Systematic review of artificial intelligence in brachytherapy

Purpose

Artificial intelligence (AI) is now a key component of medical research, with radiation oncology being a leading proponent of its incorporation into new technology [1]. Many review articles evaluated AI in radiation oncology [2-16]. Algorithms have demonstrated a decrease in the time needed for treatment planning, contouring, and image processing for radiation therapy treatments [9]. Dose prescription [17], image registration [9], tumor segmentation [18], and treatment validation [19], have been performed by AI in radiation oncology with similar or improved performance when compared with humans.
It is predicted that AI tools applied to external beam radiation therapy can potentially solve brachytherapy (BT) treatment challenges with proper adjustments [20], but data on the subject is sparse. BT treatment has several challenges, including long planning time, consistency across planners, and correct identification of target structures [1]. BT, like other forms of radiotherapy, relies on the performance of many repetitive and automatable tasks. AI has the potential to refocus experts talents on more complicated, manual tasks while increasing the accuracy and efficiency of BT treatment. Clinicians and patients would benefit from the use of AI in BT, as it has already been partially applied from planning procedures to treatment [20], with the potential to improve treatment quality on all fronts.
While many studies evaluate the use of AI in radiation oncology, there is no single comprehensive, systematic overview regarding the clinical use of AI in all aspects of BT treatment planning, treatment, imaging, and prognosis. At present, only eight review papers exist on the use of AI in BT [6, 7, 20-25], including recent works by Chen et al. [6] and Fionda et al. [25], which largely provide narrative overviews but are limited in scope. For example, Fionda et al. emphasized AI in interventional radiotherapy primarily from a patient-centered perspective, while Chen et al. summarized technical applications across planning and imaging. One review focused solely on the applications of deep learning [21]. However, neither followed the Preferred Reporting Items for Systemic reviews and Meta-Analyses (PRISMA) methodology, nor did they conduct a systematic, multi-database literature search or provide detailed extraction of study-level performance data. In the current study, a systematic, comprehensive review of AI in BT literature was conducted, and analyzed investigating application to BT, AI description, training and testing datasets, input and output of AI, treatment description, ground truth classification, accuracy compared with ground truth, and time for results. Comparison of the use of AI techniques to each other and to standard-of-care conventional BT in terms of efficiency, accuracy, and practical application, was also performed when possible.

Material and methods

Literature search

This systematic review followed the guidelines described in PRISMA (Figure 1). Inclusion criteria for literature searches included any peer-reviewed journal articles published from January 1, 1980 till August 1, 2025 in PubMed, Google Scholar, Cochrane Library as well as University of Iowa and University of Texas library catalogs. These databases were selected for their comprehensive biomedical and clinical coverage. Scopus was not included, since preliminary search showed substantial overlap with the above sources. Searches included “Brachytherapy” in combination with one or more of the terms “Artificial Intelligence”, “Deep Learning”, “Machine Learning”, “Neural Networks”, “AI”, “DL”, “automation”, “radiomics”, and “convolutional”. Selected bibliographies, citations, and meetings or presentations were also explored for relevant papers. Studies were excluded if no English version was available, they were not peer-reviewed, or if they were unrelated to AI in BT. Exclusions were performed using built-in automated filters of each database (publication date, English language, peer-reviewed journal articles, etc.). Studies only mentioning an application to BT with no elaboration or evidence were also excluded. Determination of inclusion was done by three independent investigators, who reviewed titles, abstracts, and articles. Data collected on each study included title, authors, year of publication, AI technique/method, AI application, disease site, and comparison (conventional BT vs. other AI-based BT methods).

Review and synthesis

Included studies were sorted by disease site and BT application. Training and testing dataset information, accuracy metrics, comparison with traditional techniques, and time for AI to produce output, were included if described clearly in a manuscript. A meta-analysis, a way of comparing a large number of studies quantitatively to find trends in data [26], was not performed in this paper due to studies being too heterogeneous. Heterogeneity arose from differences in disease site, applicator type, imaging modality (trans-rectal ultrasound [TRUS], magnetic resonance imaging [MRI], computed tomography [CT]), and AI methodology. Reported outcomes and performance metrics also varied widely ranging from Dice similarity coefficient (DSC)/Hausdorff distance (HD) for segmentation to AUC or DVH metrics for prediction and planning, with inconsistent ground truth definitions and validation approaches. Thus, these variations did not allow for a valid quantitative meta-analysis.
For the purpose of this review, AI was defined as a computer algorithm performing tasks, which usually require human intelligence [9]. We further categorize algorithms as deep learning (DL) or machine learning (ML). Machine learning was defined as a form of AI that did not involve neural networks, such as radiomics and unsupervised, supervised, and reinforcement learning techniques. Deep learning techniques included any approach using neural networks. Many of deep learning systems were convolutional neural networks (CNN), artificial neural networks (ANN), or variations of the two through mainly either U-Net architecture or residual networks (Res-Net) architecture. Artificial neural networks (ANNs) are used to describe fully connected, feed-forward models processing non-image inputs, such as radiomic, dosimetric, geometric, or clinical features. Convolutional neural networks (CNNs), by contrast, are architectures designed for image-based data, where convolutional filters and weight sharing enable the network to capture local spatial structure, and only the final layer is fully connected [27]. Accordingly, most image-driven applications in brachytherapy, such as segmentation or applicator localization, use CNNs, while ANNs are applied to structured or tabular data. In contrast to prior reviews of Chen et al. [6] and Fionda et al. [25], which provided primarily narrative overviews, our study applied these operational definitions to a PRISMA-guided, multi-database systematic review of AI in brachytherapy, and explicitly distinguished between algorithms evaluated retrospectively in pre-clinical settings and those tested in prospective or routine clinical use.

