A Deep Regression Model for Seed Identification in Prostate Brachytherapy

Luke Fu; Ren-Dih Sheu; Yading Yuan; Yeh-Chi Lo

arxiv: 1906.10183 · v1 · pith:LTN6SKUYnew · submitted 2019-06-24 · 📡 eess.IV · cs.CV· physics.med-ph

A Deep Regression Model for Seed Identification in Prostate Brachytherapy

Yading Yuan , Ren-Dih Sheu , Luke Fu , Yeh-Chi Lo This is my paper

Pith reviewed 2026-05-25 16:38 UTC · model grok-4.3

classification 📡 eess.IV cs.CVphysics.med-ph

keywords seed identificationprostate brachytherapydeep convolutional networkCT imagingregression modelpost-implant dosimetrymetal artifactsfully convolutional network

0 comments

The pith

A 3D deep convolutional network regresses CT images to seed probability maps and detects 94.1 percent of implanted brachytherapy seeds.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper sets out to show that framing seed detection as a supervised regression problem solved by a 3D fully convolutional network produces a clean probability map from noisy CT scans. This map reduces the effects of metal artifacts and overlapping seeds so that simple post-processing can locate the seeds reliably. A reader would care because post-implant dosimetry requires accurate seed positions to confirm the radiation dose actually delivered to the prostate. The network was trained on 5534 seeds from 70 patients and tested on 2286 seeds from 30 separate patients, reaching 94.1 percent detection and a 16 percent gain over commercial software.

Core claim

The central claim is that the 3D deep fully convolutional network, trained to map each input CT volume to a corresponding probability volume where each voxel value indicates the chance it belongs to a seed, correctly identifies 2150 of 2286 seeds (94.1 percent) in the held-out test patients and improves detection by 16 percent relative to the VariSeed commercial finder.

What carries the argument

The 3D deep fully convolutional network that outputs a probability map for seed voxels.

If this is right

The probability map output suppresses metal artifacts and overlapping seed appearances, so downstream localization steps become simpler and more controllable.
The same trained network can process a clinical database containing thousands of seeds across many patients without per-case manual tuning.
Detection performance improves by 16 percent over the widely used commercial seed finder on the reported test set.
The regression formulation avoids direct classification of every voxel and therefore reduces false positives from artifact patterns.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same regression-to-probability-map strategy could be tested on other small metallic objects in artifact-heavy CT scans such as vascular stents or orthopedic hardware.
If the model maintains accuracy across scanner vendors, clinics could adopt it to shorten the time between seed implantation and final dosimetry review.
Performance on future patients might drop if CT acquisition protocols or seed types differ from those in the 100-patient training set.

Load-bearing premise

The locations used as ground truth for training and testing are accurate and the 30 test patients match the distribution of future clinical cases.

What would settle it

Running the trained model on a new cohort of patients whose seed positions have been verified by at least two independent observers or by an orthogonal imaging modality and obtaining a detection rate below 85 percent.

Figures

Figures reproduced from arXiv: 1906.10183 by Luke Fu, Ren-Dih Sheu, Yading Yuan, Yeh-Chi Lo.

**Figure 1.** Figure 1: An example of seed appearance in CT images in axial (left), sagittal (middle) and coronal (right) view, respectively. Yellow arrows indicate the metal artifacts and red dots represent mannually annotated seed locations. Clustered seeds can be clearly seen in saggital and coronal views, as indicated by the blue arrows. PID is typically performed at day 30 following implantation that utilizes CT to image the… view at source ↗

**Figure 2.** Figure 2: (a) The target probability maps created from the dot manual annotations in [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Figure 3: Architecture of the proposed deep regression network (DRN). DRN is a fully 3D model that employs convolution and max-pooling to aggregate contextual information, and uses transpose convolution and long-range skip connection for better determination of seed locations. The numbers under each block represent the dimensions of its output, in which the first dimension denotes the feature channel. as 2 for upsca… view at source ↗

