Performance of an artificial intelligence foundation model for prostate radiotherapy segmentation

Performance of an artificial intelligence foundation model for prostate radiotherapy segmentation | medRxiv /* */ /* */ <!-- <!-- /*! * yepnope1.5.4 * (c) WTFPL, GPLv2 */ (function(a,b,c){function d(a){return"[object Function]"==o.call(a)}function e(a){return"string"==typeof a}function f(){}function g(a){return!a||"loaded"==a||"complete"==a||"uninitialized"==a}function h(){var a=p.shift();q=1,a?a.t?m(function(){("c"==a.t?B.injectCss:B.injectJs)(a.s,0,a.a,a.x,a.e,1)},0):(a(),h()):q=0}function i(a,c,d,e,f,i,j){function k(b){if(!o&&g(l.readyState)&&(u.r=o=1,!q&&h(),l.onload=l.onreadystatechange=null,b)){"img"!=a&&m(function(){t.removeChild(l)},50);for(var d in y[c])y[c].hasOwnProperty(d)&&y[c][d].onload()}}var j=j||B.errorTimeout,l=b.createElement(a),o=0,r=0,u={t:d,s:c,e:f,a:i,x:j};1===y[c]&&(r=1,y[c]=[]),"object"==a?l.data=c:(l.src=c,l.type=a),l.width=l.height="0",l.onerror=l.onload=l.onreadystatechange=function(){k.call(this,r)},p.splice(e,0,u),"img"!=a&&(r||2===y[c]?(t.insertBefore(l,s?null:n),m(k,j)):y[c].push(l))}function j(a,b,c,d,f){return q=0,b=b||"j",e(a)?i("c"==b?v:u,a,b,this.i++,c,d,f):(p.splice(this.i++,0,a),1==p.length&&h()),this}function k(){var a=B;return a.loader={load:j,i:0},a}var l=b.documentElement,m=a.setTimeout,n=b.getElementsByTagName("script")[0],o={}.toString,p=[],q=0,r="MozAppearance"in l.style,s=r&&!!b.createRange().compareNode,t=s?l:n.parentNode,l=a.opera&&"[object Opera]"==o.call(a.opera),l=!!b.attachEvent&&!l,u=r?"object":l?"script":"img",v=l?"script":u,w=Array.isArray||function(a){return"[object Array]"==o.call(a)},x=[],y={},z={timeout:function(a,b){return b.length&&(a.timeout=b[0]),a}},A,B;B=function(a){function b(a){var a=a.split("!"),b=x.length,c=a.pop(),d=a.length,c={url:c,origUrl:c,prefixes:a},e,f,g;for(f=0;f<d;f++)g=a[f].split("="),(e=z[g.shift()])&&(c=e(c,g));for(f=0;f<b;f++)c=x[f](c);return c}function g(a,e,f,g,h){var i=b(a),j=i.autoCallback;i.url.split(".").pop().split("?").shift(),i.bypass||(e&&(e=d(e)?e:e[a]||e[g]||e[a.split("/").pop().split("?")[0]]),i.instead?i.instead(a,e,f,g,h):(y[i.url]?i.noexec=!0:y[i.url]=1,f.load(i.url,i.forceCSS||!i.forceJS&&"css"==i.url.split(".").pop().split("?").shift()?"c":c,i.noexec,i.attrs,i.timeout),(d(e)||d(j))&&f.load(function(){k(),e&&e(i.origUrl,h,g),j&&j(i.origUrl,h,g),y[i.url]=2})))}function h(a,b){function c(a,c){if(a){if(e(a))c||(j=function(){var a=[].slice.call(arguments);k.apply(this,a),l()}),g(a,j,b,0,h);else if(Object(a)===a)for(n in m=function(){var b=0,c;for(c in a)a.hasOwnProperty(c)&&b++;return b}(),a)a.hasOwnProperty(n)&&(!c&&!--m&&(d(j)?j=function(){var a=[].slice.call(arguments);k.apply(this,a),l()}:j[n]=function(a){return function(){var b=[].slice.call(arguments);a&&a.apply(this,b),l()}}(k[n])),g(a[n],j,b,n,h))}else!c&&l()}var h=!!a.test,i=a.load||a.both,j=a.callback||f,k=j,l=a.complete||f,m,n;c(h?a.yep:a.nope,!!i),i&&c(i)}var i,j,l=this.yepnope.loader;if(e(a))g(a,0,l,0);else if(w(a))for(i=0;i (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0];var j=d.createElement(s);var dl=l!='dataLayer'?'&l='+l:'';j.src='//www.googletagmanager.com/gtm.js?id='+i+dl;j.type='text/javascript';j.async=true;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-P4HH5NV'); Skip to main content Home About Submit ALERTS / RSS Search for this keyword Advanced Search Performance of an artificial intelligence foundation model for prostate radiotherapy segmentation View ORCID Profile Matt Doucette , Chien-Yi Liao , Mu-Han Lin , Steve Jiang , Dan Nguyen , Daniel X. Yang doi: https://doi.org/10.1101/2025.02.23.25322754 Matt Doucette 1 Medical Artificial Intelligence and Automation (MAIA) Laboratory, University of Texas Southwestern Medical Center , Dallas, TX BS Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Matt Doucette Chien-Yi Liao 1 Medical Artificial Intelligence and Automation (MAIA) Laboratory, University of Texas Southwestern Medical Center , Dallas, TX PhD Find this author on Google Scholar Find this author on PubMed Search for this author on this site Mu-Han Lin 1 Medical Artificial Intelligence and Automation (MAIA) Laboratory, University of Texas Southwestern Medical Center , Dallas, TX 2 Department of Radiation Oncology, University of Texas Southwestern Medical Center , Dallas, TX PhD Find this author on Google Scholar Find this author on PubMed Search for this author on this site Steve Jiang 1 Medical Artificial Intelligence and Automation (MAIA) Laboratory, University of Texas Southwestern Medical Center , Dallas, TX 2 Department of Radiation Oncology, University of Texas Southwestern Medical Center , Dallas, TX PhD Find this author on Google Scholar Find this author on PubMed Search for this author on this site Dan Nguyen 1 Medical Artificial Intelligence and Automation (MAIA) Laboratory, University of Texas Southwestern Medical Center , Dallas, TX 2 Department of Radiation Oncology, University of Texas Southwestern Medical Center , Dallas, TX PhD Find this author on Google Scholar Find this author on PubMed Search for this author on this site For correspondence: daniel.yang{at}utsouthwestern.edu dan.nguyen{at}utsouthwestern.edu Daniel X. Yang 1 Medical Artificial Intelligence and Automation (MAIA) Laboratory, University of Texas Southwestern Medical Center , Dallas, TX 2 Department of Radiation Oncology, University of Texas Southwestern Medical Center , Dallas, TX MD Find this author on Google Scholar Find this author on PubMed Search for this author on this site For correspondence: daniel.yang{at}utsouthwestern.edu dan.nguyen{at}utsouthwestern.edu Abstract Full Text Info/History Metrics Supplementary material Data/Code Preview PDF Abstract Importance Artificial intelligence (AI) foundation models such as Segment Anything Model 2 (SAM 2) offer potential for semi-automated image segmentation with minimal fine-tuning, but their performance in specialized clinical tasks like radiation therapy planning are not well characterized. Objective To evaluate the performance of SAM 2 in segmenting pre-operative intact prostate and post-operative prostate fossa targets for prostate radiotherapy planning. Design, Setting, Participants Retrospective cohort study deploying and testing a foundation model for AI segmentation for prostate radiotherapy planning. CT simulation images and radiation plans were obtained from a single academic institution for patients undergoing prostate cancer treatment. Data analysis was performed from September 2024 to February 2025. Exposures AI segmentation with varying levels of human intervention, ranging from intervals of every 2nd to every 10th ground truth slice provided as input. Main Outcome and Measures Segmentation accuracy measured by Dice Similarity Coefficient (DSC) and Hausdorff Distance (HD) for intact and post-operative prostate target delineation. Results While SAM 2 outperformed interpolation in DSC and HD for both intact and post-operative prostate cancer patient cases, the AI segmentation accuracy was significantly better in the intact pre-operative patient cases where anatomic boundaries were better defined than post-operative patient cases. This is especially evident when sparse ground truth was provided simulating lower levels of human intervention. Conclusions and Relevance AI foundation models show promising application for specialized medical tasks such as prostate cancer radiotherapy segmentation with limited need for fine-tuning or retraining, although their clinical application will require further understanding of task-specific performance. Introduction In recent years, the development of robust machine learning models has been a resource-intensive endeavor, requiring significant computational power, large-scale annotated datasets, and substantial funding. 1 – 3 The emergence of foundation models has introduced new toolsets to this landscape, enabling researchers and practitioners to leverage pre-trained models and fine-tune them for specialized applications. 4 – 8 These foundation models, such as Meta’s Segment Anything Model 2 (SAM 2) and OpenAI’s ChatGPT, have demonstrated versatility across diverse domains, including medical imaging and natural language processing. Applying foundation models to specialized clinical and research tasks may offset the prohibitive costs associated with training such large-scale models from scratch. 9 – 12 Radiation therapy is a major treatment modality for localized prostate cancer. Accurate segmentation for radiation treatment planning is crucial, as it directly impacts the precision of the treatment delivered and patient outcomes. 13 , 14 While artificial intelligence (AI) tools have been introduced to aid segmentation, it remains a labor-intensive task requiring expert clinician input. AI auto-segmentation models can also be challenging to implement clinically due to data heterogeneity, lack of standardization across practice settings, variations in patient anatomy and tumor boundaries, hurdles in clinical workflow integration, amongst other factors. 15 , 16 Foundation models have been postulated to help address these challenges, potentially reducing time and costs without extensive model retraining efforts. 17 , 18 In this study, we aim to study an open-source foundation model designed to segment objects in various contexts with minimal user input. While initially developed for general-purpose segmentation tasks, the model’s adaptability has sparked interest in its application to medical imaging, including radiotherapy planning. 