Results

Databases searches yielded 7,723 cumulative results. A total of 7,231 studies were excluded based on date restrictions, language, not being a peer-reviewed journal article, full-text of study unavailable, or duplicates. An additional 308 were determined to be out of scope after reviewing of the abstract or full-text, leaving 140 accepted, peer-reviewed articles included in the final synthesis, as shown in Figure 1. A total of 35 studies used machine learning AI techniques [28-61], and 179 employed deep learning [22, 24, 62-211].

AI in BT per application

We classified studies with a major separation of AI according to the steps required for types of BT planning and treatment. The category of applicator, needle, and seeds reconstruction, considered any reconstruction of applicator or needle studies, also referred as digitization or segmentation in the literature as well as seed localization studies. Papers on contouring or segmentation of organs and target volumes were categorized as “Segmentation”. One study was included in both the applicator reconstruction and segmentation sections [111]. Studies were considered in the “Dose calculation and optimization” category if AI was used in any part of dose calculation or dose optimization (e.g., optimizing dwell times and/or dwell positions) during treatment planning. Reports involving treatment planning that did not fit into any of the listed categories were classified as “Other treatment planning”.
AI in BT outcome predictions included studies on predicting side effects or toxicity, regression or relapse of disease, or survival rates after BT treatments. Scientific research applying AI to other aspects of imaging used in BT were classified as “AI in BT imaging”, and they could be further grouped as belonging to applications of AI in image registration, image-artifact removal, or another imaging application. Tables 1-5 present study characteristics and performance results sorted by each application of BT. Additional study characteristics and performance results are available in Appendices 1-5.

AI in BT treatment planning

Intracavitary applicator, needle, and seeds reconstruction

Artificial intelligence was introduced to intracavitary applicator, needle, and seed reconstruction to reduce time needed while maintaining the accuracy of a planner. Most studies were retrospective in nature, relying on archival imaging datasets, with expert-based reconstruction (localization) of applicators, needles, and/or seeds as ground truth. However, four studies [29, 33, 42, 43] were not retrospective: Lee et al. [33] evaluated a prospective tool for automated planning volume definition in soft-tissue sarcoma brachytherapy; Shiradkar et al. [42] developed a computational radiomics-based framework for prostate planning; Stenhouse et al. [43] proposed a machine learning model for applicator selection in cervical brachytherapy; and Deufel et al. [29] validated an automated applicator digitization algorithm using thresholding and clustering approaches. Of the 38 papers using AI for reconstruction, four studies utilized machine learning techniques, while the rest employed deep learning, mainly convolutional neural networks. None of the studies in this group relied solely on non-machine learning/deep learning automation approaches (e.g., rule-based, clustering, or thresholding methods); all used either machine learning or deep learning. Machine learning methods (e.g., radiomics with classical classifiers) work with smaller datasets, and are more interpretable, though constrained by feature design and generalizability. Deep learning methods (e.g., CNNs, U-Net, Res-Net) achieve the highest accuracy and adaptability, but demand larger datasets, superior computation, and pose interpretability challenges. In 63% of studies, high-dose-rate (HDR) BT needle [33] or low-dose-rate (LDR) seeds reconstruction [18] for prostate cancer patients were investigated. The remaining fourteen studies explored the use of AI in intracavitary applicators [17] needles [19], or reconstruction for gynecological cancer patients. In 89% of studies, training datasets of less than 100 treatment plans were utilized, with the largest containing 713 plan datasets [64] and the smallest, 4 [120]. Of the studies, 26% did not indicate numbers used specifically for independent validation or testing performed with datasets not used in training or other validation methods. N-fold cross-validation was performed in seven studies [68, 69, 113, 114, 122, 133, 157], with independent testing datasets ranging in size from 2 [47] to 389 [64].

Intracavitary applicator reconstruction for gynecological cancers

Artificial intelligence-based applicator reconstruction performance was mainly reported in terms of Hausdorff distance (HD), the largest difference between two reconstructions (experts vs. AI), and Dice similarity coefficient (DSC), a measure of the amount of volume overlaps the two segmentations have in terms of percent overlap [212]. Fifteen studies [79, 80, 85, 111, 120, 122, 127, 129-131, 133, 170, 209-211] employed deep learning, with two studies using machine learning [29, 32], while nine and three utilized CT [29, 79, 85, 111, 129-131, 161, 170] and MRI [32, 120, 127], respectively. Two employed titanium tandem and ovoid (T&O) applicators [29, 85], one reported Fletcher-Williamson T&O applicator [111], with the other two using a plastic T&O applicator [79] and a plastic tandem and ring (T&R) [32], respectively. One paper focused on the reconstruction of plastic catheters in MRI-guided interstitial brachytherapy for gynecological cancer [127]. All intracavitary applicator reconstruction studies, regardless of the applicator type or type of imaging modality, presented HD values between < 1 mm and 5.8 mm, DSC values of 88.3-93.1, or less than 1 mm errors. For the three metallic T&O cases (imagined on CT), HD between AI and manual for tandem ranged from less than 1 mm to 1.9 mm [29, 85, 111]. For ovoids only, this value ranged from less than 1 mm to 3.9 mm [29, 85, 111]. For the reconstructed paths, the average accuracy was within 1 mm for all channels [85, 111]. For plastic T&O applicators on CT, Hu et al. reported a segmentation DSC of 0.89 and an average tip error for reconstruction of 0.80 mm [79]. A study that examined plastic T&R applicators on MRI, achieved dwell position errors of 0.72 mm for the ring and 0.70 mm for tandem with a machine learning algorithm, significantly (p < 0.05) better than inter-observer variability [32].

