Murine Organ Auto-Contouring in Small-Animal Precision Irradiation: A Comprehensive Approach Integrating Deep Learning and Contrast Enhancement for Onboard CBCT

Murine Organ Auto-Contouring in Small-Animal Precision Irradiation: A Comprehensive Approach Integrating Deep Learning and Contrast Enhancement for Onboard CBCT | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (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],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Murine Organ Auto-Contouring in Small-Animal Precision Irradiation: A Comprehensive Approach Integrating Deep Learning and Contrast Enhancement for Onboard CBCT Ethan Cramer, Sophie Dobiasch, Ioannis I. Verginadis, Xinmin Liu, and 2 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6891084/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 9 You are reading this latest preprint version Abstract Modern preclinical irradiators have evolved to mimic their clinical Linac counterparts in terms of 360-degree beam delivery and on-board imaging capabilities with CBCT. The primary factor preventing widespread 3D conformal small-animal RT is the necessity of manually segmentation, as this task is time-consuming and impractical for large-scale studies. Although DL-based auto-contouring methods have been explored for preclinical irradiator CBCT, these methods have been limited to high-contrast, minimally anatomically complex structures. Thus, DL-based segmentation for low-contrast abdominal structures has yet to be addressed. In combining DL with iodine-based contrast-agent, precise full-body auto-contouring was achieved. A U-net-like architecture was trained to contour kidneys, spinal cord, stomach, liver, bowels, heart, lungs, and bones in small-animal irradiator CBCT mouse scans. Post contrast-enhancement, 41 mice were manually contoured, establishing ground truths. The model was trained with 26 mice, 2 for validation, and 15 for testing. Performance was evaluated using dice, precision, HD, and MSD. The proposed model predicted high-quality contours within a second, with the median for all organs reported: dice > 97%, precision > 98%, HD 95 < 2.15 mm, and MSD < 0.55 mm. The proposed combination of a DL and contrast-enhanced model is a viable method to vastly improve efficiency of small-animal IGRT. Biological sciences/Biophysics/Computational biophysics Biological sciences/Computational biology and bioinformatics/Machine learning Health sciences/Medical research/Pre clinical studies Biological sciences/Cancer/Cancer therapy/Radiotherapy Figures Figure 1 Figure 2 Figure 3 Figure 4 INTRODUCTION Animal models are the backbone of many areas of biomedical and preclinical research and the mouse is the most commonly used model organism for studying human disease 1 – 4 . Understanding and characterizing mouse models in detail is considered key to improving the reproducibility of preclinical results in human subjects 5 . For cancer-based radiobiology, research guiding clinical treatment often requires preliminary studies on the cellular or small animal scale 6 , 7 . However, to best model the actual radiobiological, radioimmunological, and toxicity characteristics of human therapies, preclinical techniques analogous to modern clinical strategies are needed. Recent studies have shown the possibility of mimicking the 3D conformal dose distributions of clinical intensity-modulated radiation therapy (IMRT) on small animal irradiations 8 – 11 . However, a primary obstacle preventing widespread use of such techniques is the need to segment anatomic murine structures for treatment planning optimization. Depending on the site, simple single organ contouring (brain, heart, etc.) can take approximately 20 minutes, whereas, for highly complex cases involving many organs or skeleton contouring for total bone marrow irradiation can take many hours 12 , 13 . Such times are not suitable for large scale murine radiation biology studies which require many mouse deliveries per day to generate statistical significance of results. In addition, manual delineation of organs on each slice of a volumetric computed tomography (CT) scan requires not only great attention to detail, but a high-level of mouse anatomy expertise. Furthermore, there is subjectivity involved in manual segmentation due to individual bias. Recent studies have investigated the use of automated segmentation approaches using machine learning. For instance, Lappas et al. 14 used a U-net architecture with three levels of encoding to extract features and three levels of decoding to automatically predict organ segmentation. This 3D U-net deep learning (DL) architecture model successfully predicted high-quality contours for the thorax and head area in mice cone-beam CT (CBCT) scans taken using an image-guided (IG) small animal irradiator. Van der Heyden et al. used a 3D U-net DL architecture model to determine skeletal muscle mass in small animal irradiator CBCT scans 15 . However, the lower quality of preclinical irradiator CBCT has so far limited AI to only areas that are high contrast—minimally anatomically complex—structures located in the head (eyes, brain) and thorax (heart, lungs) regions. To allow DL-based auto-contouring of abdominal and other more ambiguously shaped organs has required the use of a dedicated small animal µ-CT scanner 16 . This greatly limits the use of AI organ segmentation for many preclinical radiation therapy (RT) studies as the mouse must be physically transferred from the µ-CT to the small animal irradiator leading to potential position setup errors. To address this issue, and to further expand the utility of AI contouring in preclinical RT, we present a method of using a machine learning model with CBCT irradiators by combining a DL U-net-like architecture with iodine-based contrast-enhancement to efficiently and accurately contour unambiguous structures in the thorax and bones, as well as mostly ambiguous, abdominal mouse structures. To the best of our knowledge, this is the first method to perform high quality DL based auto contouring of abdominal structures in lower quality CBCT scans and has the potential to vastly improve the workflow efficiency of small animal IGRT preclinical trials. RESULTS As means of optimizing detection of the OARs, injection of iodine contrast agent was used immediately before imaging to increase soft tissue contrast. The intestinal loops became partly air-filled, the kidneys appeared in the retroperitoneum, and the spine and all other skeleton structures appeared hyperdense. Figure 1 displays a comparison of CBCT scans of a native, non-contrast-enhanced mouse, and a contrast-enhanced mouse. As shown in the axial (top left), coronal (bottom), and sagittal (top right) slices, there was a clear advantage in using iodine-based contrast-enhancements. The DL model was trained on an optimal dataset size of 26 mice. Of the 26 mice used, one mouse was used for testing during training, and another one used for validation for the Adam optimizer during training. The model was trained on 30 epochs and tested on a test set of 15 mice. Although more training mice were available, it was found that in general, going beyond a 26-mouse dataset resulted in only marginal dice score improvements and caused increased noise. Figure 2 shows representatively a side-by-side comparison of 3D projections of human segmented ground truth file (left) and AI automatically segmented file (right). As shown, the two look remarkably similar, which has extremely positive implications for the efficacy of the DL model. In general, for the best contours, the DL generated segmentations for each structure almost perfectly match the human ground truth contours, while in the worst case there were only a few regions of disagreement. Figure 3 shows example slices of the coronal and sagittal views for one mouse from the test dataset, with the ground truth human contours outlined in bolded yellow. The DL generated contours are shown highlighting each predicted structure. Areas of disagreement came in the form of either a pixel characterized as not part of a specific structure, when the ground truth file had that pixel as a part of that structure, or a pixel is characterized as part of a specific structure by the DL model, while that pixel is not part of the organ in the ground truth file. Figure 4 displays the dice score boxplots results for each structure, as well as the median of all structures, for all 15 mice used for evaluation. Median dice scores for each organ are reported: left kidney = 99.36%, right kidney = 99.36%, spinal cord = 75.79%, stomach = 96.52%, liver = 95.92%, bowel = 95.65%, heart = 98.75%, left lung = 98.23%, right lung = 98.24%, bones = 94.77%, and average for all organs = 97.38%. With exception of the spinal cord, the median dice score was higher than 85%, meaning overall performance of the DL automatic segmentation model was very high 14 . The kidneys showed consistent accuracy of over 99%, whereas the stomach, liver, and bowel, were slightly less accurate, but still consistently over 95%. Segmented structures in the thorax region, as expected, performed very high, consistently over 98%. Nevertheless, where the model lacked in accuracy is longitudinally, in the spinal cord for instance, where there was high variance, ranging from a dice score of 40–90%. Since the spinal cord appears in so few slices, any inaccuracies are amplified. Thus, the worst contours were still accurately contoured with a dice score of 75%, which is still highly accurate for a small longitudinal structure. For all ten annotated structures–left kidney, right kidney, spinal cord, stomach, liver, bowels, heart, left lung, right lung, and bones–each evaluation metric is presented in Table 1 . Similar trends in the results can be observed in the Jaccard