**Figure 4.** Figure 4: Two examples of PID study in CT images in axial, sagittal and coronal views. The second and fourth rows are the corresponding probability maps generated by the proposed DRN model. The right column shows the overall 3D distributions of the ground truth and seeds identified by DRN. In each figure, the red dots represent the ground truth while the cyan dots are seed locations identified by DRN. seeds. For a l… view at source ↗

read the original abstract

Post-implant dosimetry (PID) is an essential step of prostate brachytherapy that utilizes CT to image the prostate and allow the location and dose distribution of the radioactive seeds to be directly related to the actual prostate. However, it it a very challenging task to identify these seeds in CT images due to the severe metal artifacts and high-overlapped appearance when multiple seeds clustered together. In this paper, we propose an automatic and efficient algorithm based on 3D deep fully convolutional network for identifying implanted seeds in CT images. Our method models the seed localization task as a supervised regression problem that projects the input CT image to a map where each element represents the probability that the corresponding input voxel belongs to a seed. This deep regression model significantly suppresses image artifacts and makes the post-processing much easier and more controllable. The proposed method is validated on a large clinical database with 7820 seeds in 100 patients, in which 5534 seeds from 70 patients were used for model training and validation. Our method correctly detected 2150 of 2286 (94.1%) seeds in the 30 testing patients, yielding 16% improvement as compared to a widely-used commercial seed finder software (VariSeed, Varian, Palo Alto, CA).

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper reports 94.1% seed detection on 30 test patients with a 16% edge over VariSeed, but supplies no ground-truth protocol or methods details.

read the letter

The main point is that this work frames seed localization in prostate brachytherapy CTs as a 3D regression task with a fully convolutional network and reports a concrete gain on held-out cases. The 70-patient training set covers 5534 seeds and the 30-patient test set covers 2286, which is a reasonable scale for the domain. The regression output is presented as making post-processing simpler than direct detection approaches in prior work. That is the actual new element here: applying the regression formulation to this specific clinical imaging problem rather than a new architecture or theory. The result shows the model can suppress metal artifacts enough to outperform a standard commercial tool on the reported metric. The paper does that part cleanly enough for an empirical report. The soft spots sit in the validation layer. The abstract gives no architecture, loss function, optimizer, or cross-validation scheme, and no error bars or statistical comparison. More critically, there is no description of how the 7820 ground-truth seed locations were established—no annotator count, no consensus process, no distance tolerance for a correct detection, and no check for label noise. If the reference standard has systematic bias on clustered seeds, the 16% improvement cannot be read as a reliable clinical advance. The 70/30 patient split also assumes the test distribution matches future cases, but no scanner or protocol details back that up. This paper is for medical physicists and radiation oncologists who need practical seed identification tools in post-implant dosimetry. A reader working on clinical translation can extract the performance numbers and decide whether to test the approach locally. Methods-focused readers will need the full details before building on it. The work shows clear engagement with the task and the literature on its own terms, so it deserves peer review. Send it out, but ask for the ground-truth protocol and full training description in revision.

Referee Report

3 major / 1 minor

Summary. The manuscript presents a 3D deep fully convolutional network that models seed localization in post-implant CT as a supervised regression task producing a per-voxel probability map. On a clinical database of 7820 seeds across 100 patients (5534 from 70 patients for training/validation), the method detects 2150 of 2286 seeds (94.1 %) in the 30-patient held-out test set and reports a 16 % improvement over the VariSeed commercial software.

Significance. If the reported detection rate holds under transparent evaluation, the approach could reduce manual effort and improve consistency in post-implant dosimetry, a critical step in prostate brachytherapy. The scale of the clinical dataset (100 patients) is a positive feature of the empirical evaluation.

major comments (3)

[Abstract] Abstract: the 94.1 % detection rate and 16 % improvement over VariSeed are reported without any description of how the 7820 ground-truth seed locations were established (number of annotators, consensus procedure, or distance tolerance for a correct detection). Because both the proposed regressor and the commercial comparator are scored against the same reference, any systematic bias in the labels renders the delta uninterpretable.
[Abstract] Abstract (methods section implied by the patient split): no architecture diagram, loss function, optimizer, training schedule, data augmentation, or cross-validation procedure is supplied, so the regression model itself cannot be assessed or reproduced.
[Abstract] Abstract: the 70/30 patient split is presented without scanner model, imaging protocol, slice thickness, or institution details, leaving the assumption that the test distribution matches future clinical cases unsupported.

minor comments (1)

[Abstract] Abstract: typographical error 'it it a very challenging task' should read 'it is a very challenging task'.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback. We will revise the manuscript to provide the requested details on ground truth annotation, model implementation, and dataset characteristics.

read point-by-point responses

Referee: [Abstract] Abstract: the 94.1 % detection rate and 16 % improvement over VariSeed are reported without any description of how the 7820 ground-truth seed locations were established (number of annotators, consensus procedure, or distance tolerance for a correct detection). Because both the proposed regressor and the commercial comparator are scored against the same reference, any systematic bias in the labels renders the delta uninterpretable.