Interstitial needles for prostate or gynecological cancers

Of the 26 studies reconstructing interstitial needles, six focused on metal needles [86, 100, 108, 129, 131, 170], and two on plastic needles [127, 211] in gynecological cancer patients, while the remaining fourteen [64, 68, 69, 80, 107, 112-114, 122, 130, 157, 158, 160, 171] focused on prostate cancer using either plastic (8/14; 57%) or metal needles. All studies, except Grigo et al. [159], utilized CNNs. Across all interstitial needle reconstruction studies, the reported performance varied by disease site and imaging modality. Gynecological HDR studies (mostly CT- or US-based) presented relatively consistent results, with HD values in the range of ~0.7-3.0 mm and DSC values of ~88-93%. In contrast, prostate HDR studies demonstrated wider variability, particularly in MRI-based reconstructions, with HD values spanning 0.7-15.9 mm and DSC values from ~50% to 95%. When aggregated, these studies collectively reported HD values of 0.71-15.9 mm, DSC values of 50-93.7%, or < 2.0 mm errors. For gynecological cases, a CT study [86] reported a needle reconstruction DSC of 0.93 and an HD of 0.71 ±0.40 mm for the needle-tip difference in needle reconstruction, reconstructing all needles, including cases with needles touched or crossed. A gynecological interstitial needle case study using MR images showed an average DSC of 60% and an error of 2.0 mm for closely spaced catheters [108], while one paper on US images achieved errors of 1.03 mm and 0.62 degrees [100]. In addition, Straathof et al. [211] reported a retrospective planning study using CT and MRI datasets to optimize curved channel configurations in 3D-printed, patient-tailored applicators for cervical cancer brachytherapy, highlighting the potential of AI-driven personalization beyond standard applicator geometry.
In a series of studies using deep learning for multiplastic needle reconstruction in TRUS images of HDR prostate cancer patients [112-115], reconstruction errors ranged from 0.33 mm to 0.44 mm, with 96-98% of needles successfully identified. Similarly, Andersén et al. [64] and Wang et al. [107] demonstrated mean errors below 1 mm for TRUS-guided prostate HDR needle digitization, with high needle detection rates. For MRI-guided HDR prostate brachytherapy with plastic needles, Dai et al. [68, 69] and Aleong et al. [157] reported that 82-87% of needles were detected within a 2 mm error, with mean reconstruction errors generally below 1 mm across independent cohorts. For TRUS- and MRI-based metallic interstitial needle reconstruction, more recent work has demonstrated further accuracy gains. Moradi et al. [160] and Goulet et al. [158] showed a mean tip errors of 0.41-0.72 mm and average errors of 0.44 ±0.32 mm, respectively, while Kobayashi et al. [171] explored MRI-based prediction of needle deflection, reporting HD95 values up to 4.9 mm, depending on extra-capsular extension. Across all included studies, expert manual digitization served as the ground truth for training and evaluation. Although explicit measures of human reconstruction accuracy were generally not reported, the observed differences between AI-based and manual reconstructions, 0.4-0.7 mm in Moradi et al. [160] and Goulet et al. [158], were consistent with the expected intra- and inter-observer variability.

Seed identification in prostate seeds implant (PSI)

Six studies [39, 47, 76, 101, 133, 159] investigated AI algorithms for iodine-125 (125I) seed detection in post-implant dosimetry. Two applied traditional machine learning approaches [39, 47], and four used convolutional neural networks [76, 101, 133, 159]. The datasets were generally small, with test sizes ranging from a single phantom plan with 20 dummy seeds [39] to 40 clinical MRI plans with varying numbers of seeds [101]. Across studies, AI-based seed identification achieved detection rates of 64-97.6%, and localization errors between 0.19-1.44 mm. In TRUS images, one study [76] reported that 64% of seeds could be detected, whereas another [47] found a mean localization error of 1.44 mm with 4.3° angular deviation. In an MRI-based study of Sanders et al. [101], a recall of 97.6% with 1.9% precision was achieved, with a mean localization error of 0.19 mm, while Nosrati et al. [39] demonstrated comparable accuracy using an unsupervised ML method on phantom data. More recently, Yuan et al. [133] applied deep learning to fluoroscopy images, resulting in a mean average precision of > 0.91. Grigo et al. [159] developed an MRI-only post-implant dosimetry workflow using a 3D nnU-Net, reporting mean centroid errors of 1.8 ±1.3 mm and mean local deviations of 0.9 ±0.9 mm, demonstrating feasibility of CNN-based MRI-only workflows for prostate LDR brachytherapy.

