In interstitial brachytherapy, a physician inserts a dozen or more hollow needles into a tumor, such as an advanced cervical cancer, and then drives a radioactive source through them. Dose falls off roughly with the inverse square of distance from the source, so a needle that sits a few millimeters off its intended track can either leave tumor underdosed or burn a nearby bladder or rectum. Clinics already reconstruct every needle precisely, but they do it afterward, on a post-implant MRI scan. During insertion, the only live view is transrectal ultrasound (TRUS), a probe placed in the rectum. TRUS is noisy, shadowed, and has never been annotated at scale. This proposal trains a TRUS needle segmenter with no hand-drawn ultrasound masks. The training labels are assembled from information the clinic already records about each implant: the MRI reconstruction made for treatment planning, the template hole each needle went through, the recorded insertion depth, and the needle count. Each of these is mapped into TRUS space and treated as a separate noisy "labeling function." A statistical label model estimates how much to trust each one, without ground truth, and merges them into soft voxel labels for the ultrasound network.
The approach adapts a known recipe to a new kind of task. Cross-modal data programming, in which rules are written over an easy modality and used to train a model on a hard one, works for medical diagnosis. Here it is moved from report-level diagnosis to millimeter-scale localization. The work also asks whether that move holds up when what matters is the dose a patient receives, not classification accuracy.
Core Claim
In a paired TRUS–MRI brachytherapy cohort, the planning MRI and procedure records can serve as the "easy modality" for weakly supervising intraoperative TRUS needle segmentation. When the registration-dependent sources are modeled explicitly as correlated, a label model that combines several disagreeing sources should give more accurate needle tips than the obvious shortcut of warping the MRI reconstruction into TRUS and calling it ground truth. The advantage should hold both geometrically and on a dose surrogate.
The claim rests on where the obvious shortcut is likely to fail. In the cohort that compared TRUS-guided needle measurements with post-implant MRI, agreement was high for needle-to-cervix distances (ICC > 0.90) but lowest for insertion depth (ICC 0.81), with the overall range 0.81 to 0.93 (Application of transrectal ultrasound in guiding interstitial brachytherapy for advanced cervical cancer). Depth, and therefore the needle tip, is exactly where a single registered label would carry the most error. It is also where a second, independent source, the recorded insertion length measured from the template, can correct it.
Prior Work and Its Roles
Three papers supply the method. Data programming (Data Programming: Creating Large Training Sets, Quickly) introduced the generative label model. This model infers how accurate each labeling function is from patterns of agreement and disagreement among them on unlabeled data, and it accepts a user-supplied graph of dependencies between functions. That second feature is essential here, because every source warped into TRUS space shares the same registration error. Snorkel (Snorkel: A System for Lightweight Extraction) turned this into a working system. Cross-Modal Data Programming (Cross-Modal Data Programming Enables Rapid Medical Machine Learning.) wrote rules over radiology text to train image and EEG classifiers. On radiographs it outperformed medium-sized hand-labeled sets by 8.50 ROC-AUC points. We keep its central move, supervising from an auxiliary modality that is absent at inference. We replace text with spatial sources and classification with dense segmentation, which that work explicitly did not attempt. The aortic-valve study (Weakly supervised classification of aortic valve malformations using unlabeled cardiac MRI sequences) showed that labeling functions can be written over geometric primitives extracted from images. Our ultrasound-side rules have the same character.
Two papers motivate the gap. The closest instrument work is Yang et al. (Weakly-supervised learning for catheter segmentation in 3D frustum ultrasound). They turned 3D bounding boxes into pseudo-labels using class activation maps, Frangi tubular filtering, and a conditional random field, and reached 66.6% Dice against 70.0% for full supervision on 60 ex-vivo porcine volumes. That pipeline has one supervision source, gives no account of how reliable its sources are, and handles a single catheter. Ours has multiple disagreeing sources, estimates their accuracy, and handles multi-needle in-vivo implants. Kostoulas et al. (Dealing with Segmentation Errors in Needle Reconstruction for MRI-Guided Brachytherapy) reached a median shaft error of 0.75 mm and tip error of 1.07 mm on prostate MRI. They used a fully supervised nnU-Net followed by geometric fitting and pruning that requires the exact needle count, and they left dose evaluation to future work. Their numbers set the accuracy bar; they are not a direct competitor on TRUS.