19 , 20 Specifically, we aim to characterize the segmentation performance of SAM 2 in pre-operative intact prostate cancer and post-operative prostate fossa cancer cases. We subsequently assess SAM 2’s capability to perform effectively with varying levels of human intervention by providing sparse ground truth annotations. By applying a foundation model to radiotherapy segmentation, this work highlights the potential for future foundation models to assist clinicians in specialized medical segmentation tasks, particularly in scenarios with limited resources or minimal expert input. The findings could have broader implications for the adoption of foundation models in medical imaging and clinical practice. Methods RT Planning and Data Acquisition This study included 282 prostate cancer patient cases from a single academic institution, divided into pre-operative (N=139) and post-operative (N=143) cohorts. The pre-operative patient cases include the clinical target volume (CTV) of the intact prostate, while the post-operative cases include the CTV of the prostate fossa following prostatectomy. Data was retrospectively collected and includes CT volumes obtained during simulation and treatment planning for definitive radiotherapy to the prostate and for adjuvant or salvage post-operative radiotherapy to the prostatic fossa. 21 All patients underwent CT imaging for simulation and treatment planning between 2021 and 2024. 22 The CTVs range from 10 to 39 slices across all the pre-operative and post-operative cases. The voxel size of the CT images is 1.17 × 1.17 × 2 mm 3 , ensuring sufficient resolution for delineating treatment targets and surrounding structures. Model and Input Raw DICOM data was parsed to create a folder with the JPG images and corresponding three-dimensional binary mask of the anatomical structure of interest (intact prostate or prostate fossa). The raw DICOM data was normalized using a sigmoid function transformation with α = 0.02 to create images that maximize the dynamic range near the average pixel value of the structure of interest. 23 There are 4 different-sized SAM 2 models called tiny, small, base plus, and large with sizes of 38.9, 46, 80.8, and 224.4 Mb respectively. The large model was used, which offers the best performance at the slowest inference rate. Experiment Design The SAM 2 model supports multiple input modalities, including positive and negative clicks, bounding boxes, and masks. 9 Input masks were used for this study. For each case, the physician-created CTV contour was used as the ground truth segmentation. A subset of these ground truth slices was provided as input to the SAM 2 model, simulating how a human clinician would typically segment every few slices in clinical practice and subsequently interpolate to create the CTV. In this study, the first and last ground truth slices were always included, with additional ground truth slices supplied at intervals to emulate different levels of human intervention. For example, in a case where the CTV contains 14 slices total, using an interval of every 4th slice would mean the 1st, 4th, 8th, 12th, and 14th slices were provided as input to the SAM 2 model. The remaining slices (2nd, 3rd, 5th, 6th, 7th, 9th, 10th, 11th, and 13th) were predicted by the model. CTV segmentation accuracy for intervals ranging from every 2nd to every 10th slice were evaluated for each patient case using the Dice Similarity Coefficient (DSC) and Hausdorff Distance (HD). To serve as a baseline comparison to SAM 2 performance, the missing slices were estimated using the interpolation of the given ground truth slices. The interpolated slices were calculated as the weighted linear combination of the Euclidean distance transforms of the last given and next given ground truth masks. 24 The average SAM 2 segmentation DSC scores across superior-inferior anatomical positions for both pre-operative (prostate) and post-operative (prostate fossa) cases were graphed and visualized to better understand how anatomic location influenced model performance. Statistical Analysis Performance comparisons were conducted at each level of ground truth provided (different levels of human intervention). Within each cohort, the performance of SAM 2 was compared to that of interpolation for intervals ranging from every 2nd to every 10th slice given. To assess statistically significant differences in the DSC and HD scores, the Mann-Whitney U test was used. 25 Statistical analysis was performed using the scipy.stats module in Python 3.11.3. Results Patient and Treatment Characteristics The pre-operative patient cohort had a median age of 65 years (IQR 56-74) and the post-operative cohort had a median age of 64 (IQR 54-76). Across both clinical settings, patients underwent simulation and treatment between 2021 to 2024. CTV volume was on average 40.22 cm 3 (STD: 16.89 cm 3 ) in the pre-operative setting compared to 146.51 cm 3 (STD: 45.78 cm 3 ) in the post-operative setting ( Table 1 ). View this table: View inline View popup Download powerpoint Table 1. Patient Characteristics. Performance Trends As the interval between provided ground truth slices