Organs at risk and target volumes segmentation

Artificial intelligence has also been broadly applied to the segmentation of organs at risk (OARs) and target volumes in BT across both prostate and gynecologic tumor sites [63, 73-75, 77, 82, 84, 88, 90-92, 95-98, 102, 110, 111, 118, 119, 121, 123-126, 128, 132, 134, 135, 137-140, 142, 150, 155, 162-170, 172-189]. In total, 74 studies employed deep learning techniques, with approximately 85% implementing CNN-based architectures [63, 73-75, 77, 82, 88, 90, 91, 95, 97, 98, 110, 111, 123, 125, 126, 128, 132, 134, 135, 137-140, 155, 162-170, 172-189]. For prostate brachytherapy, most efforts focused on TRUS segmentation [88, 90-92, 96-98, 110, 123, 138], supplemented by MRI- and CT-based investigations [57, 58, 63, 73-75, 102, 118, 125, 126]. For gynecologic brachytherapy, segmentation was explored primarily on MRI datasets [95, 111, 119, 124, 132, 134, 135, 137, 139, 140, 142, 155, 162-170, 172, 174-189], with several complementary CT-based studies addressing post-operative or intracavitary workflows [77, 84, 121, 128, 139, 155]. A smaller sub-set of studies targeted endometrial carcinoma [170] and parotid gland brachytherapy [141]. A domain-adversarial neural network was employed in one study for the segmentation of clinical target volume (CTV) in gynecological cancer patients undergoing post-operative vaginal brachytherapy with CT imaging [155]. Most research (36%) involved prostate cancer patients [57, 58, 63, 73-75, 88, 90-92, 96-98, 102, 110, 118, 123, 125, 126, 138, 166, 170, 172, 176, 185, 189]. Studies used training datasets ranging from 7 patients [124] to 771 patients [92], with validation dataset sizes ranging from 7 [110] to 193 patients [92]. Some authors (30%) did not use an independent testing dataset, but testing dataset sizes as small as 5 patients [77] and as large as 280 patients [96] were reported.
As far as the type of validation methods concerned, N-fold cross-validation [90, 91, 96, 128, 142, 165, 166, 172, 174, 175, 181, 188] and cross-institution validation [82, 95, 119, 125, 134, 135] were reported. Segmentation performance was shown in terms of HD only (18/63), DSC only (39/63), and both HD and DSC (31/63). In 32% of studies, a DSC of 90% or higher was reported [57, 58, 63, 73, 74, 84, 88, 90, 91, 95, 97, 98, 101, 110, 123, 152, 165, 166, 170, 172, 174, 175, 185, 188, 189]. Regarding HD, all studies presented a value of less than 16 mm, with most (35%) being under 5 mm [63, 73, 77, 88, 90, 91, 95, 98, 102, 121, 123, 128, 132, 134, 137, 141, 155, 165, 166, 170, 172, 174-176, 185, 188, 189]. All research demonstrated computation times under 300 seconds, with majority (33%) of studies [63, 73, 75, 88, 96-98, 165, 172, 182, 184, 187, 189] reporting times under 1 second. For the 17 studies reporting prostate cancer segmentation on TRUS images [57, 58, 63, 73, 74, 88, 90, 91, 96-98, 123, 138, 155, 166, 170, 184], DSCs ranged from 88% [75] to 97% [73], whereas HDs for these studies ranged from 2.3 mm [91] to 8.4 mm [75]. Four studies used deep learning to segment the prostate on CT, with DSCs reported as 83% [118], 85% [92], 89% [74], and 95% [73], while HDs ranged from 5.2 mm to 6.3 mm. A study by Sanders et al. compared the performance of 120 different deep learning algorithms on segmentation for prostate cancer [102]. They found that the highest performing model, fully convolutional network (FCN), resulted in DSCs ranging from 70% (external urinary sphincter) to 96% (bladder). Another case that utilized MRI, but as priors for TRUS segmentation, achieved a DSC of 90.9% on the prostate with a CNN [110]. For all prostate segmentation cases, the base and apex regions were reported as being segmented with lower accuracy than the mid-gland region.
For OAR segmentations, the bladder, rectum, and sigmoid had DSCs of 85-96%, 66-97%, and 66-92% on CT for gynecological cancer, respectively [95, 111, 128, 139, 165]. For OAR segmentations on MR images for prostate cancer, the highest DSCs values were 96% for the bladder and 91% for the rectum [102], while a study on TRUS images reported DSCs of 75%, 86%, and 90% for the bladder, urethra, and rectum, respectively [91]. A study by Iantsen et al. used a U-Net to segment the sub-tumor region to identify functional uptake regions in PET images for gynecological cancer patients, achieving an average DSC of 80% [82]. Another paper by Jeuthe et al. employed CT images to train and test their machine learning and deep learning models that segmented pelvic bones, and observed that the highest DSC of 98% was found using a deep learning method [84]. A study by Li et al. [141] reported on deep learning-based two-step auto-segmentation model for OAR in parotid gland carcinoma patients using CT. Clinical acceptability was defined quantitatively, with most OARs achieving a DSC greater than 80%, which the authors considered the threshold for clinically satisfactory segmentation. The method produced results in under one minute per patient.