and precision scores. According to the Jaccard scores, the least accurate segmentations came in the spinal cord. The model produced extremely good precision scores, with the least precise predicted volume segmentation came in the spinal cord, with a precision score of about 80%. The false negative rate (FNR) or miss rate of the DL-enabled mouse automatic segmentation model for each organ is shown in the fifth column of Table 1 . Overall, there was a median miss rate of about 4%, which shows high accuracy of the DL automatic contours. The HD50 median for all organs was 0.00 mm, while the HD95 averaged for all organs was about 2.12 mm. The median of the mean surface distance for all organs was 0.55 mm. The median standard deviation for all organs was 0.83 mm. Overall, distance measurements showed strong agreement with quality metrics, showing high accuracy of DL automatic contours. According to the surface distance measurements, consistently, the least accurately contoured organ was the bowel. This is because the bowel is the largest area to contour, as the bowel is spread along the whole abdomen. The spinal cord, however, had the largest miss rate at 21%, as shown in Table 1 . Table 1 Segmentation accuracy evaluated using the dice, Jaccard score, precision score, FNR, 50th percentile and 95th percentile HDs, mean, median, and standard deviation of surface distances (SD). The values displayed are the medians for each organ over the while 15-mice test dataset. Median of Metric Organ Dice Jaccard Precision FNR HD 50 HD 95 Mean SD Median SD Std. SD Left Kidney 99.36 98.73 98.67 0.01 0.00 1.00 0.18 0.00 0.48 Right Kidney 99.36 98.73 99.62 0.01 0.00 1.00 0.16 0.00 0.48 Spinal Cord 75.79 61.02 79.46 0.21 1.00 3.90 1.40 1.00 1.32 Stomach 96.52 93.28 98.70 0.06 0.00 2.83 0.77 0.00 0.99 Liver 95.92 92.16 97.06 0.04 1.00 4.47 1.31 1.00 1.69 Bowel 95.65 91.67 95.79 0.05 2.00 9.05 2.80 2.00 2.99 Heart 98.75 97.53 99.73 0.02 0.00 2.00 0.39 0.00 0.67 Left Lung 98.23 96.53 99.03 0.03 0.00 2.00 0.23 0.00 0.55 Right Lung 98.24 96.54 99.44 0.02 0.00 1.00 0.19 0.00 0.48 Bones 94.77 90.05 96.02 0.07 0.00 2.24 0.71 0.00 1.59 All 97.38 94.91 98.69 0.04 0.00 2.12 0.55 0.00 0.83 DISCUSSION As preclinical irradiators continue to advance to closely mimic their clinical Linac counterparts, there is a requirement for high-quality automated 3D contours for large-scale 3D conformal RT small animal studies. Not only does automatic organ segmentation vastly increase efficiency of these preclinical murine studies, but it also eliminates human bias and fatigue which can lead to errors. Previous DL-enabled automated segmentation for small animal methods have either required dedicated micro-CT scanners for whole body murine imaging or have been limited to only high-contrast, non-anatomically complex thoracic and head regions when using small animal irradiator CBCT 14 , 16 – 19 . This work demonstrates that despite the image resolution and contrast limitations of small animal irradiator CBCT, the combination of iodine-based contrast-enhancement and U-net-like DL architecture can allow high-quality contours of more difficult to see abdominal organs without the need for a micro-CT. Additionally, unlike previous DL contouring models, which required large training datasets of 75 + mice, this work demonstrated the ability to create a high accuracy full-body segmentation models with a training dataset of only 26 mice 14 . Manual structure segmentation can take upwards of 10 minutes per organ, making the total segmentation time in this study over 75 hours 14 . Conversely, the DL model can accurately segment all eleven mouse organs and skeletal structures in under a second, vastly reducing contouring time and increasing efficiency of the image-guided precision RT workflow. The U-net-like DL architecture proved as a method to accurately, automatically contour murine abdominal and thorax organs, as well as skeletal and heterogenous tumor structures, right and left kidney’s, spinal cord, stomach, liver, bowel, heart, left and right lungs, and bones in under a second. Upon training on a 26-mouse dataset of manual expert and semi-autonomous segmentation, performance of the test mice for most of the organs showed high-quality automatic contours, consistently way over a dice score of 85% and HD95 of less than 2.15 mm. A dataset size of 26 mice was chosen based on preliminary results, showing a saturation effect in performance when including additional mice in the training dataset. There are sources for uncertainty in this DL-enabled automatic contouring architecture that influences the quality of the generated segmentations 14 . Most notably, there is uncertainty that can arise in the input. Manual, human contours may not represent a true ground truth segmentation due to subjective interpretation and bias, which must be taken into account 16 . This subjective interpretation effect is quantified using inter-observer variability (IOV) studies, which potentially causes regions of disagreement. Additionally, the observer performing the manual segmentation of organs may suffer from fatigue and time pressure, causing them to miss contours over time, causing incompleteness and disagreements. Limitations of using DL for automated delineations of small animal structures mainly stem from CBCT artifacts causing the model to predict pixels that are a part of the CT bench as part of a structure. This effect, however, can be quickly and easily resolved in most treatment planning software by simply deleting the contours that are clearly out of the image of the small animal. The model lacked accuracy mainly in areas often considered ambiguous for manual segmentation specialists. The metrics presented are extremely sensitive and consider all differences between the DL automatic contours and human ground truth segmentation contours. Thus, relatively poorer scores can be observed longitudinally in the spinal cord, as found in previous studies. Nevertheless, this model is applicable to other contrast agent enhanced CBCTs. In this work, a deep learning (DL) model was successfully trained to accurately predict contours for low-contrast abdominal structures in preclinical irradiator CBCT for the first time. The hybrid approach combining DL with an iodine contrast agent is efficient and enhances setup accuracy since the mouse does not need to be moved from the irradiator for dedicated µ-CT low-contrast tissue imaging. By significantly reducing the manual contouring burden, the proposed method makes advanced 3D conformal preclinical radiotherapy (RT) techniques, such as IMRT, more practical for large-scale mouse studies. METHODS Dataset and Image Conditioning The dataset consisted of 41 6-week-old immunosuppressed CD-1 nude mice partly taken from a previously completed pancreatic cancer study 20 . In this study, all animal procedures were carried out in accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. Moreover, the protocol was officially approved and authorized by German law for the Care and Use of Laboratory Animals and the regional government of Upper Bavaria, Germany (ref. 55.2-1-54-2532-217-2025 from April 13, 2016). All surgical procedures were performed under anesthesia, and all efforts were made to minimize suffering 20 . Additionally, experimental descriptions in this manuscript comply with ARRIVE guidelines. Treatment planning and irradiation of the mice were performed when the GTV reached 90 mm 3 , which took about eight weeks post injection of Panc-1 (CRL-1469) and three weeks for MiaPaCa-2 (CRL-1420). Anesthesia was used during the entirety of the imaging process to immobilize the mice. Each mouse inhaled isoflurane anesthesia at a concentration of 1.5% with a 6% volume of oxygen as a carrier of the gas. Imaging was done using the onboard cone-beam computed tomography (CBCT) system of a small animal irradiator (SARRP, Xstrahl, GA) with the operating x-ray source at a voltage of 60 kV and current of 0.8 mA. In addition, each mouse was injected with 5 mL per 1 kg mice body weight of iodine-containing contrast agent (Imeron, Bracco, Italy) intravenously (iv) through the lateral tail vessel to improve soft tissue contrast and optimize the definition of the OARs. CBCT imaging was taken immediately after iv injection using 1440 projections and a voxel size of 0.115 x 0.115 x 0.115 mm 3 . Mice were ethnized one-hour post-irradiation treatment by cervical dislocation 20 . Treatment planning, including manual segmentation of mice structures, was done using the preclinical treatment planning software ‘MuriPlan’. The bones, along with six OARs, including both kidneys, bowel, stomach, spinal cord, and liver, were manually delineated slice-by-slice of the CBCT image by a segmentation specialist. Contouring of the heart, lungs, and bones was done semi-automatically, using a previously built model to predict contours, then carefully edited slice-by-slice to fix any potential errors. When inputting a native CBCT scan into the DL-based automatic segmentation model, performance is very poor, and limits quality contours to only the thorax region, as well as the bone structure. For this reason, the use of contrast-enhancement immediately before CBCT imaging was imperative to building an accurate abdominal organ automatic segmentation model, with this U-net-like DL architecture. As a means of standardizing the 3D Nifti encoded CBCT images before model training, a script was first used to check for completeness of the available scans, verifying them for consistency in that both the CBCT scan and annotated 3D Nifti volumes are of the same size. After this each volume is turned into coronal slices to allow for faster processing by slicing the 3D Nifti volume into 2D tiff images. The signal intensity was then normalized by subtracting the mean and