increased, the performance of the SAM 2 model declined in both pre-operative and post-operative patient cases ( Figure 1 ). When every 2nd slice was provided, the model demonstrated strong performance with DSC scores averaging 0.956 and 0.952 and HD values of 3.666 and 4.020 in the pre-operative and post-operative settings respectively ( Table 2 , Supplement eTable 1). However, as fewer ground truth slices were given (e.g., every 10th slice), DSC scores decreased to 0.856 in pre-operative cases and 0.751 in post-operative cases, while HD values increased (worse performance) to 7.450 and 15.36, respectively. View this table: View inline View popup Download powerpoint Table 2. Dice similarity coefficient (DSC) for intact pre-operative and post-operative prostate cancer patient cases at varying levels of human segmentation. Download figure Open in new tab Figure 1. DSC scores and Hausdorff distances for foundation model performance for preoperative and post-operative prostate cancer radiotherapy target segmentation. Overall, when fewer ground truth slices were provided, the SAM 2 model output and interpolation exhibited steadily decreasing performance on average for both DSC and HD. With decreased ground truth slices provided simulating reduced human intervention, SAM 2 outperformed interpolation and demonstrated better performance in the pre-operative setting compared to the post-operative setting in both DSC and HD metrics ( Figure 1 ). DSC Performance As the level of human intervention decreased (i.e. fewer ground truth slices provided), the average DSC for the SAM 2 model output in the pre-operative setting decreased from 0.956 to 0.856 for intervals ranging from every 2nd slice to every 10th slice ( Table 2 ). Similarly, the DSC for interpolation decreased from 0.968 to 0.842 over the same interval. This trend highlights a consistent decline in AI segmentation accuracy as the spacing between ground truth slices widened. Notably, the pre-operative cases maintained higher DSC values than interpolation, with the performance gap widening at larger intervals. For example, at an interval of every 9th slice, the model achieved an average DSC of 0.866 (STD: 0.049), outperforming interpolation, which scored 0.851 (p < 0.001*). In the post-operative cohort, the SAM 2 model also demonstrated a decline in DSC as the interval between ground truth slices increased. The average DSC decreased from 0.952 to 0.751 for intervals ranging from every 2nd slice to every 10th slice ( Table 2 ). Interpolation in the post-operative setting showed a similar trend, with DSC decreasing from 0.949 to 0.705 over the same interval. The model consistently outperformed interpolation, also with the difference in performance widening at larger intervals. For example, at an interval of every 7th slice, the model achieved an average DSC of 0.846 (STD: 0.066), significantly outperforming interpolation, which scored 0.784 (p < 0.001*). The pre-operative cases consistently achieved higher DSC values than the post-operative cases, particularly at lower levels of human intervention (i.e., larger intervals between ground truth slices). While both cohorts showed an approximately linear decrease in DSC as the interval increased, the decline was more pronounced in the post-operative setting ( Table 2 , Figure 2 ). For example, at an interval of every 10th slice, the pre-operative DSC of 0.856 was significantly higher than the post-operative DSC of 0.751 (p < 0.001*). Additionally, the standard deviation in DSC scores had a greater increase in the post-operative cohort (from 0.015 to 0.098) compared to the pre-operative cohort (from 0.010 to 0.054). These trends suggest that the SAM 2 model is more robust in the pre-operative setting, particularly when human intervention is minimal. Download figure Open in new tab Download figure Open in new tab Figure 2. Violin plots of pre-operative and post-operative foundation model prostate cancer radiotherapy segmentation performance at selected intervals. HD Performance The HD performance mirrored the DSC trends, with segmentation accuracy declining as the interval between ground truth slices increased. In the pre-operative setting, the average HD for the SAM 2 model increased from 3.666 to 7.450 (worse performance) for intervals ranging from every 2nd to every 10th slice. Similarly, in the post-operative cohort, the model’s HD increased from 4.020 to 15.36. The pre-operative cases maintained lower HD values than the post-operative cases, with the performance gap widening at larger intervals. For example, at every 10th slice, the pre-operative HD of 7.450 was significantly lower than the post-operative HD of 15.36 (p < 0.001*). Variability also increased more sharply in the post-operative cohort, with standard deviations rising from 1.098 to 8.422 at intervals between every 2nd slice and every 10th slice, compared to 1.228 to 