Dose calculation and optimization

Forty studies [37, 38, 49-51, 53, 55, 70, 89, 93, 94, 99, 103, 106, 117, 136, 143, 144, 146, 148, 150, 151, 153, 154, 156, 190-201, 213-215] investigated AI applications for dose calculation, dose prediction, and treatment optimization in BT.
Needle placement optimization: Five studies focused on optimizing or assisting with needle placement and localization prior to planning [37, 38, 143, 148, 154]. These attempts primarily addressed geometric optimization of applicator or needle positioning as a foundation for subsequent planning.
Dwell time and treatment plan optimization: Nine studies investigated treatment plan optimization, primarily in the form of dwell time distribution or overall plan quality enhancement [94, 99, 103, 106, 143, 148, 154, 190, 193]. Among these, five explicitly examined optimization strategies using AI [37, 38, 143, 148, 154].
Dose calculation and dose prediction: Four papers directly addressed dose calculation tasks [70, 89, 93, 153], while seven focused on dose prediction models [52, 136, 144, 146, 150, 151, 156]. In this context, “prediction” referred to models forecasting dose distributions or dose volume histogram (DVH) parameters, which were then used to guide dwell time optimization or validate plan quality. One investigation specifically reported dose calculation with tissue heterogeneity modeling using machine learning [117].
The majority of research (20/23) employed deep learning techniques [50, 70, 89, 93, 94, 99, 103, 106, 136, 143, 144, 146, 148, 150, 151, 153, 154, 190, 193]. One study used machine learning [117], another reported a neural network approach [156], and one applied a knowledge-based method [52]. No study showed an independent testing dataset larger than 200 patients. Across all clinical sites, median p-values were > 0.05, indicating no significant difference between AI-predicted outputs and ground truth clinical plans.
For the optimization cases, twelve studies [37, 38, 50, 53, 55, 99, 106, 190, 198, 200, 210, 213] used data input from prostate cancer patients, of which three focused on LDR prostate brachytherapy [37, 38, 190], eight on HDR prostate brachytherapy [50, 53, 55, 99, 106, 198, 200, 213], and one included both HDR prostate and cervix [210]. Barten et al. [213] demonstrated the first prospective clinical experience of AI-based automated HDR prostate planning, while Dickhoff et al. [210] developed and validated a framework that extended multi-objective AI optimization for both HDR prostate and cervical BT. Rajković et al. used an ANN to optimize treatment parameters (dose, coverage index, prostate-specific antigen [PSA] nadir 1, and the risk of recurrence) in prostate HDR-BT patients of varied risks of biochemical failure [99]. The model achieved an R value of greater than 0.7, and a mean squared error and deviation from reference that both, along with the R value, were within acceptable limits. The other five studies reported DVH parameter errors of between 0.1% and 5.1% [37, 38, 106, 143, 148], with only one significant DVH parameter difference in one study [37]. There was also a significant decrease in planning time [38]. One optimization paper focused on gynecological cancer, and improved the quality score of human-created plans by 10.7% by using an AI-based reinforcement learning framework to adjust organ weights [103]. Another optimization paper that did not specify a clinical site [94], optimized needle locations in BT treatment plans under a simulated phantom setting without validation via clinical imaging and treatment plans, and reported the root mean square errors of each needle’s position to be 3 mm.
Deep learning has been applied for diminishing the time needed to perform heterogeneity-corrected dose calculations using Monte Carlo simulations [106, 117, 136]. Five studies [89, 93, 136, 146, 151] used CNNs with CT images as inputs to generate a predicted dose distribution. Using the clinically delivered dose as ground truth, one study reported p-values of greater than 0.01 for all DVH metrics, suggesting no difference between deep learning and clinical dose [89]. The other study [93] reported the Monte Carlo simulation as ground truth, with differences between ground truth and their algorithm of under 3% for all DVH metrics. Yet another study [70] employed a large consensus dataset and the inverse problem theory in combination with ANNs to improve the Sievert integration model, achieving an approximation error of 3.3% in radial dose-rate.

Other planning procedures

Fourteen studies used AI for other planning procedures [30, 33, 42, 43, 61, 65, 66, 71, 83, 105, 149, 202, 203, 216]. Shiradkar et al. was able to reduce the prescribed dose delivered to prostate cancer patients to the whole gland by 43% using a new plan based on intensity and radiomic features of the tumor [42]. Gynecological treatment plans were also corrected, with an ANN estimating variations in intra-fractional organs doses, and then modifying the treatment plan in 10-14 of 30 plans while not changing the target dose [83]. Another study involving the verification of gynecological plans achieved an accuracy of around a 1-pixel error for position, and came within 2% of planning system calculations for dwell time [71].
Stenhouse et al. used machine learning methods to develop a model that predicted whether intracavitary or interstitial applicators should be applied in gynecological HDR-BT, and were able to predict which type of applicator should be utilized with a 91.5% predictive accuracy [43]. A study reconstructing catheters and removing artifacts while considering patient movement via electromagnetic tracking data in CT images for BT of breast cancer, achieved results qualitatively similar to human planners [30]. Another paper used deep learning to predict the insertion force of needles for prostate BT, and reported errors of 4.5% [66]. Similarly, Avila-Carrasco et al. developed an ANN to estimate the deflection of needles in soft tissue phantoms, with prediction errors ranging from 0.29 to 0.62 mm [65].
A study by Sina et al. utilized an ANN and CT numbers of tissues to determine the composition of tissues in treatment planning, and predicted mass density with errors of only 0.52% for bone and –0.95% for muscle [105]. Another application was studied by Lee et al., who smoothly outlined planning volumes of soft-tissue sarcoma patients with machine learning inputs of catheter positions and PTV volume, achieving clinically acceptable outlines in 2 minutes [33].
Wang et al. demonstrated that HDR prostate BT plan quality can be predicted directly from TRUS images, enabling real-time feedback during the procedure and potentially reducing reliance on time-intensive post-planning evaluations [216]. Beyond prostate, Wang et al. also developed a machine learning-guided, 3D-printed 32P hydrogel applicator for skin hemangiomas, which allowed customizable dose distributions and highlighted the versatility of AI-driven personalized applicator design in non-traditional brachytherapy sites [61]. In cervical cancer HDR-BT, Jiang et al. applied deep learning to select between Syed and T&O applicators, achieving high accuracy in applicator choice and offering a tool to support clinicians in complex treatment decisions [203]. Complementing planning and applicator selection, Hoffer et al. introduced automated thermal imaging to monitor local response to cervical cancer brachytherapy, providing a non-invasive and real-time method to assess treatment effectiveness during therapy [202].