Two results constrain the design. The Motion-Aware needle paper (Motion-Aware Needle Segmentation in Ultrasound Images) found that ImageNet-pretrained encoders did worse than training from scratch on needle ultrasound. We therefore do not assume generic pretraining closes the gap, though we test it. BOXWRENCH (Stronger Than You Think: Benchmarking Weak Supervision on Realistic Tasks) defines weak supervision's value as a "crossover point," the number of clean labels a supervised model needs to catch up (for example 1,600 on ChemProt). It also shows that weak pretraining followed by clean fine-tuning is the fair way to compare. We adopt both. The data programming paper itself warns that unmodeled correlated sources make the label model misestimate accuracies, which is a direct threat in our setting and is addressed below.
Research Program
The unit of labeling is a voxel inside a corridor around each planned needle. The corridor is a cylinder a few millimeters wide, centered on the line from that needle's template hole along the planned direction, and extended past the recorded depth. Each voxel in the corridor gets a binary label: shaft or background. A separate set of ordinal "past the tip or not" votes along the corridor axis localizes the tip. Restricting labels to corridors keeps the class imbalance manageable and pairs every vote with a needle identity, so the needle count is built into the problem rather than enforced afterward.
The labeling functions come in two families. The first family is cross-modal. It includes the MRI reconstruction deformably registered into TRUS space, and the template-hole line paired with the recorded insertion depth. It also includes the same MRI track under an alternative rigid registration, which exists to expose registration disagreement. The second family works directly on ultrasound. It includes a bright-line-with-distal-shadow detector, a tubular filter in the spirit of Yang et al., and a parallelism rule against neighboring needles. It also includes a sweep-consistency rule that abstains on voxels whose response flickers across adjacent frames, borrowing the intuition behind the Kalman-style block in the Motion-Aware paper.
The label model is the data programming generative model with an explicit dependency graph. All registration-derived functions are tied into one correlated block, so their agreement is not mistaken for independent confirmation. Accuracies are estimated separately for depth bins along each corridor, which lets the model learn that the MRI source is trustworthy near the cervix and less so at the tip.
The resulting soft labels train a 3D segmentation network with a noise-aware loss, meaning a loss that weights each voxel by its label confidence. At inference the network sees only TRUS, followed by a light curve fit per needle that does not need the needle count. We also run the network in the raw polar "frustum" geometry, as Yang et al. did, to check that their tenfold compute saving survives.
Evaluation measures dose consequences as well as geometry. The reference is the clinical MRI reconstruction, with a held-out subset re-annotated directly on TRUS by a physicist. For each needle we place the planned source stops along the predicted track and along the reference track. We compute dose with a simplified point-source kernel that falls off as inverse square of distance and ignores tissue attenuation and source anisotropy, and we report it explicitly as a surrogate. We then compare target coverage and the maximum dose to the bladder and rectum between the two tracks. This addresses the dosimetric gap Kostoulas et al. left open, and it tests whether the noise in weak labels moves dose boundaries in a clinically meaningful way.
Baselines
The naive reference point is the classical pipeline with no learning: the bright-line and tubular-filter detectors applied directly to TRUS within the template corridors, followed by the same curve fit. It runs on the same test patients. To count as beaten, the proposal must improve median tip error by more than the spread between annotators on the TRUS-annotated subset.
The strongest competitor is single-source cross-modal supervision. This means deformably registering the clinical MRI reconstruction into TRUS, rasterizing it into hard masks, and training the identical network on them. It is the method any practitioner would try first, and it uses the same paired data. It gets the same architecture, the same training budget, and a registration-tuning budget equal to our label-model tuning budget, all selected on the same validation patients. No reported result on Lacuna exists for this baseline or for any TRUS needle segmentation in brachytherapy. The closest reported numbers come from MRI, not TRUS: 0.75 mm shaft and 1.07 mm tip error (Dealing with Segmentation Errors in Needle Reconstruction for MRI-Guided Brachytherapy). The closest reported number for weakly supervised 3D ultrasound is 66.6% Dice, on ex-vivo cardiac catheters (Weakly-supervised learning for catheter segmentation in 3D frustum ultrasound). Beating this baseline means a statistically significant reduction in median tip error across patients (paired test) together with no loss in shaft error. A shaft-only gain would not count.
