Refining deep learning segmentation in gallium-68-prostate-specific membrane antigen-11 positron emission tomography: evaluation of small lesion filtering and intersection-over-union thresholds
Clinical Snapshot
PICO Framework
| P — Population | 115 patients with prostate cancer undergoing Ga-68-PSMA-11 PET imaging |
| I — Intervention | 3D U-Net deep learning segmentation with small lesion filtering (<8 or <27 voxels) and varying IoU thresholds (10-50%) |
| C — Comparator | Manual expert contours as ground truth reference standard |
| O — Outcomes | Segmentation accuracy measured by Dice coefficient, sensitivity, specificity, and positive predictive value at voxel, lesion, and patient levels |
Bottom Line
This study demonstrates that 3D U-Net deep learning can achieve high accuracy for automated segmentation of PSMA PET lesions, with patient-level sensitivity and positive predictive value exceeding 94%. The key finding is that filtering out very small lesions (<27 voxels, ~658 mm³) improves segmentation reliability and should be considered in clinical implementation. However, the single-center design, lack of confidence intervals, and absence of clinical impact assessment limit immediate clinical adoption. The optimal IoU threshold of 20-40% for defining true positive lesions provides guidance for future validation studies. While promising for workflow efficiency in nuclear medicine departments, multicenter validation and regulatory approval would be required before clinical implementation in Australian practice.
Key Findings
P Value: Not reported
Effect Size: Dice coefficient improvement of 0.0198 (2.5% relative improvement)
Primary Outcome: Voxel-level Dice coefficient improved from 0.7975 to 0.8173 with 27-voxel filtering
Nnt Or Sensitivity: Patient-level sensitivity 96.6%, positive predictive value 94.5%
Confidence Interval: Not reported
Clinical Application
Requires specialized deep learning infrastructure and validation on local scanner protocols Relevant for Australian nuclear medicine departments using PSMA PET, which is PBS-subsidized for biochemically recurrent prostate cancer. Would need TGA approval for clinical implementation and validation against local protocols. Prostate cancer patients undergoing Ga-68-PSMA-11 PET imaging for staging or restaging
Abstract
PURPOSE: Evaluate how small lesion filtering and different intersection-over-union (IoU) thresholds influence deep learning segmentation of Ga-68-PSMA-11 PET images in prostate cancer. METHODS: A 3D U-Net was trained on 115 patient scans with manual contours as ground truth. Performance was assessed at voxel, lesion and patient levels. Lesions less than 8 voxels (195 mm 3 ) or less than 27 voxels (658 mm 3 ) were optionally discarded, and lesion-level metrics were computed across 10-50% IoU thresholds. RESULTS: Excluding lesions less than 27 voxels increased voxel-level Dice from 0.7975 to 0.8173 and improved precision. Lesion-level metrics were stable across IoU thresholds of 20-40% after 27-voxel filtering, while sensitivity declined at higher overlap thresholds. Patient-level sensitivity and positive predictive value reached 96.6 and 94.5%, respectively. CONCLUSION: Implementing small lesion filtering criteria improves the reliability of automated segmentation outputs, particularly for quantitative evaluation. For lesion-level metrics, defining true positives within the range of 20-40% IoU thresholds is optimal. Further validation through multicenter studies and larger datasets is essential to ensure the generalizability of these findings.
References
- 1.Huang, Y.-Y., Yang, S.-H., Chen, C.-Y., Shen, Y.-J., Chen, Y.-H., Liou, F.-M., & Chen, J.-C. (2026). Refining deep learning segmentation in gallium-68-prostate-specific membrane antigen-11 positron emission tomography: evaluation of small lesion filtering and intersection-over-union thresholds. Nuclear Medicine Communications. https://doi.org/10.1097/MNM.0000000000002114
Related Research
European journal of radiology
A CT-based deep learning model to differentiate between benign and malignant adrenal lesions
3 Aug 2026
European respiratory review : an official journal of the European Respiratory Society
Prediction of PD-L1 expression in nonsmall cell lung cancer using artificial intelligence models based on radiomics: a systematic review and meta-analysis
10 July 2026
Radiation protection dosimetry
Impact of neutron sieve therapy for deep site in boron neutron capture therapy using Monte Carlo calculation
8 July 2026
This content is for educational purposes for healthcare professionals only and does not constitute clinical advice. Clinical decisions should be based on individual patient assessment, current guidelines, and appropriate specialist consultation. Editorial Standards · Privacy Policy · Terms of Service