AI in BT outcome predictions

Artificial intelligence was broadly applied for predicting outcomes of different aspects of BT in 27 papers. Of the 27 studies, nine [34, 35, 45, 46, 56, 60, 204, 206, 217] focused on prostate, and sixteen [28, 31, 36, 40, 41, 44, 48, 59, 104, 116, 145, 205, 218-221] were dedicated to gynecological cancer. Most [28, 31, 34-36, 40, 41, 44-46, 48, 54, 56, 59, 206, 218-221] papers used machine learning/radiomics techniques, with four utilizing deep learning [87, 104, 116, 145, 204]. AUC (area under the receiver operating characteristic [ROC] curve; a quantitative accuracy measurement where higher values indicate better classification performance), p-values, and sensitivity/specificity, were the most commonly reported performance metrics, with studies focusing on predicting failure, toxicity, or recurrence/relapse after brachytherapy. Most prediction studies reporting AUCs had values of ≥ 0.70; however, a minority showed AUCs in the range of 0.63-0.69 or lower for specific endpoints/sub-sets [40, 220]. All AUC values were reported in the respective studies while were not re-calculated in this review.

Predicting toxic side effects

Using AI to predict toxicity or formation of toxic structures, such as fistulas or urethral strictures after BT treatment, was studied in eight papers [36, 44, 45, 59, 116, 145, 217, 219]. Six [36, 44, 59, 116, 145, 219] involved gynecological cancer cases, while the other [45] considered prostate cancer patients. For the gynecological cases, studies predicting fistula formation [44], rectum toxicity [116], and late bladder toxicity [145] achieved AUCs of 0.904 and 0.7 and AUROC of 0.81, respectively. For acute and late toxicity per area, radiomics resulted in averaged sensitivity and specificity of 73-78% and 64-78% [36]. Additional research applied AI to predict broader toxicity outcomes, including grade 3+ toxicities in gynecologic cancer patients treated with HDR brachytherapy, with F1 scores up to 0.63 and accuracies of 0.80-0.85 [59], whereas radiation-induced vaginal injury after cervical cancer radiotherapy achieved an AUC of 0.975 [219]. For the prostate case, machine learning methods predicted structure formation, and found only two MRI radiomic features, i.e., homogeneity and contrast, to be significant [45]. More recent work has also applied AI to identify rectal sub-regions associated with toxicities in prostate radiotherapy, observing significant associations between posterior rectal regions and increased toxicity risk [217].

Failure after treatment, progression, or survival prediction

In fourteen papers, AI was used to predict progression of disease, recurrence, or failure/mortality after BT treatment [31, 35, 40, 41, 46, 48, 54, 56, 87, 104, 204, 206, 218, 221]. The two prostate cancer studies in this section employed machine learning, and predicted biochemical failure using the Phoenix definition [46] and local recurrence [35]. The biochemical failure rate was 0.75 in patients with algorithm-designated parameters (0.38 for conventional population), and 17 selected radiomic features for local recurrence had p-values of 0.05. One choroidal melanoma study [87] used clinical and demographic inputs to predict 5-year mortality post-BT treatment, and achieved an AUC value of 0.81. In another study, a machine learning model was employed to forecast a four-year risk of metastasis and mortality in uveal melanoma using ultrasound datasets, achieving values of 0.730 and 0.846 for predicting metastasis based on one-year and three-year follow-up data, respectively [54]. Five research focused on predicting response to treatment in gynecological cancers [31, 40, 41, 48, 104]. Local and distant relapse were projected using deep learning, with accuracies of 89% and 87% using fluorodeoxyglucose positron emission tomography (FDG PET) and CT data as inputs [104]. Machine learning was employed to predict local recurrence, distant failure, and overall survival/ local control, showing AUCs of 0.63-0.76 with 8 features [40], 0.83 with a created shell feature [31], 0.75 with 11 features [41], and 0.75-0.82 with 10 features [48]. Hao et al. [31] also used their features to predict distant failure in non-small cell lung cancer patients, with an AUC value of 0.82.

Other predictions

More broadly, one study used machine learning to identify changes in apparent diffusion coefficient (ADC)-based radiomic features, such as histogram and texture descriptors (i.e., mean, skewness, kurtosis) in prostate cancer patients undergoing external beam radiation therapy (EBRT) and BT with dose escalation [34]. They were able to determine which features from 101 tested underwent the most change during treatment. Similarly, Bowen et al. performed a study to track tumor heterogeneity using intensity histogram and texture features derived from multi-parametric MRI/PET imaging after treatment [28], and were able to determine features undergoing the greatest changes.

AI in BT imaging

Image artifact removal and multi-image registration

Artifacts can decrease image quality, hence implementing deep learning may help to remove artifacts from images. One paper showed a reduction of metal artifacts from gynecological CT images in less than one second, with a resulting mean peak sound-to-noise ratio of 38.09 dB, which was higher than the 37.79 dB obtained from the conventional CNN [81]. In medical imaging, PSNR values above ~30 dB are generally considered good image quality, while values below ~20 dB are considered poor. Thus, the reported PSNR values in this study indicated high-quality reconstruction. For US, Allman et al. applied a CNN to identify noise artifacts [62]. With no excess noise, the CNN was able to identify 100% of sources and artifacts from 3,468 test images, but added noise greatly decreased performance with 98% identified at –9 dB or greater, and only 4.35% identified at –21 dB. Image reconstruction was also performed by deep learning in prostate cancer patients, with two studies registering MR and US images, achieving a DSC of 97% [67] and a target registration error of 2.53 mm [109], respectively. Salehi et al. demonstrated the viability of employing deep learning for deformable image registration to address inter-fraction organ motion in locally advanced cervical cancer patients undergoing HDR brachytherapy. This study utilized radiotherapy structure data to facilitate local image deformation. Notably, the mean Dice indices attained for the cervix, bladder, and rectum were 0.89 ±0.02, 0.96 ±0.01, and 0.93 ±0.02, respectively [152]. In another study, Ecker et al. trained 3D and 2D neural networks to predict applicator structures from MR images, followed by automatic applicator-based image registration for MRI-guided cervical HDR brachytherapy. The results showed mean Dice coefficients of 0.70 ±0.07 and 0.58 ±0.04 for 3D and 2D neural networks, and registration algorithms achieved mean Dice error ranging from 8.1 ±3.7 mm to 0.7 ±0.5 mm [147].