dividing by the standard deviation. The images were then resampled to a resolution of 240 µm/vx, even though the DL U-Net-like architecture works over a broad range of 120–1120 µm/vx 16 . Devices and Python Libraries The algorithm, developed in python, was implemented using these open-source libraries: PyTorch, SciPy, NiBabel, Nrrd, Numpy, and Matplotlib. The deep learning framework pipeline uses PyTorch. Additionally, a NVIDIA Titan Xp GPU was used, but not necessary as the program can be run using a CPU, at the cost of time. The pipeline, proposed by Schoppe et al. 16 , consisted of three parts, a preprocessor, deep learning backbone, and a postprocessor. Deep Learning Model The DL backbone follows a U-Net-like architecture, consisting of an encoding and decoding path connected with skip connections. The varying convolutional sizes help to extract features at various resolutions and without skip connections, upsampling back to the original resolution would be very difficult. Moreover, the use of skip connections resolved the vanishing gradient problem by providing an uninterrupted gradient flow through all layers 21 . Encoding levels consisted of two convolutions (kernel size: 3, padding: 1, stride: 1), batch normalization, a rectifying linear activation function (ReLu), and a max-pooling operation (kernel size: 2, stride: 2). Additionally, each encoding level had twice the number of featured channels compared to its previous level, starting at 32 feature channels. The decoding units consisted of three convolutions (kernel size: 3, padding: 1, stride: 1) and received concatenation of upsampled input from the previous level, including input from the skip connection from the corresponding encoding level (consisting of the same number of feature channels). The final convolution maps the 32 feature channels to the number of structures predicted. The last pass was through a sigmoid function which creates a volumetric probability map for each structure. The DL model was trained with a soft-Dice loss function and used the Adam optimizer 16 . A soft-Dice loss function was used to directly predict probabilities of target structure position, as opposed to converting them into a binary mask. The soft-Dice loss function is a differential reformation of the commonly used Dice Similarity Coefficient (DSC) evaluation metric. If \(\:A\) represents pixels from the human segmentation ground truth file, and \(\:B\) consists of pixels from the DL automatically segmented file, the DSC can be calculated as two times the intersection of A and B divided by the sum of the two sets: \(\:DSC\:=\:\frac{2|A\cap\:B|}{\left|A\right|\:+\:\left|B\right|}\:\) 2 2 . Letting \(\:2\left|A\cap\:B\right|={\sum\:}_{i}{a}_{i}*{b}_{i}=\:\) and \(\:\left|A\right|=\:{\sum\:}_{i}{a}_{i}*a=\:\) , a differential soft-Dice loss function can be determined as: $$\:{L}_{Dice}=1-\:\frac{\:2\:}{\:+\:}\:$$ The Adam optimizer was used as a method for computing individual adaptive learning rates for different parameters from estimates of first and second moments of the gradients 23 . Furthermore, a nested k-fold cross-validation procedure was used to split the dataset into a training set (for model weight optimization), a validation set (for hyper-parameter optimization), and a test set (for evaluation) 23 . The DL model was trained for 30 epochs, with an initial learning rate of 10 − 3 , on the training set, with the learning rate being gradually reduced as validation performance was unchanged over five epochs. The postprocessor utilizes ensemble dash voting, where the neural network uses information from many independently trained networks and merges them to make the most accurate prediction possible 16 . Finally, the volumetric probability map from the DL backbone is used for volumetric reconstruction to create a predicted segmentation file in the postprocessor. Evaluation Metrics Both quality and distance metrics are presented for performance evaluation. Quality metrics, such as the Dice Similarity Score, or just dice score, is an evaluation of the qualitative performance of the algorithm with respect to the human segmentation for each structure. Quality metrics use voxel classification of the DL-predicted contouring of a CBCT scan for each structure, denoting each voxel as either ‘True Positive’ (TP), ‘False Positive’ (FP), or ‘False Negative’ (FN). Voxels are classified as TP if it was correctly predicted compared to the ground truth file, FP if the voxel was incorrectly predicted as part of an organ, and FN if the voxel was predicted to be a non-organ but is a part of an organ in the ground truth file. With this classification of the pixels, additional metrics were utilized to assess accuracy of automatically segmented files against human segmentation, such as those presented in the Table 2 . The quality metrics are based on two sets of pixels, A and B, which consists of pixels from the DL-predicted segmentation and the human ground truth segmentation, respectively. The dice index gives the percent overlap, or reproducibility–how well the segmentation and ground truth files match–for each structure. The Jaccard index is the intersection of A and B over the union of A and B. The precision gives the rate at which pixels are correctly classified as part of the correct structure. FP rate and FN rate give the rate at which false positives classifications are found and false negative predictions are found, respectively 24 . Table 2 Quality metrics used for performance evaluation of the DL model (left) and distance metrics used for evaluation of DL model (right) Quality Metrics Formula 24 Distance Metrics Formula 26 \(\:Dice\:Score\:=\:\frac{2TP}{2TP\:+\:FP\:+\:FN\:+\:ϵ}\times\:100\%\) \(\:h(A,B)\:=\:{max}_{a\in\:A}{min}_{b\in\:B}|A-B|\) \(\:HD(A,B)\:=\:max\left(h\right(A,B),\:h(B,A)\) \(\:Jaccard\:Index\:=\:\frac{TP}{TP\:+\:FP\:+\:FN\:+\:ϵ}\) \(\:{HD}_{p}(A,B)\:=\:percentile\left[max\right(h(A,B),\:h(B,A)\left)\right]\) \(\:Precision\:=\:\frac{TP}{TP\:+\:FP\:+\:ϵ}\) \(\:Mean\:SD\:=\:mean\left(h\right(A,B),\:h(B,A\left)\right)\) \(\:FP\:rate\:=\:\frac{FP}{FP\:+\:TP\:+\:ϵ}\) \(\:Median\:SD\:=\:median\left(h\right(A,B),\:h(B,A\left)\right)\) \(\:FN\:rate\:=\:\frac{FN}{FN\:+\:TP\:+\:ϵ}\) \(\:std\:SD\:=\:std\left(h\right(A,B),\:h(B,A\left)\right)\) \(\:ϵ\:=\:0.0001\:\) for numerical stability SD = Surface Distance, HD = Hausdorff Distance Spatial distance metrics were used to quantitatively evaluate the accuracy of the predicted contours, taking into consideration the spatial position of voxels. The directed Hausdorff (h(A,B)), as shown in the Table 2 , where |a-b| is the difference in the Euclidean distance between pixels in their respective pixel sets from predicted contouring of a structure and ground truth contours. The Hausdorff Distance (HD) is the distance between crisp volumes (two finite pixel sets, A and B). Since noise and outlier pixels are very common in medical segmentation, it is not recommended to use the maximum HD. Instead, the quantile method proposed by Huttenlocher et al. 25 is one way to handle outliers, where HD is defined as some percentile of the maximum distance 24 . For this reason, the 50th (HD50) and 95th (HD95) percentile HDs are presented. The mean surface distance gives the mean distances between predicted and human segmentations for each structure. The median and standard deviation of these distances are also reported. Declarations Competing interests No conflict exists for any author. Author Contribution E.C. fabricated DL segmentation model, conducted experiments, and drafted manuscript. S.D. provided mouse CBCT data, administered contrast agent, contributed to manual segmentation, and results analysis. I.V. contributed to manual segmentation efforts and analyzed results. X.L. and S.C. analyzed results. R.W. conceived of experiment and analyzed results. All authors reviewed the manuscript. Data Availability Data is available from the corresponding author upon reasonable request. References Rosenthal, N. & Brown, S. The mouse ascending: perspectives for human-disease models. Nat. Cell. Biol. 9 , 993–999 (2007). Osuchowski, M. F. et al. Abandon the mouse research ship? Not just yet! Shock Augusta Ga. 41 , 463–475 (2014). Schoppe, O. et al. Deep learning-enabled multi-organ segmentation in whole-body mouse scans. Nat. Commun. 11 , 5626 (2020). Hackam, D. G. & Redelmeier, D. A. Translation of research evidence from animals to humans. JAMA 296 , 1731–1732 (2006). Perrin, S. Preclinical research: Make mouse studies work. Nature 507 , 423–425 (2014). Kahn, J., Tofilon, P. J. & Camphausen, K. Preclinical models in radiation oncology. Radiat. Oncol. Lond. Engl. 7 , 223 (2012). Van Dyke, T. & Jacks, T. Cancer modeling in the modern era: progress and challenges. Cell 108 , 135–144 (2002). Woods, K., Neph, R., Nguyen, D. & Sheng, K. A sparse orthogonal collimator for small animal intensity-modulated radiation therapy. Part II: hardware development and commissioning. Med. Phys. 46 , 5733–5747 (2019). Jiang, L., Lyu, Q., Abdelhamid, A. M. H., Hui, S. & Sheng, K. An efficient rectangular optimization method for sparse orthogonal collimator based small animal irradiation. Phys. Med. Biol. 10.1088/1361-6560/ac910b (2022). Redler, G. et al. Small Animal IMRT Using 3D-Printed Compensators. Int. J. Radiat. Oncol. Biol. Phys. 110 , 551–565 (2021). Liu, X. et al. Improving the efficiency of small animal 3D-printed compensator IMRT with beamlet intensity total variation regularization. Med. Phys. 49 , 5400–5408 (2022). Wang, H., Stout, D. B. & Chatziioannou, A. F. Estimation of mouse organ locations through registration of a statistical mouse atlas with micro-CT images. IEEE Trans. Med. Imaging . 31 , 88–102 (2012). van der Heyden, B. et al. Automatic multiatlas based organ at risk segmentation in mice. Br. J. Radiol. 