Authors: We agree that details on how the ground-truth seed locations were established are necessary to interpret the detection rates and the improvement over VariSeed. We will add this information to the revised manuscript, specifying the annotation protocol, number of annotators, consensus procedure, and distance tolerance. This will demonstrate that the same reference was used for both the proposed method and the commercial software, making the comparison valid. revision: yes
Referee: [Abstract] Abstract (methods section implied by the patient split): no architecture diagram, loss function, optimizer, training schedule, data augmentation, or cross-validation procedure is supplied, so the regression model itself cannot be assessed or reproduced.

Authors: We acknowledge the lack of these methodological details in the current manuscript. The revised version will include an architecture diagram and a comprehensive description of the loss function, optimizer, training schedule, data augmentation strategies, and cross-validation procedure to allow full assessment and reproduction of the regression model. revision: yes
Referee: [Abstract] Abstract: the 70/30 patient split is presented without scanner model, imaging protocol, slice thickness, or institution details, leaving the assumption that the test distribution matches future clinical cases unsupported.

Authors: We will include the missing information on the imaging acquisition in the revised Data section, detailing the scanner model, imaging protocol, slice thickness, and institution. This will support the relevance of the 70/30 split to clinical practice. revision: yes

Circularity Check

0 steps flagged

No circularity; purely empirical ML performance report on held-out data

full rationale

The paper trains a 3D FCN regression model on 5534 seeds from 70 patients and reports detection performance (2150/2286 seeds, 94.1%) on an independent 30-patient test set. No equations, derivations, fitted parameters renamed as predictions, or self-citation chains appear. The central claim is a direct empirical measurement against a commercial baseline on held-out cases; it does not reduce to its own inputs by construction. Ground-truth label quality is a separate validity concern outside the circularity definition.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on supervised training with accurate seed labels and on the assumption that the held-out test distribution matches future use cases; the network weights themselves are free parameters learned from the 70-patient training set.

free parameters (1)

network weights and hyperparameters
All convolutional filter values and training choices are fitted to the 5534 training seeds.

axioms (1)

domain assumption Ground-truth seed locations in the CT volumes are accurate and free of labeling bias
The regression targets are derived from these labels; any systematic error would propagate directly into the reported detection rate.

pith-pipeline@v0.9.0 · 5762 in / 1318 out tokens · 34424 ms · 2026-05-25T16:38:10.283358+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

13 extracted references · 13 canonical work pages

[1]

L., et al.: Cancer Statistics, 2019

Siegel, R. L., et al.: Cancer Statistics, 2019. Cancer J Clin. 69, 7–34 (2019)

work page 2019
[2]

Chin, J., et al.: Brachytherapy for patients with prostate cancer: american society of clinical oncology/cancer care ontario joint guideline update. J. Clin Oncol. 35, 1737-43 (2017)

work page 2017
[3]

Eur Urology 41, 434-439 (2002)

Stock, R., et al.: Importance of post-implant dosimetry in permanent prostate brachytherapy. Eur Urology 41, 434-439 (2002)

work page 2002
[4]

Phys Med Biol 48(9), 1191 - 1203 (2003)

Liu, H., et al.: Automatic localization of implanted seeds from post-implant CT images. Phys Med Biol 48(9), 1191 - 1203 (2003)

work page 2003
[5]

J., et al.: An automatic seed ﬁner for brachytherapy CT postplans based on the Hough transform

Holupka, E. J., et al.: An automatic seed ﬁner for brachytherapy CT postplans based on the Hough transform. Med Phys. 31(9), 2672-2679 (2004)

work page 2004
[6]

G., et al.: Automatic 3D seed location and orientation detection in CT image for prostate brachytherapy

Nguyen, H. G., et al.: Automatic 3D seed location and orientation detection in CT image for prostate brachytherapy. In: IEEE ISBI 2014, pp. 1320-1323. (2014)

work page 2014
[7]

Int J CARS

Zhang, G., et al.: Automatic seed picking for brachytherapy postimplant validation with 3D CT images. Int J CARS. 12, 1985 - 1993 (2017)

work page 1985
[8]

Med Imag Anal

Litjens, et al.: A survey on deep learning in medical image analysis. Med Imag Anal. 42, 60-88 (2017)

work page 2017
[9]