Improvement in TRUS image quality

In one study, deep learning was used to construct high-resolution TRUS images from low-resolution TRUS images [78]. Comparing network-generated images to original high-resolution images, the network was able to reconstruct images very close to the original image, with a mean absolute error (MAE) of 6.5 ±0.5 dB. Feleppa et al. employed ANN tissue-type classifiers to help improve image quality by using training data of cancerous and non-cancerous tissues, with an ROC curve area of 0.804 ±0.052, which was over 20% higher than that in non-AI classification systems [72].

AI in BT per disease site

Breast

Currently, only one study reported the use of AI in fractionated high-dose-rate brachytherapy (fractionated HDR-BT for breast cancer via electromagnetic tracking (EMT) data to determine if the source movement aligns with the treatment plan, and reported that deviations can be removed automatically [30]. AI has not been applied in intracavitary catheter reconstruction for APBI (accelerated partial breast irradiation) or in segmentation of OARs or treatment volumes for breast cancer BT.

Gynecology

Of the three main gynecological cancer HDR-BT techniques, i.e., intracavitary, interstitial, and hybrid (intracavitary-interstitial), the first two were explored in this study. Intracavitary paths and applicators were reconstructed with less than a 1 mm error. For interstitial needles, results were shown for metal needles only. At present, no studies have reported on the application of AI-based reconstruction methods for hybrid (intracavitary–interstitial) applicators in gynecological HDR-BT. Jung et al. [85, 86] investigated AI-based interstitial needle reconstruction, and obtained mean localization errors ranging from 0.63 to 2.0 mm, depending on algorithm and dataset. Despite variability across methods, all studies consistently demonstrated high needle-identification rates. Gynecological cancer was comparatively less represented in the OAR segmentation studies compared with prostate cancer. Segmentation of organs and ROI in gynecological cancer showed that within the same imaging modalities, ranges of DSCs were similar across cervical and prostate cancer patients. Eight studies provided segmentation of OARs and targets for gynecological cancer on MRI [124, 132, 134, 135, 142, 167, 169, 181]. For imaging in BT for gynecological cancer cases, the predominant focus was on artifact removal, with only two studies exploring image registration [78, 79].

Prostate

Prostate cancer represented majority of the studies in this review. AI has been shown to have the potential to improve both plastic and metal interstitial needle BT cases as well as PSI LDR-BT. In prostate HDR-BT, Andersén et al. [64] evaluated AI-based reconstruction of interstitial metal needles on ultrasound, and reported a mean localization error of 0.55 mm. In terms of segmentation, prostate cancer cases were presented on TRUS, MRI, and CT. When measured against expert manual contours, AI-generated segmentations achieved DSCs of approximately 0.88-0.93 on TRUS [88, 90, 91, 97, 98, 123, 166], 0.90-0.94 on MRI [102, 171], and 0.84-0.90 on CT [92], indicating slightly higher agreement for TRUS and MRI than for CT. Prostate cancer cases were primarily centered on image construction and registration techniques.