92 , 20180364 (2019). Lappas, G. et al. Automatic contouring of normal tissues with deep learning for preclinical radiation studies. Phys. Med. Biol. 67 , 044001 (2022). Van Der Heyden, B. et al. Automated CT-derived skeletal muscle mass determination in lower hind limbs of mice using a 3D U-Net deep learning network. J. Appl. Physiol. 128 , 42–49 (2020). Schoppe, O. et al. Deep learning-enabled multi-organ segmentation in whole-body mouse scans. Nat. Commun. 11 , 5626 (2020). Lappas, G. et al. Inter-observer variability of organ contouring for preclinical studies with cone beam Computed Tomography imaging. Phys. Imaging Radiat. Oncol. 21 , 11–17 (2022). van de Worp, W. R. P. H. et al. Deep Learning Based Automated Orthotopic Lung Tumor Segmentation in Whole-Body Mouse CT-Scans. Cancers 13 , 4585 (2021). Rosenhain, S. et al. A preclinical micro-computed tomography database including 3D whole body organ segmentations. Sci. Data . 5 , 180294 (2018). Dobiasch, S. et al. Histopathological Tumor and Normal Tissue Responses after 3D-Planned Arc Radiotherapy in an Orthotopic Xenograft Mouse Model of Human Pancreatic Cancer. Cancers 13 , 5656 (2021). Neupane, B., Aryal, J. & Rajabifard, A. Dual skip connections in U-Net, ResUnet and U-Net3 + for remote extraction of buildings. Preprint at (2023). http://arxiv.org/abs/2303.09064 Galdran, A., Carneiro, G. & Ballester, M. Á. G. On the Optimal Combination of Cross-Entropy and Soft Dice Losses for Lesion Segmentation with Out-of-Distribution Robustness. Preprint at (2022). http://arxiv.org/abs/2209.06078 Kingma, D. P., Ba, J. & Adam A Method for Stochastic Optimization. Preprint at (2017). http://arxiv.org/abs/1412.6980 Taha, A. A. & Hanbury, A. Metrics for evaluating 3D medical image segmentation: analysis, selection, and tool. BMC Med. Imaging . 15 , 29 (2015). Huttenlocher, D. P., Klanderman, G. A. & Rucklidge, W. J. Comparing images using the Hausdorff distance. IEEE Trans. Pattern Anal. Mach. Intell. 15 , 850–863 (1993). Müller, D. et al. MISeval: a Metric Library for Medical Image Segmentation Evaluation. Preprint at (2022). http://arxiv.org/abs/2201.09395 Additional Declarations No competing interests reported. Cite Share Download PDF Status: Under Review Version 1 posted Reviews received at journal 18 Oct, 2025 Reviews received at journal 08 Oct, 2025 Reviewers agreed at journal 06 Oct, 2025 Reviewers agreed at journal 29 Sep, 2025 Reviewers invited by journal 18 Sep, 2025 Editor assigned by journal 17 Sep, 2025 Editor invited by journal 17 Jun, 2025 Submission checks completed at journal 17 Jun, 2025 First submitted to journal 13 Jun, 2025 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-6891084","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":548933977,"identity":"410e2c0b-0891-4320-b7b1-88342acf4e4a","order_by":0,"name":"Ethan Cramer","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA+ElEQVRIiWNgGAWjYBADOSA2gDDZQQQbYS3GEC0JQIqZSC2JDURr4Wc/+/DTjYp76Wvbm7c9+PiDIbG/mfkBw4eywzi1SPakG0vnnCnO3XbmWLnhjASGxBmH2QwYZ5zDrcXgBhuDdG5bQu62Gzlm0jxALRuYGQyYedtwa7G/wcb8O/dfQrrZ/Tdm0n/AWtg/MP/Fo8VAgo1NOrchIcHsBo+ZNANYC48BMyMeLRJn0tisc44lGG47k1Ym2ZMmYTzjME/BwZ5z6Ti18LcfY76dU5Mgb3b88DaJHzY2sv3t7Rsf/CizxqkFw1YweYBo9aNgFIyCUTAKsAIAxp1Q7tq90EcAAAAASUVORK5CYII=","orcid":"","institution":"University of California, Los Angeles","correspondingAuthor":true,"prefix":"","firstName":"Ethan","middleName":"","lastName":"Cramer","suffix":""},{"id":548933978,"identity":"4b73a01d-f062-4e57-9d1d-7a27f55fc4f6","order_by":1,"name":"Sophie Dobiasch","email":"","orcid":"","institution":"Technical University of Munich","correspondingAuthor":false,"prefix":"","firstName":"Sophie","middleName":"","lastName":"Dobiasch","suffix":""},{"id":548933979,"identity":"cf2a7758-263a-40da-900d-5b6fab2d89c0","order_by":2,"name":"Ioannis I. Verginadis","email":"","orcid":"","institution":"University of Pennsylvania","correspondingAuthor":false,"prefix":"","firstName":"Ioannis","middleName":"I.","lastName":"Verginadis","suffix":""},{"id":548933980,"identity":"48a39a75-1370-400c-a14e-e41c25d554de","order_by":3,"name":"Xinmin Liu","email":"","orcid":"","institution":"University of California, Los Angeles","correspondingAuthor":false,"prefix":"","firstName":"Xinmin","middleName":"","lastName":"Liu","suffix":""},{"id":548933981,"identity":"6437332c-7bd8-4f66-85ae-cf99258a281b","order_by":4,"name":"Stephanie E. Combs","email":"","orcid":"","institution":"Technical University of Munich","correspondingAuthor":false,"prefix":"","firstName":"Stephanie","middleName":"E.","lastName":"Combs","suffix":""},{"id":548933982,"identity":"5ee16d09-69b0-4899-ac12-c450429bda29","order_by":5,"name":"Rodney D. Wiersma","email":"","orcid":"","institution":"University of California, Los Angeles","correspondingAuthor":false,"prefix":"","firstName":"Rodney","middleName":"D.","lastName":"Wiersma","suffix":""}],"badges":[],"createdAt":"2025-06-14 00:23:12","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-6891084/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6891084/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":98768409,"identity":"0e850615-4e6e-482f-b314-e33878031c63","added_by":"auto","created_at":"2025-12-22 10:25:13","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":3388802,"visible":true,"origin":"","legend":"","description":"","filename":"AutoSegNetPaperScientificReportsCramer.docx","url":"https://assets-eu.researchsquare.com/files/rs-6891084/v1/9732066a57aa00ffd20c3274.docx"},{"id":98768595,"identity":"3a9934fb-7f2e-48c2-94fe-1400b2777476","added_by":"auto","created_at":"2025-12-22 10:25:27","extension":"json","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":7484,"visible":true,"origin":"","legend":"","description":"","filename":"85fe77a5742c43ea98587606b64f24c1.json","url":"https://assets-eu.researchsquare.com/files/rs-6891084/v1/d33de5deeba620c9e62f35de.json"},{"id":98768410,"identity":"868c1482-3e87-433d-9fe9-236be4a388fd","added_by":"auto","created_at":"2025-12-22 10:25:13","extension":"xml","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":82121,"visible":true,"origin":"","legend":"","description":"","filename":"85fe77a5742c43ea98587606b64f24c11enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-6891084/v1/9abaa6439871273ea756577e.xml"},{"id":98768387,"identity":"c5112867-7778-407e-a937-cebc7ccbd106","added_by":"auto","created_at":"2025-12-22 10:25:10","extension":"png","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":621046,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-6891084/v1/35d81c43367e554f72914222.png"},{"id":98768517,"identity":"e8f98f4a-5a77-4490-8095-9cb77cd1f599","added_by":"auto","created_at":"2025-12-22 10:25:21","extension":"png","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":190993,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-6891084/v1/2fe4ae16ae172a778c79b8a1.png"},{"id":98768601,"identity":"7423136a-fec3-4aa5-a5b7-ebe6a2f78cbe","added_by":"auto","created_at":"2025-12-22 10:25:27","extension":"png","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":157177,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-6891084/v1/53535340179f23a420fe9b24.png"},{"id":98768528,"identity":"66b53918-e176-4c0d-9341-969876b5e584","added_by":"auto","created_at":"2025-12-22 10:25:22","extension":"png","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":21622,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-6891084/v1/6c448a3d66dd5cef5c449371.png"},{"id":98768598,"identity":"ffe3b98f-3dad-412b-9883-8c24f4091335","added_by":"auto","created_at":"2025-12-22 10:25:27","extension":"xml","order_by":11,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":81094,"visible":true,"origin":"","legend":"","description":"","filename":"85fe77a5742c43ea98587606b64f24c11structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-6891084/v1/d7cf3a88fcddc154af2ad7aa.xml"},{"id":98768413,"identity":"cf5559a0-b4ef-4239-a9f9-f662d71b13f8","added_by":"auto","created_at":"2025-12-22 10:25:14","extension":"html","order_by":12,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":90505,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-6891084/v1/6a39d3ddaeb145f070a1c6f8.html"},{"id":98780352,"identity":"55be5a9d-6563-4605-97a5-e379ee84c5fa","added_by":"auto","created_at":"2025-12-22 12:31:14","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":745260,"visible":true,"origin":"","legend":"\u003cp\u003eComparison of non-contrast-enhanced, native, CBCT scan versus iodine-based contrast-enhanced CBCT scan.\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-6891084/v1/184b400c5d525416976a3877.png"},{"id":98768538,"identity":"ec462d55-6d93-4ff0-b603-cb7d0251658c","added_by":"auto","created_at":"2025-12-22 10:25:24","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":1290510,"visible":true,"origin":"","legend":"\u003cp\u003eHuman ground truth 3D projection of the segmentation file (left) juxtaposed to the AI-contoured file (right) for a sample mouse. The left kidney is orange, right kidney is gold, spinal cord is aqua blue, bowel, stomach is blue, liver is purple, bowels are magenta, heart is light blue, left lung is bright green, right lung is chartreuse, and the bones are pink.\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-6891084/v1/88d122047fe71b81387fa170.png"},{"id":98768511,"identity":"20d8b857-10f6-41f5-b95d-96a006940909","added_by":"auto","created_at":"2025-12-22 10:25:21","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":1190149,"visible":true,"origin":"","legend":"\u003cp\u003e2D slice depictions in the coronal and sagittal views for one mouse in the 15-mouse testing dataset. The ground truth human segmentations are shown as a bold yellow outline. AI automatic segmentations are shown in the highlighted colors. The left kidney is orange, right kidney is gold, spinal cord is aqua blue, bowel, stomach is blue, liver is purple, bowels are magenta, heart is light blue, left lung is bright green, right lung is chartreuse, and the bones are pink.