Med Imag Anal

Setio, A., et al.: Validation, comparison, and combination of algorithms for au- tomatic detection of pulmonary nodules in computed tomography images: the LUNA16 challenge. Med Imag Anal. 42, 1 - 13 (2017)

work page 2017
[10]

IEEE Trans

Xu, Y., et al: Gland instance segmentation using deep multichannel neural net- works. IEEE Trans. Med. Imaging 64(12), 2901-2912 (2017)

work page 2017
[11]

: The liver tumor segmentation benchmark (LiTS)

Bilic, P., et al. : The liver tumor segmentation benchmark (LiTS). arXiv preprint arXiv:1901.04056 (2019)

work page arXiv 1901
[12]

IEEE Trans

Yuan, Y., et al.: Automatic skin lesion segmentation using deep fully convolutional networks with Jaccard distance. IEEE Trans. Med. Imaging 36(9), 1876-1886 (2017)

work page 2017
[13]

: Automated melanoma recognition in dermoscopy images via very deep residual networks

Yu, L., et al. : Automated melanoma recognition in dermoscopy images via very deep residual networks. IEEE Trans. Med. Imaging 36(4), 994-1004 (2017)

work page 2017

[1] [1]

L., et al.: Cancer Statistics, 2019

Siegel, R. L., et al.: Cancer Statistics, 2019. Cancer J Clin. 69, 7–34 (2019)

work page 2019

[2] [2]

Chin, J., et al.: Brachytherapy for patients with prostate cancer: american society of clinical oncology/cancer care ontario joint guideline update. J. Clin Oncol. 35, 1737-43 (2017)

work page 2017

[3] [3]

Eur Urology 41, 434-439 (2002)

Stock, R., et al.: Importance of post-implant dosimetry in permanent prostate brachytherapy. Eur Urology 41, 434-439 (2002)

work page 2002

[4] [4]

Phys Med Biol 48(9), 1191 - 1203 (2003)

Liu, H., et al.: Automatic localization of implanted seeds from post-implant CT images. Phys Med Biol 48(9), 1191 - 1203 (2003)

work page 2003

[5] [5]

J., et al.: An automatic seed ﬁner for brachytherapy CT postplans based on the Hough transform

Holupka, E. J., et al.: An automatic seed ﬁner for brachytherapy CT postplans based on the Hough transform. Med Phys. 31(9), 2672-2679 (2004)

work page 2004

[6] [6]

G., et al.: Automatic 3D seed location and orientation detection in CT image for prostate brachytherapy

Nguyen, H. G., et al.: Automatic 3D seed location and orientation detection in CT image for prostate brachytherapy. In: IEEE ISBI 2014, pp. 1320-1323. (2014)

work page 2014

[7] [7]

Int J CARS

Zhang, G., et al.: Automatic seed picking for brachytherapy postimplant validation with 3D CT images. Int J CARS. 12, 1985 - 1993 (2017)

work page 1985

[8] [8]

Med Imag Anal

Litjens, et al.: A survey on deep learning in medical image analysis. Med Imag Anal. 42, 60-88 (2017)

work page 2017

[9] [9]

Med Imag Anal

Setio, A., et al.: Validation, comparison, and combination of algorithms for au- tomatic detection of pulmonary nodules in computed tomography images: the LUNA16 challenge. Med Imag Anal. 42, 1 - 13 (2017)

work page 2017

[10] [10]

IEEE Trans

Xu, Y., et al: Gland instance segmentation using deep multichannel neural net- works. IEEE Trans. Med. Imaging 64(12), 2901-2912 (2017)

work page 2017

[11] [11]

: The liver tumor segmentation benchmark (LiTS)

Bilic, P., et al. : The liver tumor segmentation benchmark (LiTS). arXiv preprint arXiv:1901.04056 (2019)

work page arXiv 1901

[12] [12]

IEEE Trans

Yuan, Y., et al.: Automatic skin lesion segmentation using deep fully convolutional networks with Jaccard distance. IEEE Trans. Med. Imaging 36(9), 1876-1886 (2017)

work page 2017

[13] [13]

: Automated melanoma recognition in dermoscopy images via very deep residual networks

Yu, L., et al. : Automated melanoma recognition in dermoscopy images via very deep residual networks. IEEE Trans. Med. Imaging 36(4), 994-1004 (2017)

work page 2017