The Radiation Oncology Feed For radiation oncologists · physicists · RTTs · dosimetrists
technical-relevance cervical auto-segmentation

PocketNet matched nnU-Net segmentation with 40-fold lower model storage

PocketNet matched nnU-Net accuracy for cervical MRI segmentation while using 40-fold less storage and approximately half the training time.

RadOncToday · ·1 min read

KEY POINTS

  • The retrospective study used pretreatment T2-weighted MRI from 102 patients receiving definitive cervical cancer chemoradiotherapy. The two-stage model first segmented a combined uterus–cervix–vagina–tumour region and then constrained tumour segmentation within that predicted anatomy.
  • Manual contours were reviewed by an abdominal radiologist with 22 years of experience. Five-fold patient-level cross-validation was used internally, while external validation included 18 evaluable patients from the multicentre TCIA Cervical Cancer Tumor Heterogeneity collection.
  • PocketNet achieved mean Dice scores of 0.81 ± 0.12 for the combined organ region and 0.71 ± 0.21 for tumour internally. External performance remained similar at 0.81 ± 0.11 and 0.67 ± 0.23, respectively.
  • Internal 95th-percentile Hausdorff distances were highly variable at 64.2 ± 65.9 mm for the organ region and 38.6 ± 56.3 mm for tumour. External values were lower at 14.8 ± 7.5 mm and 16.0 ± 13.1 mm, indicating substantial case-level outliers despite acceptable average Dice scores.
  • PocketNet contained 0.15 million parameters and occupied 0.57 MB, compared with 5.5 million parameters and 20.9 MB for nnU-Net. Training required 696.6 versus 1,362.8 seconds, while no segmentation metric differed significantly between architectures (all p>0.05).
  • The successful examples in Figure 3 achieved tumour Dice scores of 0.88 and 0.89 and organ scores of 0.92 and 0.95. Figure 4 shows clinically important failure cases, including tumour Dice scores of 0.53 and 0.02, particularly for small lesions near the rectosigmoid.
  • Performance deteriorated for tumours smaller than approximately 3 cm, lesions with limited T2 contrast, very large masses with adnexal invasion, and motion-degraded images. The model used T2-weighted imaging alone and requires prospective testing, broader multicentre training, and mandatory clinician review.

CLINICAL TAKEAWAY

PocketNet offers similar accuracy to a much larger nnU-Net while sharply reducing computational requirements, which may help resource-constrained or time-sensitive MRI workflows. Its tumour contours remain too inconsistent for autonomous planning, particularly for small or poorly visualized lesions.

SOURCE

Physics and Imaging in Radiation Oncology