3.521 in the pre-operative cohort. These findings corroborate the DSC results, demonstrating that segmentation accuracy and consistency deteriorate more significantly in the post-operative setting as human intervention decreases. Anatomical Position Analysis For pre-operative cases, the superior part of the target segmentation had average DSC values ranging from 0.77 to 0.81 for the every 10th slice interval, 0.85 to 0.90 for every 5th slice, and 0.90 to 0.93 for every 2nd slice. The middle and inferior regions showed slightly higher segmentation accuracy at the every 10th slice interval, with DSC values ranging from 0.82 to 0.87, while the ranges of 0.85 to 0.91 for every 5th slice and 0.88 to 0.94 for every 2nd slice were in line with the superior region ranges ( Figure 3 ). The difference in performance at the every 10th slice interval is likely reflective of seminal vesicles being present at the superior part of the intact CTV, leading to variations in ground truth and model outputs. Download figure Open in new tab Figure 3. Foundation model pre-operative and post-operative prostate cancer radiotherapy segmentation DSC performance by relative anatomic position. For post-operative cases, superior regions show average DSC values ranging from 0.56 to 0.71 for the every 10th slice interval, 0.81 to 0.83 for every 5th slice, and 0.87 to 0.90 for every 2nd slice ( Figure 3 ). The middle and inferior regions achieve significantly better segmentation accuracy at the every 10th slice interval, with average DSC values ranging from 0.82 to 0.88, while the every 5th slice and every 2nd slice intervals ranged from 0.86 to 0.90 and 0.85 to 0.93 respectively. The more pronounced drop in target segmentation performance in the superior portion of the CTV within the post-operative cohort at the every 10th slice interval reflects a greater degree of variability and less well-defined boundaries than the pre-operative cases. Figure 3 shows representative image pairs illustrating the anatomical dependence of segmentation accuracy. In superior regions of the CTV, AI predictions (blue) frequently deviated from the ground truth (green), particularly in areas involving complex anatomical structures such as the seminal vesicles or post-surgical changes in the prostate fossa. By contrast, middle and inferior regions exhibit closer alignment between predicted and ground truth contours, reflecting the model’s improved performance in areas with more distinct and stable anatomical boundaries. Discussion Our study demonstrates that foundation models can achieve promising performance in prostate radiotherapy segmentation tasks with reduced human intervention and no domain-specific training. Unlike commercially available auto-segmentation tools designed specifically for this domain, SAM 2 offers a versatile, open-source solution that can be readily adapted to various medical imaging tasks. By testing the model on pre-operative and post-operative patient cases across a range of levels of human guidance, we provide a nuanced understanding of its capabilities across distinct clinical contexts. While SAM 2 does not outperform existing commercial prostate radiotherapy segmentation solutions, it maintained reasonable accuracy across varying levels of human guidance in pre-operative cases with clear anatomical boundaries. 26 Our results highlight a promising paradigm shift in medical image analysis. Rather than developing specialized models for each clinical task, foundation models offer a more scalable approach through fine-tuning. 27 , 28 This could significantly reduce development costs and accelerate the deployment of AI solutions across different medical imaging applications. 29 This corroborates previous work with SAM 2 demonstrating feasibility in medical imaging, achieving superior performance through targeted modifications to the base model. While SAM 2 scored an average DSC of 0.86 on various anatomical structure segmentation tasks, a model specialized for medical imaging achieved an average of 0.89 through the use of a “self-sorting memory bank” that gives higher weight to more informative embeddings of the surrounding slices rather than giving the highest weight to the embeddings of the closest slices. 19 These advancements demonstrate domain-specific adaptations can potentially improve foundation model performance without complete architectural redesign. Despite these promising results, there are limitations to this approach. As a foundation model trained on general video and image data, SAM 2 is likely optimized to recognize clearly defined boundaries in everyday settings. However, in medical image analysis and radiotherapy segmentation, well-defined boundaries may not be representative of the data used to train the model and are not always present. A previous study found that SAM struggled to segment organs of interest with less well-defined boundaries mirroring our finding of worse performance on post-operative prostate fossa segmentation than intact prostate segmentation, particularly at lower levels of human intervention. 