A weakly-supervised deep learning model for fast localisation and delineation of the skeleton, internal organs, and spinal canal on whole-body diffusion-weighted MRI (WB-DWI)

Antonio Candito ¹, Alina Dragan ², Richard Holbrey ¹

⁰The Institute of Cancer Research, London, UK.

Computer Methods and Programs in Biomedicine +

|

September 10, 2025

View abstract on PubMed

Summary

Automated deep learning accurately maps anatomical structures on whole-body MRI (WB-DWI) for cancer biomarkers. This fast, reproducible method aids disease staging and treatment response assessment.

Area Of Science

Radiology and Medical Imaging
Artificial Intelligence in Healthcare
Oncology Imaging

Background

Whole-body diffusion-weighted MRI (WB-DWI) provides cancer imaging biomarkers like Apparent Diffusion Coefficient (ADC) and Total Diffusion Volume (TDV).
Manual delineation of anatomical structures for these biomarkers is clinically impractical, necessitating automated solutions.
Accurate segmentation of the skeleton, internal organs, and spinal canal is crucial for quantitative analysis.

Purpose Of The Study

To develop and validate an automated algorithm for generating probability maps of key anatomical regions on WB-DWI.
To enable fast and reproducible quantification of imaging biomarkers for cancer staging and treatment monitoring.
To overcome the limitations of manual segmentation in clinical WB-DWI analysis.

Main Methods

A 3D patch-based Residual U-Net deep learning architecture was employed for automated localization and delineation.
The model was trained using soft-labels derived from an atlas-based approach on a multi-center WB-DWI dataset (532 scans).
The pipeline was tested on 45 patient scans, evaluating segmentation accuracy and speed compared to atlas-based methods.

Main Results

The automated model achieved high segmentation performance, with Dice scores up to 0.86 for the spinal canal and 0.83 for internal organs.
Average surface distances for delineations were below 3 mm, and relative median ADC differences were under 10% compared to manual segmentations.
The deep learning model demonstrated a 12x speed improvement over atlas-based methods (25 seconds vs. 5 minutes) and received positive radiologist assessments.

Conclusions

The developed deep learning model provides fast and reproducible probability maps for anatomical structures on WB-DWI.
This automation facilitates the non-invasive quantification of imaging biomarkers, supporting clinical decision-making in cancer care.
The approach holds potential for enhancing disease staging and treatment response assessment using WB-DWI.

Keywords:

Deep learning Metastatic bone disease Multiple Myeloma Quantitative Imaging Biomarkers Segmentation Whole-Body Diffusion-Weighted MRI (WB-DWI)

Related Concept Videos

Magnetic Resonance Imaging

Magnetic resonance imaging (MRI) is a noninvasive medical imaging technique based on a phenomenon of nuclear physics discovered in the 1930s, in which matter exposed to magnetic fields and radio waves was found to emit radio signals. In 1970, a physician and researcher named Raymond Damadian noticed that malignant (cancerous) tissue gave off different signals than normal body tissue. He applied for a patent for the first MRI scanning device in clinical use by the early 1980s. The early MRI...