. Imaging Inform. Medicine},
  year = {2025},
  doi = {10.1007/s10278-024-01158-y},
}
Figure