30 These findings suggest that if foundation models like SAM 2 were to be adopted in medical image analysis and radiation segmentation, special attention should be given to scenarios that do not closely align with the data used to train the model. There are other limitations to our study. This is a single institution analysis, with patients treated across a multiple year time period with shifts in practice patterns and radiotherapy technology. Our data was retrospectively collected with inherent limitations of such analysis. Nevertheless, the heterogeneity of the data may reflect potential of the foundation model in real-world settings. Future research will be needed to develop efficient fine-tuning strategies that can quickly adapt such foundation models to specific clinical tasks while preserving their generalization capabilities. 31 – 33 This includes creating methods to incorporate domain expertise without requiring extensive retraining of the base model. Additionally, establishing robust frameworks for evaluating clinical reliability and safety will be crucial for healthcare adoption. 34 , 35 If future models are found to be safe and reliable, integration with existing clinical workflows presents another key challenge, requiring careful consideration of user interface design, computational requirements, and compatibility with current clinical software systems. 36 These developments will need to address both technical performance and practical implementation challenges to realize the full potential of foundation models in medical applications. Conclusions Foundation models represent a significant AI advancement with promising applications for specialized medical segmentation tasks. For prostate cancer radiotherapy planning, its strengths were more evident in the pre-operative intact setting, where clear radiographic boundaries facilitate accurate segmentation with reduced human intervention. Further research is needed to refine such models for radiotherapy task-specific applications such as addressing challenges of post-operative segmentation. These advancements could pave the way for broader adoption of foundation models in medical imaging and personalized treatment planning. Data Availability All data produced in the present study are available upon reasonable request to the authors Footnotes Updated supplemental material section and figure title. 1 Physician-created ground truth mask provided as input every nth slice 2 DSC reported as mean (std) 3 P-value calculated by Mann-Whitney U test comparing SAM 2 and Interpolation performance within corresponding cohort 4 P-value calculated by Mann-Whitney U test comparing SAM 2 performance across cohorts References 1. ↵ Rae JW , Borgeaud S , Cai T , et al. Scaling Language Models: Methods, Analysis & Insights from Training Gopher . Published online January 21 , 2022 . doi: 10.48550/arXiv.2112.11446 OpenUrl CrossRef 2. Thoppilan R , Freitas DD , Hall J , et al. LaMDA: Language Models for Dialog Applications . Published online February 10 , 2022 . doi: 10.48550/arXiv.2201.08239 OpenUrl CrossRef 3. ↵ Justus D , Brennan J , Bonner S , McGough AS . Predicting the Computational Cost of Deep Learning Models . Published online November 28 , 2018 . doi: 10.48550/arXiv.1811.11880 OpenUrl CrossRef 4. ↵ Brown TB , Mann B , Ryder N , et al. Language Models are Few-Shot Learners . Published online July 22 , 2020 . doi: 10.48550/arXiv.2005.14165 OpenUrl CrossRef 5. Bommasani R , Hudson DA , Adeli E , et al. On the Opportunities and Risks of Foundation Models . Published online July 12 , 2022 . doi: 10.48550/arXiv.2108.07258 OpenUrl CrossRef 6. Khan W , Leem S , See KB , Wong JK , Zhang S , Fang R. A Comprehensive Survey of Foundation Models in Medicine . Published online June 15 , 2024 . doi: 10.48550/arXiv.2406.10729 OpenUrl CrossRef 7. Esteva A , Chou K , Yeung S , et al. Deep learning-enabled medical computer vision . Npj Digit Med . 2021 ; 4 ( 1 ): 1 – 9 . doi: 10.1038/s41746-020-00376-2 OpenUrl CrossRef PubMed 8. ↵ Yu KH , Beam AL , Kohane IS . Artificial intelligence in healthcare . Nat Biomed Eng . 2018 ; 2 ( 10 ): 719 – 731 . doi: 10.1038/s41551-018-0305-z OpenUrl CrossRef PubMed 9. ↵ Ravi N , Gabeur V , Hu YT , et al. SAM 2: Segment Anything in Images and Videos . Published online October 28 , 2024 . doi: 10.48550/arXiv.2408.00714 OpenUrl CrossRef 10. Grattafiori A , Dubey A , Jauhri A , et al. The Llama 3 Herd of Models . Published online November 23 , 2024 . doi: 10.48550/arXiv.2407.21783 OpenUrl CrossRef 11. Xiao B , Wu H , Xu W , et al. Florence-2: Advancing a Unified Representation for a Variety of Vision Tasks . Published online November 10 , 2023 . doi: 10.48550/arXiv.2311.06242 OpenUrl CrossRef 12. ↵ Chia MA , Zhou Y , Keane PA . A New Foundation Model for Multimodal Ophthalmic Images: Advancing Disease Detection and Prediction . NEJM AI . 