Advanced Deep Learning in Medical Imaging
A deep learning-based automated workflow for accurate segmentation and measurement of abdominal organs in CT scans, providing comprehensive organ analysis capabilities that support diagnostic workflows and treatment planning.
Key findings
- Automated segmentation and measurement of the kidneys, liver, spleen and prostate on CT
- nnU-Net and U-Net++ segment the organs; a 3D RCNN model measures organ volumes and dimensions
- Precision and recall exceeded 95% for all targeted organs
- Low mean squared error between predicted and ground-truth measurements
Plain-English summary
- The question
- Measuring abdominal organs by hand on CT is slow and varies between readers. Can a deep learning workflow segment and measure the kidneys, liver, spleen and prostate automatically and reliably?
- The data
- Abdominal CT datasets, with segmentation quality checked against variations in patient anatomy and scanner settings.
- The method
- A two-stage workflow: nnU-Net and U-Net++ segment the organs, then a 3D RCNN model measures their volumes and dimensions. Performance was scored with precision, recall and mean squared error.
- The result
- Precision and recall above 95% for all targeted organs, with low mean squared error between predicted and ground-truth measurements. The authors plan to extend the approach to other organs and to complex pathological cases.
Why it matters for hospitals and radiologists
The paper's starting point is that manual organ measurement is slow and varies between readers. An automated pipeline gives radiologists a consistent first measurement to confirm rather than one to produce from scratch.
Related at 5C: Hybrid intelligence: AI + radiologist
Citation
Praveen Shastry, Ashok Sharma, Kavya Mohan, Naveen Kumarasami, Anandakumar D, Mounigasri M, Keerthana R, Kishore Prasath Venkatesh, Bargava Subramanian, Kalyan Sivasailam. Deep Learning-Based Automated Workflow for Accurate Segmentation and Measurement of Abdominal Organs in CT Scans. arXiv preprint arXiv:2503.10717, 2025. https://doi.org/10.48550/arXiv.2503.10717