\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-6891084/v1/a5ba7f2e403dc23fe5051566.png"},{"id":98768542,"identity":"b6ae241c-a2dd-486c-9ab4-0e8f46488f66","added_by":"auto","created_at":"2025-12-22 10:25:24","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":81017,"visible":true,"origin":"","legend":"\u003cp\u003eDice scores of DL generated contours compared to manual human ground truth contours. Boxplots show the median (solid red line), interquartile range (top and bottom of the boxes), 1.5 times the interquartile range (whiskers), and outliers (circles). The left kidney is orange, right kidney is gold, spinal cord is aqua blue, bowel, stomach is blue, liver is purple, bowels are magenta, heart is light blue, left lung is bright green, right lung is chartreuse, and the bones are pink.\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-6891084/v1/7e2436b11a475cb8384d98fb.png"},{"id":98786727,"identity":"d66bd8e5-b655-46d8-9af8-18d0f3772ba3","added_by":"auto","created_at":"2025-12-22 12:43:41","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":4068321,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6891084/v1/a0edd64a-622d-4320-b61b-d95156889227.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Murine Organ Auto-Contouring in Small-Animal Precision Irradiation: A Comprehensive Approach Integrating Deep Learning and Contrast Enhancement for Onboard CBCT","fulltext":[{"header":"INTRODUCTION","content":"\u003cp\u003eAnimal models are the backbone of many areas of biomedical and preclinical research and the mouse is the most commonly used model organism for studying human disease\u003csup\u003e\u003cspan additionalcitationids=\"CR2 CR3\" citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e\u003c/sup\u003e. Understanding and characterizing mouse models in detail is considered key to improving the reproducibility of preclinical results in human subjects\u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e. For cancer-based radiobiology, research guiding clinical treatment often requires preliminary studies on the cellular or small animal scale\u003csup\u003e\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e,\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e\u003c/sup\u003e. However, to best model the actual radiobiological, radioimmunological, and toxicity characteristics of human therapies, preclinical techniques analogous to modern clinical strategies are needed. Recent studies have shown the possibility of mimicking the 3D conformal dose distributions of clinical intensity-modulated radiation therapy (IMRT) on small animal irradiations\u003csup\u003e\u003cspan additionalcitationids=\"CR9 CR10\" citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e\u003c/sup\u003e. However, a primary obstacle preventing widespread use of such techniques is the need to segment anatomic murine structures for treatment planning optimization. Depending on the site, simple single organ contouring (brain, heart, etc.) can take approximately 20 minutes, whereas, for highly complex cases involving many organs or skeleton contouring for total bone marrow irradiation can take many hours\u003csup\u003e\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e,\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e\u003c/sup\u003e. Such times are not suitable for large scale murine radiation biology studies which require many mouse deliveries per day to generate statistical significance of results. In addition, manual delineation of organs on each slice of a volumetric computed tomography (CT) scan requires not only great attention to detail, but a high-level of mouse anatomy expertise. Furthermore, there is subjectivity involved in manual segmentation due to individual bias.\u003c/p\u003e \u003cp\u003eRecent studies have investigated the use of automated segmentation approaches using machine learning. For instance, Lappas et al.\u003csup\u003e\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u003c/sup\u003e used a U-net architecture with three levels of encoding to extract features and three levels of decoding to automatically predict organ segmentation. This 3D U-net deep learning (DL) architecture model successfully predicted high-quality contours for the thorax and head area in mice cone-beam CT (CBCT) scans taken using an image-guided (IG) small animal irradiator. Van der Heyden et al. used a 3D U-net DL architecture model to determine skeletal muscle mass in small animal irradiator CBCT scans\u003csup\u003e\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e\u003c/sup\u003e. However, the lower quality of preclinical irradiator CBCT has so far limited AI to only areas that are high contrast\u0026mdash;minimally anatomically complex\u0026mdash;structures located in the head (eyes, brain) and thorax (heart, lungs) regions. To allow DL-based auto-contouring of abdominal and other more ambiguously shaped organs has required the use of a dedicated small animal \u0026micro;-CT scanner\u003csup\u003e\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e\u003c/sup\u003e. This greatly limits the use of AI organ segmentation for many preclinical radiation therapy (RT) studies as the mouse must be physically transferred from the \u0026micro;-CT to the small animal irradiator leading to potential position setup errors.\u003c/p\u003e \u003cp\u003eTo address this issue, and to further expand the utility of AI contouring in preclinical RT, we present a method of using a machine learning model with CBCT irradiators by combining a DL U-net-like architecture with iodine-based contrast-enhancement to efficiently and accurately contour unambiguous structures in the thorax and bones, as well as mostly ambiguous, abdominal mouse structures. To the best of our knowledge, this is the first method to perform high quality DL based auto contouring of abdominal structures in lower quality CBCT scans and has the potential to vastly improve the workflow efficiency of small animal IGRT preclinical trials.\u003c/p\u003e"},{"header":"RESULTS","content":"\u003cp\u003eAs means of optimizing detection of the OARs, injection of iodine contrast agent was used immediately before imaging to increase soft tissue contrast. The intestinal loops became partly air-filled, the kidneys appeared in the retroperitoneum, and the spine and all other skeleton structures appeared hyperdense. Figure\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e displays a comparison of CBCT scans of a native, non-contrast-enhanced mouse, and a contrast-enhanced mouse. As shown in the axial (top left), coronal (bottom), and sagittal (top right) slices, there was a clear advantage in using iodine-based contrast-enhancements.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe DL model was trained on an optimal dataset size of 26 mice. Of the 26 mice used, one mouse was used for testing during training, and another one used for validation for the Adam optimizer during training. The model was trained on 30 epochs and tested on a test set of 15 mice. Although more training mice were available, it was found that in general, going beyond a 26-mouse dataset resulted in only marginal dice score improvements and caused increased noise. Figure\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e shows representatively a side-by-side comparison of 3D projections of human segmented ground truth file (left) and AI automatically segmented file (right). As shown, the two look remarkably similar, which has extremely positive implications for the efficacy of the DL model. In general, for the best contours, the DL generated segmentations for each structure almost perfectly match the human ground truth contours, while in the worst case there were only a few regions of disagreement.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e shows example slices of the coronal and sagittal views for one mouse from the test dataset, with the ground truth human contours outlined in bolded yellow. The DL generated contours are shown highlighting each predicted structure. Areas of disagreement came in the form of either a pixel characterized as not part of a specific structure, when the ground truth file had that pixel as a part of that structure, or a pixel is characterized as part of a specific structure by the DL model, while that pixel is not part of the organ in the ground truth file.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e displays the dice score boxplots results for each structure, as well as the median of all structures, for all 15 mice used for evaluation. Median dice scores for each organ are reported: left kidney\u0026thinsp;=\u0026thinsp;99.36%, right kidney\u0026thinsp;=\u0026thinsp;99.36%, spinal cord\u0026thinsp;=\u0026thinsp;75.79%, stomach\u0026thinsp;=\u0026thinsp;96.52%, liver\u0026thinsp;=\u0026thinsp;95.92%, bowel\u0026thinsp;=\u0026thinsp;95.65%, heart\u0026thinsp;=\u0026thinsp;98.75%, left lung\u0026thinsp;=\u0026thinsp;98.23%, right lung\u0026thinsp;=\u0026thinsp;98.24%, bones\u0026thinsp;=\u0026thinsp;94.77%, and average for all organs\u0026thinsp;=\u0026thinsp;97.38%. With exception of the spinal cord, the median dice score was higher than 85%, meaning overall performance of the DL automatic segmentation model was very high\u003csup\u003e\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u003c/sup\u003e. The kidneys showed consistent accuracy of over 99%, whereas the stomach, liver, and bowel, were slightly less accurate, but still consistently over 95%. Segmented structures in the thorax region, as expected, performed very high, consistently over 98%. Nevertheless, where the model lacked in accuracy is longitudinally, in the spinal cord for instance, where there was high variance, ranging from a dice score of 40\u0026ndash;90%. Since the spinal cord appears in so few slices, any inaccuracies are amplified. Thus, the worst contours were still accurately contoured with a dice score of 75%, which is still highly accurate for a small longitudinal structure.