2024 ; 1 ( 12 ): AIe2401024 . doi: 10.1056/AIe2401024 OpenUrl CrossRef 13. ↵ Barber N , Ali A Wang T , Lewis B , Ruscetti M , et al. Prostate Cancer: Advances in Radiation Oncology, Molecular Biology, and Future Treatment Strategies . In: Barber N , Ali A , eds. Urologic Cancers . Exon Publications ; 2022 . Accessed January 5, 2025 . http://www.ncbi.nlm.nih.gov/books/NBK585981/ 14. ↵ Malicki J. The importance of accurate treatment planning, delivery, and dose verification . Rep Pract Oncol Radiother . 2012 ; 17 ( 2 ): 63 – 65 . doi: 10.1016/j.rpor.2012.02.001 OpenUrl CrossRef PubMed 15. ↵ Dal Pra A , Dirix P , Khoo V , et al. ESTRO ACROP guideline on prostate bed delineation for postoperative radiotherapy in prostate cancer . Clin Transl Radiat Oncol . 2023 ; 41 : 100638 . doi: 10.1016/j.ctro.2023.100638 OpenUrl CrossRef PubMed 16. ↵ Latorzeff I , Sargos P , Loos G , Supiot S , Guerif S , Carrie C. Delineation of the Prostate Bed: The “Invisible Target” Is Still an Issue? Front Oncol . 2017 ; 7 . doi: 10.3389/fonc.2017.00108 OpenUrl CrossRef 17. ↵ Zhang Y , Shen Z , Jiao R. Segment anything model for medical image segmentation: Current applications and future directions . Comput Biol Med . 2024 ; 171 : 108238 . doi: 10.1016/j.compbiomed.2024.108238 OpenUrl CrossRef PubMed 18. ↵ Lee HH , Gu Y , Zhao T , et al. Foundation Models for Biomedical Image Segmentation: A Survey . Published online January 15 , 2024 . doi: 10.48550/arXiv.2401.07654 OpenUrl CrossRef 19. ↵ Zhu J , Hamdi A , Qi Y , Jin Y , Wu J. Medical SAM 2: Segment medical images as video via Segment Anything Model 2 . Published online December 4 , 2024 . doi: 10.48550/arXiv.2408.00874 OpenUrl CrossRef 20. ↵ Dong H , Gu H , Chen Y , Yang J , Chen Y , Mazurowski MA . Segment anything model 2: an application to 2D and 3D medical images . Published online August 22 , 2024 . doi: 10.48550/arXiv.2408.00756 OpenUrl CrossRef 21. ↵ Bolla M , Van Poppel H , Tombal B , et al. Postoperative radiotherapy after radical prostatectomy for high-risk prostate cancer: long-term results of a randomised controlled trial (EORTC trial 22911) . The Lancet . 2012 ; 380 ( 9858 ): 2018 – 2027 . doi: 10.1016/S0140-6736(12)61253-7 OpenUrl CrossRef PubMed 22. ↵ Hall WA , Paulson E , Davis BJ , et al. NRG Oncology Updated International Consensus Atlas on Pelvic Lymph Node Volumes for Intact and Postoperative Prostate Cancer . Int J Radiat Oncol . 2021 ; 109 ( 1 ): 174 – 185 . doi: 10.1016/j.ijrobp.2020.08.034 OpenUrl CrossRef PubMed 23. ↵ Pathak A , Siddiqui S , Singh HK , Bhardwaj M , Singh EM . Enhancement of Low Contrast Medical Images Using Adaptive Sigmoid Function , Histogram Transformation and Swarm Optimization . 6 ( 3 ). 24. ↵ Strutz T. The Distance Transform and its Computation . ArXiv . Published online June 7 , 2021 . Accessed January 5, 2025 . https://www.semanticscholar.org/paper/The-Distance-Transform-and-its-Computation-Strutz/a3b9abfee2c3f165ad94ed0731919d8bdbdca08c 25. ↵ Nachar N. The Mann-Whitney U: A Test for Assessing Whether Two Independent Samples Come from the Same Distribution . Tutor Quant Methods Psychol . 2008 ; 4 ( 1 ): 13 – 20 . doi: 10.20982/tqmp.04.1.p013 OpenUrl CrossRef 26. ↵ Duan J , Tegtmeier RC , Vargas CE , et al. Achieving accurate prostate auto-segmentation on CT in the absence of MR imaging . Radiother Oncol . 2025 ; 202 : 110588 . doi: 10.1016/j.radonc.2024.110588 OpenUrl CrossRef PubMed 27. ↵ Litjens G , Kooi T , Bejnordi BE , et al. A survey on deep learning in medical image analysis . Med Image Anal . 2017 ; 42 : 60 – 88 . doi: 10.1016/j.media.2017.07.005 OpenUrl CrossRef PubMed 28. ↵ Azad B , Azad R , Eskandari S , et al. Foundational Models in Medical Imaging: A Comprehensive Survey and Future Vision . CoRR . Published online January 1 , 2023 . Accessed January 5, 2025 . https://openreview.net/forum?id=fcpc099PzZ 29. ↵ Yu X , Wang J , Hong QQ , Teku R , Wang SH , Zhang YD . Transfer learning for medical images analyses: A survey . Neurocomputing . 2022 ; 489 : 230 – 254 . doi: 10.1016/j.neucom.2021.08.159 OpenUrl CrossRef 30. ↵ Zhang L , Liu Z , Zhang L , et al. Segment Anything Model (SAM) for Radiation Oncology . Published online July 4 , 2023 . doi: 10.48550/arXiv.2306.11730 OpenUrl CrossRef 31. ↵ Celebi ME , Salekin MS , Kim H , et al. Silva-Rodríguez J , Dolz J , Ayed IB . Towards Foundation Models and Few-Shot Parameter-Efficient Fine-Tuning for Volumetric Organ Segmentation . In: Celebi ME , Salekin MS , Kim H , et al. , eds. Medical Image Computing and Computer Assisted Intervention – MICCAI 2023 Workshops . Springer Nature Switzerland ; 2023 : 213 – 224 . doi: 10.1007/978-3-031-47401-9_21 OpenUrl CrossRef 32. Dosovitskiy A , Beyer L , Kolesnikov A , et al. An Image is Worth 16×16 Words: Transformers for Image Recognition at Scale . Published online June 3 , 2021 . doi: 10.48550/arXiv.2010.11929 OpenUrl CrossRef 33. ↵ Curriculum Fine-tuning of Vision Foundation Model for Medical Image Classification Under Label Noise . Accessed January 5, 2025 . https://arxiv.org/html/2412.00150v1 34. ↵ Keni S. Evaluating artificial intelligence for medical imaging: a primer for clinicians . Br J Hosp Med . 