Quantized TotalSegmentator for Fast Abdominal Organ Segmentation on CPUs
Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
BACKGROUND
TotalSegmentator is designed for a graphics processing unit (GPU), runs slowly on a central processing unit (CPU), and depends on Python, limiting CPU-only clinical deployment.
OBJECTIVE
The purpose of this study was to evaluate whether TS-Portable, a self-contained CPU executable running 8-bit integer (INT8)–quantized TotalSegmentator models, is noninferior to the reference implementation across 15 abdominal organs in an external multicenter cohort, quantify CPU-only runtime, and demonstrate spleen volumetry on a picture archiving and communication system (PACS) reading station.
METHODS
This retrospective study compared TS-Portable with the reference Python implementation on 100 external multicenter CT scans and 28 chronic liver disease series. Per-case mean Dice similarity coefficient (DSC) and 3-mm normalized surface distance (NSD@3 mm) across 15 organs (1.5-mm resolution) and spleen-only mode were tested against a 0.01 noninferiority margin (paired bootstrap 95% CIs). Workstation and PACS runtimes were recorded; spleen volumes were compared via Lin’s concordance correlation coefficient (CCC) and Bland-Altman analysis.
RESULTS
The external cohort included 100 patients (53 men, 47 women; mean age, 53 ± 16 [SD] years). Across 15 organs, TS-Portable was noninferior for mean DSC (difference, 0.00017; 95% CI, 0.00003–0.00031; p < .001) and NSD@3 mm (difference, 0.00015; 95% CI, −0.00002 to 0.00031; p < .001) and in spleen-only mode (both p < .001). Workstation 1.5-mm median runtime was 48.0 seconds for TS-Portable on CPU versus 342.5 seconds on CPU (ratio, 7.14) and 61.5 seconds on GPU (ratio, 1.28) for the reference. On PACS, CPU spleen volumetry took 5.8 versus 43.4 seconds, agreeing with the reference (CCC, 0.9999; bias, 0.2%; 95% limits of agreement, −7.2% to 7.6%) and manual contours (CCC, 0.989; bias, −6.8%). In 28 chronic liver disease series, TS-Portable ran from DICOM in 6.5 seconds (median), agreeing with the reference (CCC, 0.99999) and manual contours (CCC, 0.992; bias, −3.4%).
CONCLUSION
TS-Portable achieved noninferior abdominal organ segmentation accuracy compared with the reference implementation while running faster on a workstation CPU than the reference on a GPU.
CLINICAL IMPACT
Self-contained deployment enables fast multiorgan segmentation on CPU-only computers and spleen volumetry within seconds on a standard reading station, without dedicated GPUs or software installation.
Highlights
Key Finding
Across 100 external multicenter CT examinations, a quantized, self-contained CPU executable of TotalSegmentator was noninferior to the reference implementation for 15-organ segmentation accuracy and 7.14-fold faster on a workstation CPU, and completed reading-station spleen volumetry in a median of 5.8 seconds.
Importance
Compiling and quantizing three-dimensional segmentation models into a standalone CPU executable removes GPU and software-installation barriers, enabling rapid quantitative CT analysis on standard clinical workstations.