Discussion

Currently, the state of AI in BT literature reveals that both machine learning and deep learning are well tested pre-clinically and retrospectively, with confirmed accuracy and improvements in applicator and needle reconstruction, OAR and target value segmentation, treatment optimization, outcome prediction, and other BT procedures. AI-assisted applicator reconstruction minimizes errors and standardizes digitization, achieving sub millimeter agreement with expert reconstructions for intracavitary applicators. For interstitial needles in gynecologic HDR BT, the reported mean localization errors are 0.63-2.0 mm across datasets and algorithms [29, 39, 85, 86]. AI-assisted auto-segmentation and digitization demonstrate inference times of seconds in retrospective and pre-clinical settings. These results show potential to streamline planning steps, but do not equate to full clinical workflows, which still require validation, optimization, QA, and physician review. Early prospective studies [172, 213] suggest that efficiency gains are feasible, though routine deployment remains contingent on workflow validation and expert oversight. Similarly, AI can help to produce organ segmentations even on images with poor quality, and has been adopted to do so in many applications in and outside of BT [1]. Delineation is a key part of the overall success of BT, and AI has been shown to perform calculations consistently and in times of less than one minute.
Across gynecologic HDR BT, automated segmentation of organs at risk on CT/MRI is generally robust: typical DSCs for bladder and rectum are ~0.84-0.96, whereas for sigmoid and small bowel are lower (~0.60-0.80). High risk clinical target volume (HR CTV) DSCs are more variable, commonly ~0.79-0.89 with conventional CNN/nnU Net models, while recent prompt guided or ensemble approaches have reported higher values (~0.92-0.96) in single center evaluations; per structure inference is usually in the order of seconds [95, 111, 119, 124, 132, 134, 135, 137, 139, 169, 175, 176, 181, 185-187, 189]. These observations are consistent with the studies summarized in Table 2.
For prostate brachytherapy, auto segmentation performance is the strongest on TRUS, where prostate DSCs typically range ~0.88-0.97 across multi institution/device diverse datasets; MRI based pelvic structures cluster near ~0.86-0.96 (structure dependent), and CT based prostate segmentation commonly falls approximately ~0.82-0.90 [88, 90-92, 97-100, 102, 118, 121, 123, 125, 126, 166, 170]. Most frameworks report second level inference, in line with Table 2.
AI can accurately predict dose prescription and treatment plans, with the potential to decrease extraneous dose outside of the target while optimizing dose to the intended area, and adapt dwell times to catheter movements and internal anatomy [30]. In addition, AI based prognostic models have been used to estimate the risk of oncologic failure, defined here as local/regional recurrence or distant metastasis, enabling clinicians to use the data for personalized care. Follow-up frequency or BT technique can be adjusted to account for the patient’s potential for local or distant failure, and dosage can be prescribed based on the potential for toxicity [20].
Future studies focusing on the active use of AI in BT in clinical setting are necessary before clinical implementation. Research also report algorithm difficulties in hard cases, suggesting both the potential for AI algorithm refinement and the lasting necessity of physicians in non-trivial planning procedures and input of clinical parameters to AI algorithms. Studies indicate that AI may help with planning and validation procedures, but in terms of procedure performance, there still lies a heavy reliance on the clinician for the success of treatment. While AI will not eradicate the physician’s role, it is hopeful that technology will be used to standardize and strengthen the routine and repetitive tasks in BT to reduce the associated time and costs [20].
As the only comprehensive systematic review on the use of AI in BT adhered to PRISMA methodology, this study is subject to some limitations. Specifically, a publication bias could be present, as most studies were performed with previous patient cases, so cases with affirmative outcomes could have been specially selected for use. Also, the studies reported in this review varied widely in methods, population, disease site, treatment, and accuracy metrics reported, making comparison across different papers and a robust statistical analysis of results challenging. Determination of clinical applicability or performance predication is made difficult by the nature of this study being retrospective. Another important limitation is the lack of consensus on AI methodologies as best suited for clinical use. Numerous approaches, including radiomics-based models, CNN/nnU-Net architectures, and reinforcement learning frameworks, are being developed and optimized for different datasets and clinical goals, but no standardized strategy has yet been established [20, 22, 23]. This variation in methodology makes comparison across studies difficult, highlighting the need for consensus approaches and multi-institutional validation before clinical adoption. Most AI algorithm development or testing studies rely on retrospective data, and a retrospective approach can be advantageous in allowing models to be trained or improved by re-using the same patient data. Comparison of different AI methods is possible with this kind of study design, but limited patient characteristics or input datasets inhibit prediction of clinical performance. Only few studies have reported clinical use, including Nicolae et al.’s randomized trial in prostate brachytherapy [37, 38], Barten et al.’s first prospective experience in HDR prostate brachytherapy [213], and more recently Dickhoff et al. [209, 210], who used brachytherapy via artificially intelligent GOMEA-heuristic-based treatment planning (BRIGHT) framework, demonstrating that AI-based multi-objective planning can be feasibly applied in clinical HDR prostate and cervical brachytherapy. However, the limited number of clinical reports suggests either that AI has not yet been widely implemented in practice, or that clinical use is not consistently documented, both of which constrain patient benefit and hinder progress in the field. Across the 232 included studies, only 4 (1.7%) reported prospective or intra-operative application in patients: two in planning/optimization [38, 213], one in clinical needle reconstruction for prostate HDR [158], and one prospective evaluation of automated contouring for CT-based gynecologic brachytherapy [172]. All clinical implementations were single-center; no multi-center prospective validations were identified. In addition, several studies provided only qualitative comparisons with ground truth or conventional BT methods, further complicating clinical applicability prediction. This review is also limited by the low number of studies in clinical sites outside of gynecological or prostate cancer, and low number of independent testing cases or a lack of a testing/validation dataset. The current focus of AI applications on prostate and gynecological brachytherapy likely reflects both the global distribution of brachytherapy practice, where these sites account for the majority of treated patients, and the need for large, well-annotated datasets to train and validate AI models [20, 22, 23]. Smaller brachytherapy cohorts, such as head and neck or breast, are underrepresented in part because of data limitations as well as differences in institutional practice patterns and infrastructure [20, 22].

Conclusions

Artificial intelligence in BT has been applied across imaging modalities, such as MRI, CT, and TRUS. The reported performance includes applicator and needle reconstruction with localization errors of 0.7-16 mm, and OAR or target segmentation with DSC ranging from 0.60 to 0.97. Outcome prediction studies achieved AUCs of 0.60-0.90, while treatment optimization produced plans, which remained consistent with clinically delivered doses. This systematic review, the first to apply PRISMA methodology to AI in brachytherapy, demonstrates that AI consistently produces clinically acceptable and comparable treatment planning, with substantially reduced processing time compared with standard methods. Importantly, our findings also highlight persistent gaps, including the absence of large-scale prospective validations and a lack of consensus on which approaches are most suitable for clinical adoption. Addressing these gaps will be essential to transform AI from pre-clinical development into routine practice. Most available evidence remains retrospective. Adoption into routine practice will require prospective, multi institutional studies, which validate complete, end to end workflows (reconstruction, segmentation, planning, and QA) against pre specified clinical performance targets, report efficiency and safety metrics, and demonstrate generalizability across vendors and imaging modalities.

Funding

The research reported in this manuscript was supported by the University of Iowa ICRU (Iowa Center for Research by Undergraduates).

Disclosures

Approval of the Bioethics Committee was not required. The authors report no conflict of interest.
Supplementary material is available on the journal’s website.