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eFor all ten annotated structures\u0026ndash;left kidney, right kidney, spinal cord, stomach, liver, bowels, heart, left lung, right lung, and bones\u0026ndash;each evaluation metric is presented in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. Similar trends in the results can be observed in the Jaccard and precision scores. According to the Jaccard scores, the least accurate segmentations came in the spinal cord. The model produced extremely good precision scores, with the least precise predicted volume segmentation came in the spinal cord, with a precision score of about 80%. The false negative rate (FNR) or miss rate of the DL-enabled mouse automatic segmentation model for each organ is shown in the fifth column of Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. Overall, there was a median miss rate of about 4%, which shows high accuracy of the DL automatic contours. The HD50 median for all organs was 0.00 mm, while the HD95 averaged for all organs was about 2.12 mm. The median of the mean surface distance for all organs was 0.55 mm. The median standard deviation for all organs was 0.83 mm. Overall, distance measurements showed strong agreement with quality metrics, showing high accuracy of DL automatic contours. According to the surface distance measurements, consistently, the least accurately contoured organ was the bowel. This is because the bowel is the largest area to contour, as the bowel is spread along the whole abdomen. The spinal cord, however, had the largest miss rate at 21%, as shown in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eSegmentation accuracy evaluated using the dice, Jaccard score, precision score, FNR, 50th percentile and 95th percentile HDs, mean, median, and standard deviation of surface distances (SD). The values displayed are the medians for each organ over the while 15-mice test dataset.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"10\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c9\" colnum=\"9\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c10\" colnum=\"10\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colspan=\"9\" nameend=\"c10\" namest=\"c2\"\u003e \u003cp\u003eMedian of Metric\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eOrgan\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDice\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eJaccard\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003ePrecision\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eFNR\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eHD\u003csub\u003e50\u003c/sub\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eHD\u003csub\u003e95\u003c/sub\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eMean SD\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003eMedian SD\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003eStd. SD\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLeft Kidney\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e99.36\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e98.73\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e98.67\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.01\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e1.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e0.18\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e0.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e0.48\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRight Kidney\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e99.36\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e98.73\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e99.62\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.01\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e1.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e0.16\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e0.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e0.48\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSpinal Cord\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e75.79\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e61.02\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e79.46\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.21\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e1.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e3.90\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e1.40\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e1.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e1.32\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eStomach\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e96.52\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e93.28\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e98.70\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.06\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e2.83\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e0.77\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e0.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e0.99\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLiver\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e95.92\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e92.16\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e97.06\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.04\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e1.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e4.47\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e1.31\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e1.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e1.69\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBowel\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e95.65\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e91.67\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e95.79\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.05\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e2.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e9.05\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e2.80\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e2.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e2.99\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHeart\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e98.75\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e97.53\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e99.73\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.02\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e2.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e0.39\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e0.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e0.67\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLeft Lung\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e98.23\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e96.53\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e99.03\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.03\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e2.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e0.23\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e0.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e0.55\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRight Lung\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e98.24\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e96.54\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e99.44\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.02\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e1.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e0.19\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e0.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e0.48\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBones\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e94.77\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e90.05\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e96.02\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.07\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e2.24\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e0.71\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e0.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e1.59\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAll\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e97.38\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e94.91\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e98.69\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.04\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e2.12\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e0.55\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e0.