2024 ; 85 ( 7 ): 1 – 13 . doi: 10.12968/hmed.2024.0312 OpenUrl CrossRef 35. ↵ Bundele V , Çal OA , Kargi B , et al. Evaluating Self-Supervised Learning in Medical Imaging: A Benchmark for Robustness, Generalizability, and Multi-Domain Impact . Published online December 26 , 2024 . doi: 10.48550/arXiv.2412.19124 OpenUrl CrossRef 36. ↵ Lotter W , Hassett MJ , Schultz N , Kehl KL , Van Allen EM , Cerami E. Artificial Intelligence (AI) in Oncology: Current Landscape, Challenges, and Future Directions . Cancer Discov . 2024 ; 14 ( 5 ): 711 – 726 . doi: 10.1158/2159-8290.CD-23-1199 OpenUrl CrossRef PubMed View the discussion thread. Back to top Previous Next Posted February 24, 2025. Download PDF Supplementary Material Data/Code Email Thank you for your interest in spreading the word about medRxiv. NOTE: Your email address is requested solely to identify you as the sender of this article. Your Email * Your Name * Send To * Enter multiple addresses on separate lines or separate them with commas. You are going to email the following Performance of an artificial intelligence foundation model for prostate radiotherapy segmentation Message Subject (Your Name) has forwarded a page to you from medRxiv Message Body (Your Name) thought you would like to see this page from the medRxiv website. Your Personal Message CAPTCHA This question is for testing whether or not you are a human visitor and to prevent automated spam submissions. Share Performance of an artificial intelligence foundation model for prostate radiotherapy segmentation Matt Doucette , Chien-Yi Liao , Mu-Han Lin , Steve Jiang , Dan Nguyen , Daniel X. Yang medRxiv 2025.02.23.25322754; doi: https://doi.org/10.1101/2025.02.23.25322754 Share This Article: Copy Citation Tools Performance of an artificial intelligence foundation model for prostate radiotherapy segmentation Matt Doucette , Chien-Yi Liao , Mu-Han Lin , Steve Jiang , Dan Nguyen , Daniel X. Yang medRxiv 2025.02.23.25322754; doi: https://doi.org/10.1101/2025.02.23.25322754 Citation Manager Formats BibTeX Bookends EasyBib EndNote (tagged) EndNote 8 (xml) Medlars Mendeley Papers RefWorks Tagged Ref Manager RIS Zotero Tweet Widget Facebook Like Google Plus One Subject Area Oncology Subject Areas All Articles Addiction Medicine (568) Allergy and Immunology (863) Anesthesia (300) Cardiovascular Medicine (4440) Dentistry and Oral Medicine (444) Dermatology (383) Emergency Medicine (608) Endocrinology (including Diabetes Mellitus and Metabolic Disease) (1510) Epidemiology (15229) Forensic Medicine (30) Gastroenterology (1126) Genetic and Genomic Medicine (6605) Geriatric Medicine (668) Health Economics (998) Health Informatics (4541) Health Policy (1369) Health Systems and Quality Improvement (1613) Hematology (543) HIV/AIDS (1265) Infectious Diseases (except HIV/AIDS) (15923) Intensive Care and Critical Care Medicine (1103) Medical Education (623) Medical Ethics (147) Nephrology (668) Neurology (6604) Nursing (346) Nutrition (998) Obstetrics and Gynecology (1145) Occupational and Environmental Health (957) Oncology (3334) Ophthalmology (974) Orthopedics (369) Otolaryngology (420) Pain Medicine (436) Palliative Medicine (130) Pathology (663) Pediatrics (1693) Pharmacology and Therapeutics (692) Primary Care Research (711) Psychiatry and Clinical Psychology (5448) Public and Global Health (9235) Radiology and Imaging (2199) Rehabilitation Medicine and Physical Therapy (1370) Respiratory Medicine (1196) Rheumatology (594) Sexual and Reproductive Health (712) Sports Medicine (530) Surgery (712) Toxicology (99) Transplantation (289) Urology (265) (function(){function c(){var b=a.contentDocument||a.contentWindow.document;if(b){var d=b.createElement('script');d.innerHTML="window.__CF$cv$params={r:'a0165f87988b0de1',t:'MTc3OTczMTMyOQ=='};var a=document.createElement('script');a.src='/cdn-cgi/challenge-platform/scripts/jsd/main.js';document.getElementsByTagName('head')[0].appendChild(a);";b.getElementsByTagName('head')[0].appendChild(d)}}if(document.body){var a=document.createElement('iframe');a.height=1;a.width=1;a.style.position='absolute';a.style.top=0;a.style.left=0;a.style.border='none';a.style.visibility='hidden';document.body.appendChild(a);if('loading'!==document.readyState)c();else if(window.addEventListener)document.addEventListener('DOMContentLoaded',c);else{var e=document.onreadystatechange||function(){};document.onreadystatechange=function(b){e(b);'loading'!==document.readyState&&(document.onreadystatechange=e,c())}}}})();