Second, Yang et al.'s pseudo-labeling pipeline, reimplemented on TRUS with template-corridor boxes in place of hand boxes. This tests whether any well-built single-mechanism scheme already suffices.
Third, the fully supervised upper reference. This is nnU-Net trained on TRUS masks drawn by hand for varying numbers of patients, and also fine-tuned starting from our weakly supervised weights. Together these trace the crossover curve in the manner of BOXWRENCH.
Two ablations isolate the key ingredient. One runs the label model with the dependency graph removed, which tests the warning from data programming about correlated sources. The other removes the depth-binned accuracies, which tests whether per-depth trust is what fixes the tips.
Why not just…?
Why not just warp the MRI reconstruction into ultrasound and train on it? This is the real threat, and it may win. If deformable TRUS–MRI registration is good to well under a millimeter, extra sources add nothing. The available evidence points to depth as the weak axis (depth ICC 0.81 versus > 0.90 for lateral distances, from the TRUS cervical cohort), and the recorded insertion depth gives an independent check exactly there. The head-to-head comparison against this baseline, stratified by distance along the needle, settles the question. If the two tie at the tip, the honest conclusion is that a registration pipeline is all a clinic needs, and we will report it that way.
Why not just hand-label TRUS for thirty patients? Kostoulas et al. show that clean supervision plus geometric cleanup reaches sub-millimeter error on MRI. On noisy TRUS, though, annotators cannot see the tip any better than the network can, so hand labels inherit the depth ambiguity they are meant to remove. The crossover experiment measures how many hand-labeled patients a supervised model needs to match the weakly supervised one, and whether weak pretraining plus fine-tuning beats both. A small crossover, a handful of patients, would mean the label model is not worth its complexity.
Why not just use a pretrained foundation backbone? The Motion-Aware paper reports that ImageNet initialization hurt needle ultrasound segmentation, which argues against expecting much from generic features. Pretraining is also orthogonal to the labeling question: it changes the starting weights, not the targets. We include a pretrained-encoder variant of the strongest baseline. If it closes the tip-error gap, the gain did not come from the label model.
Decisive Experiments
The primary study needs a paired TRUS–MRI cervical interstitial cohort of the design in Application of transrectal ultrasound in guiding interstitial brachytherapy for advanced cervical cancer, which covered 48 patients and 337 needles. The file does not list that cohort as public, so the study assumes institutional access to a cohort of that design. Splits are by patient, and a TRUS-annotated test subset is labeled by two physicists.
Metrics are median and 90th-percentile tip and shaft error, the needle detection rate when the needle count is not given, and calibration of the label model's per-depth accuracy estimates against TRUS annotations. On the dose surrogate, we report the change in target coverage and in maximum bladder and rectum dose.
The proposal is supported if three things hold. The multi-source label model beats MRI-warping in tip error. Removing the dependency graph measurably worsens it. The geometric gain carries over to smaller changes in dose-surrogate coverage.
A secondary check reuses the Amsterdam prostate MRI setting from Kostoulas et al. There we replace their hard needle-count pruning with the same label model over template, depth, and count sources, and ask whether count-free reconstruction approaches their reported 0.75 mm shaft and 1.07 mm tip errors. A clear failure there would indicate that the label model's statistical denoising cannot substitute for hard geometric priors at sub-millimeter scales. That would be a useful negative result on the tension between weak supervision and dose-critical precision.
References
- Application of transrectal ultrasound in guiding interstitial brachytherapy for advanced cervical cancer
- Data Programming: Creating Large Training Sets, Quickly
- Snorkel: A System for Lightweight Extraction
- Cross-Modal Data Programming Enables Rapid Medical Machine Learning
- Weakly supervised classification of aortic valve malformations using unlabeled cardiac MRI sequences
- Weakly-supervised Learning For Catheter Segmentation in 3D Frustum Ultrasound
- Dealing with Segmentation Errors in Needle Reconstruction for MRI-Guided Brachytherapy
- Motion-aware Needle Segmentation in Ultrasound Images
- Stronger Than You Think: Benchmarking Weak Supervision on Realistic Tasks