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e0.83\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e"},{"header":"DISCUSSION","content":"\u003cp\u003eAs preclinical irradiators continue to advance to closely mimic their clinical Linac counterparts, there is a requirement for high-quality automated 3D contours for large-scale 3D conformal RT small animal studies. Not only does automatic organ segmentation vastly increase efficiency of these preclinical murine studies, but it also eliminates human bias and fatigue which can lead to errors. Previous DL-enabled automated segmentation for small animal methods have either required dedicated micro-CT scanners for whole body murine imaging or have been limited to only high-contrast, non-anatomically complex thoracic and head regions when using small animal irradiator CBCT\u003csup\u003e\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e,\u003cspan additionalcitationids=\"CR17 CR18\" citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e\u003c/sup\u003e. This work demonstrates that despite the image resolution and contrast limitations of small animal irradiator CBCT, the combination of iodine-based contrast-enhancement and U-net-like DL architecture can allow high-quality contours of more difficult to see abdominal organs without the need for a micro-CT. Additionally, unlike previous DL contouring models, which required large training datasets of 75\u0026thinsp;+\u0026thinsp;mice, this work demonstrated the ability to create a high accuracy full-body segmentation models with a training dataset of only 26 mice\u003csup\u003e\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eManual structure segmentation can take upwards of 10 minutes per organ, making the total segmentation time in this study over 75 hours\u003csup\u003e\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u003c/sup\u003e. Conversely, the DL model can accurately segment all eleven mouse organs and skeletal structures in under a second, vastly reducing contouring time and increasing efficiency of the image-guided precision RT workflow. The U-net-like DL architecture proved as a method to accurately, automatically contour murine abdominal and thorax organs, as well as skeletal and heterogenous tumor structures, right and left kidney\u0026rsquo;s, spinal cord, stomach, liver, bowel, heart, left and right lungs, and bones in under a second. Upon training on a 26-mouse dataset of manual expert and semi-autonomous segmentation, performance of the test mice for most of the organs showed high-quality automatic contours, consistently way over a dice score of 85% and HD95 of less than 2.15 mm. A dataset size of 26 mice was chosen based on preliminary results, showing a saturation effect in performance when including additional mice in the training dataset.\u003c/p\u003e \u003cp\u003eThere are sources for uncertainty in this DL-enabled automatic contouring architecture that influences the quality of the generated segmentations\u003csup\u003e\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u003c/sup\u003e. Most notably, there is uncertainty that can arise in the input. Manual, human contours may not represent a true ground truth segmentation due to subjective interpretation and bias, which must be taken into account\u003csup\u003e\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e\u003c/sup\u003e. This subjective interpretation effect is quantified using inter-observer variability (IOV) studies, which potentially causes regions of disagreement. Additionally, the observer performing the manual segmentation of organs may suffer from fatigue and time pressure, causing them to miss contours over time, causing incompleteness and disagreements. Limitations of using DL for automated delineations of small animal structures mainly stem from CBCT artifacts causing the model to predict pixels that are a part of the CT bench as part of a structure. This effect, however, can be quickly and easily resolved in most treatment planning software by simply deleting the contours that are clearly out of the image of the small animal.\u003c/p\u003e \u003cp\u003eThe model lacked accuracy mainly in areas often considered ambiguous for manual segmentation specialists. The metrics presented are extremely sensitive and consider all differences between the DL automatic contours and human ground truth segmentation contours. Thus, relatively poorer scores can be observed longitudinally in the spinal cord, as found in previous studies. Nevertheless, this model is applicable to other contrast agent enhanced CBCTs.\u003c/p\u003e \u003cp\u003eIn this work, a deep learning (DL) model was successfully trained to accurately predict contours for low-contrast abdominal structures in preclinical irradiator CBCT for the first time. The hybrid approach combining DL with an iodine contrast agent is efficient and enhances setup accuracy since the mouse does not need to be moved from the irradiator for dedicated \u0026micro;-CT low-contrast tissue imaging. By significantly reducing the manual contouring burden, the proposed method makes advanced 3D conformal preclinical radiotherapy (RT) techniques, such as IMRT, more practical for large-scale mouse studies.\u003c/p\u003e"},{"header":"METHODS","content":"\u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003eDataset and Image Conditioning\u003c/h2\u003e \u003cp\u003eThe dataset consisted of 41 6-week-old immunosuppressed CD-1 nude mice partly taken from a previously completed pancreatic cancer study\u003csup\u003e\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e\u003c/sup\u003e. In this study, all animal procedures were carried out in accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. Moreover, the protocol was officially approved and authorized by German law for the Care and Use of Laboratory Animals and the regional government of Upper Bavaria, Germany (ref. 55.2-1-54-2532-217-2025 from April 13, 2016). All surgical procedures were performed under anesthesia, and all efforts were made to minimize suffering\u003csup\u003e\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e\u003c/sup\u003e. Additionally, experimental descriptions in this manuscript comply with ARRIVE guidelines. Treatment planning and irradiation of the mice were performed when the GTV reached 90 mm\u003csup\u003e\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e\u003c/sup\u003e, which took about eight weeks post injection of Panc-1 (CRL-1469) and three weeks for MiaPaCa-2 (CRL-1420). Anesthesia was used during the entirety of the imaging process to immobilize the mice. Each mouse inhaled isoflurane anesthesia at a concentration of 1.5% with a 6% volume of oxygen as a carrier of the gas. Imaging was done using the onboard cone-beam computed tomography (CBCT) system of a small animal irradiator (SARRP, Xstrahl, GA) with the operating x-ray source at a voltage of 60 kV and current of 0.8 mA. In addition, each mouse was injected with 5 mL per 1 kg mice body weight of iodine-containing contrast agent (Imeron, Bracco, Italy) intravenously (iv) through the lateral tail vessel to improve soft tissue contrast and optimize the definition of the OARs. CBCT imaging was taken immediately after iv injection using 1440 projections and a voxel size of 0.115 x 0.115 x 0.115 mm\u003csup\u003e\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e\u003c/sup\u003e. Mice were ethnized one-hour post-irradiation treatment by cervical dislocation\u003csup\u003e\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eTreatment planning, including manual segmentation of mice structures, was done using the preclinical treatment planning software \u0026lsquo;MuriPlan\u0026rsquo;. The bones, along with six OARs, including both kidneys, bowel, stomach, spinal cord, and liver, were manually delineated slice-by-slice of the CBCT image by a segmentation specialist. Contouring of the heart, lungs, and bones was done semi-automatically, using a previously built model to predict contours, then carefully edited slice-by-slice to fix any potential errors. When inputting a native CBCT scan into the DL-based automatic segmentation model, performance is very poor, and limits quality contours to only the thorax region, as well as the bone structure. For this reason, the use of contrast-enhancement immediately before CBCT imaging was imperative to building an accurate abdominal organ automatic segmentation model, with this U-net-like DL architecture.\u003c/p\u003e \u003cp\u003eAs a means of standardizing the 3D Nifti encoded CBCT images before model training, a script was first used to check for completeness of the available scans, verifying them for consistency in that both the CBCT scan and annotated 3D Nifti volumes are of the same size. After this each volume is turned into coronal slices to allow for faster processing by slicing the 3D Nifti volume into 2D tiff images. The signal intensity was then normalized by subtracting the mean and dividing by the standard deviation. The images were then resampled to a resolution of 240 \u0026micro;m/vx, even though the DL U-Net-like architecture works over a broad range of 120\u0026ndash;1120 \u0026micro;m/vx\u003csup\u003e16\u003c/sup\u003e.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eDevices and Python Libraries\u003c/h3\u003e\n\u003cp\u003eThe algorithm, developed in python, was implemented using these open-source libraries: PyTorch, SciPy, NiBabel, Nrrd, Numpy, and Matplotlib. The deep learning framework pipeline uses PyTorch. Additionally, a NVIDIA Titan Xp GPU was used, but not necessary as the program can be run using a CPU, at the cost of time. The pipeline, proposed by Schoppe et al.\u003csup\u003e\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e\u003c/sup\u003e, consisted of three parts, a preprocessor, deep learning backbone, and a postprocessor.\u003c/p\u003e\n\u003ch3\u003eDeep Learning Model\u003c/h3\u003e\n\u003cp\u003eThe DL backbone follows a U-Net-like architecture, consisting of an encoding and decoding path connected with skip connections. The varying convolutional sizes help to extract features at various resolutions and without skip connections, upsampling back to the original resolution would be very difficult. Moreover, the use of skip connections resolved the vanishing gradient problem by providing an uninterrupted gradient flow through all layers\u003csup\u003e\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e\u003c/sup\u003e. Encoding levels consisted of two convolutions (kernel size: 3, padding: 1, stride: 1), batch normalization, a rectifying linear activation function (ReLu), and a max-pooling operation (kernel size: 2, stride: 2). Additionally, each encoding level had twice the number of featured channels compared to its previous level, starting at 32 feature channels. The decoding units consisted of three convolutions (kernel size: 3, padding: 1, stride: 1) and received concatenation of upsampled input from the previous level, including input from the skip connection from the corresponding encoding level (consisting of the same number of feature channels). The final convolution maps the 32 feature channels to the number of structures predicted. The last pass was through a sigmoid function which creates a volumetric probability map for each structure.\u003c/p\u003e \u003cp\u003eThe DL model was trained with a soft-Dice loss function and used the Adam optimizer\u003csup\u003e\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e\u003c/sup\u003e. A soft-Dice loss function was used to directly predict probabilities of target structure position, as opposed to converting them into a binary mask. The soft-Dice loss function is a differential reformation of the commonly used Dice Similarity Coefficient (DSC) evaluation metric. If \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:A\\)\u003c/span\u003e\u003c/span\u003e represents pixels from the human segmentation ground truth file, and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:B\\)\u003c/span\u003e\u003c/span\u003e consists of pixels from the DL automatically segmented file, the DSC can be calculated as two times the intersection of A and B divided by the sum of the two sets: \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:DSC\\:=\\:\\frac{2|A\\cap\\:B|}{\\left|A\\right|\\:+\\:\\left|B\\right|}\\:\\)\u003c/span\u003e\u003c/span\u003e\u003csup\u003e2\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e\u003c/sup\u003e. Letting \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:2\\left|A\\cap\\:B\\right|={\\sum\\:}_{i}{a}_{i}*{b}_{i}=\\:\u0026lt;A,B\u0026gt;\\)\u003c/span\u003e\u003c/span\u003e and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\left|A\\right|=\\:{\\sum\\:}_{i}{a}_{i}*a=\\:\u0026lt;A,A\u0026gt;\\)\u003c/span\u003e\u003c/span\u003e, a differential soft-Dice loss function can be determined as:\u003cdiv id=\"Equa\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equa\" name=\"EquationSource\"\u003e\n$$\\:{L}_{Dice}=1-\\:\\frac{\\:2\u0026lt;A,B\u0026gt;\\:}{\u0026lt;A,A\u0026gt;\\:+\\:\u0026lt;B,B\u0026gt;}\\:$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eThe Adam optimizer was used as a method for computing individual adaptive learning rates for different parameters from estimates of first and second moments of the gradients\u003csup\u003e\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e\u003c/sup\u003e. Furthermore, a nested k-fold cross-validation procedure was used to split the dataset into a training set (for model weight optimization), a validation set (for hyper-parameter optimization), and a test set (for evaluation)\u003csup\u003e\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e\u003c/sup\u003e. The DL model was trained for 30 epochs, with an initial learning rate of 10\u003csup\u003e\u0026minus;\u0026thinsp;3\u003c/sup\u003e, on the training set, with the learning rate being gradually reduced as validation performance was unchanged over five epochs. The postprocessor utilizes ensemble dash voting, where the neural network uses information from many independently trained networks and merges them to make the most accurate prediction possible\u003csup\u003e\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e\u003c/sup\u003e. Finally, the volumetric probability map from the DL backbone is used for volumetric reconstruction to create a predicted segmentation file in the postprocessor.\u003c/p\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003eEvaluation Metrics\u003c/h2\u003e \u003cp\u003eBoth quality and distance metrics are presented for performance evaluation. Quality metrics, such as the Dice Similarity Score, or just dice score, is an evaluation of the qualitative performance of the algorithm with respect to the human segmentation for each structure. Quality metrics use voxel classification of the DL-predicted contouring of a CBCT scan for each structure, denoting each voxel as either \u0026lsquo;True Positive\u0026rsquo; (TP), \u0026lsquo;False Positive\u0026rsquo; (FP), or \u0026lsquo;False Negative\u0026rsquo; (FN). Voxels are classified as TP if it was correctly predicted compared to the ground truth file, FP if the voxel was incorrectly predicted as part of an organ, and FN if the voxel was predicted to be a non-organ but is a part of an organ in the ground truth file. With this classification of the pixels, additional metrics were utilized to assess accuracy of automatically segmented files against human segmentation, such as those presented in the Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e. The quality metrics are based on two sets of pixels, A and B, which consists of pixels from the DL-predicted segmentation and the human ground truth segmentation, respectively. The dice index gives the percent overlap, or reproducibility\u0026ndash;how well the segmentation and ground truth files match\u0026ndash;for each structure. The Jaccard index is the intersection of A and B over the union of A and B. The precision gives the rate at which pixels are correctly classified as part of the correct structure. FP rate and FN rate give the rate at which false positives classifications are found and false negative predictions are found, respectively\u003csup\u003e\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eQuality metrics used for performance evaluation of the DL model (left) and distance metrics used for evaluation of DL model (right)\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"2\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eQuality Metrics Formula\u003csup\u003e\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDistance Metrics Formula\u003csup\u003e\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:Dice\\:Score\\:=\\:\\frac{2TP}{2TP\\:+\\:FP\\:+\\:FN\\:+\\:ϵ}\\times\\:100\\%\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:h(A,B)\\:=\\:{max}_{a\\in\\:A}{min}_{b\\in\\:B}|A-B|\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e \u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:HD(A,B)\\:=\\:max\\left(h\\right(A,B),\\:h(B,A)\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:Jaccard\\:Index\\:=\\:\\frac{TP}{TP\\:+\\:FP\\:+\\:FN\\:+\\:ϵ}\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{HD}_{p}(A,B)\\:=\\:percentile\\left[max\\right(h(A,B),\\:h(B,A)\\left)\\right]\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:Precision\\:=\\:\\frac{TP}{TP\\:+\\:FP\\:+\\:ϵ}\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:Mean\\:SD\\:=\\:mean\\left(h\\right(A,B),\\:h(B,A\\left)\\right)\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:FP\\:rate\\:=\\:\\frac{FP}{FP\\:+\\:TP\\:+\\:ϵ}\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:Median\\:SD\\:=\\:median\\left(h\\right(A,B),\\:h(B,A\\left)\\right)\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:FN\\:rate\\:=\\:\\frac{FN}{FN\\:+\\:TP\\:+\\:ϵ}\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:std\\:SD\\:=\\:std\\left(h\\right(A,B),\\:h(B,A\\left)\\right)\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:ϵ\\:=\\:0.0001\\:\\)\u003c/span\u003e\u003c/span\u003efor numerical stability\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eSD\u0026thinsp;=\u0026thinsp;Surface Distance, HD\u0026thinsp;=\u0026thinsp;Hausdorff Distance\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eSpatial distance metrics were used to quantitatively evaluate the accuracy of the predicted contours, taking into consideration the spatial position of voxels. The directed Hausdorff (h(A,B)), as shown in the Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e, where |a-b| is the difference in the Euclidean distance between pixels in their respective pixel sets from predicted contouring of a structure and ground truth contours. The Hausdorff Distance (HD) is the distance between crisp volumes (two finite pixel sets, A and B). Since noise and outlier pixels are very common in medical segmentation, it is not recommended to use the maximum HD. Instead, the quantile method proposed by Huttenlocher et al.\u003csup\u003e\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e\u003c/sup\u003e is one way to handle outliers, where HD is defined as some percentile of the maximum distance\u003csup\u003e\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e\u003c/sup\u003e. For this reason, the 50th (HD50) and 95th (HD95) percentile HDs are presented. The mean surface distance gives the mean distances between predicted and human segmentations for each structure. The median and standard deviation of these distances are also reported.\u003c/p\u003e \u003c/div\u003e"},{"header":"Declarations","content":"\u003cp\u003e \u003ch2\u003eCompeting interests\u003c/h2\u003e \u003cp\u003eNo conflict exists for any author.\u003c/p\u003e \u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eE.C. fabricated DL segmentation model, conducted experiments, and drafted manuscript. S.D. provided mouse CBCT data, administered contrast agent, contributed to manual segmentation, and results analysis. I.V. contributed to manual segmentation efforts and analyzed results. X.L. and S.C. analyzed results. R.W. conceived of experiment and analyzed results. All authors reviewed the manuscript.\u003c/p\u003e\u003ch2\u003eData Availability\u003c/h2\u003e\u003cp\u003eData is available from the corresponding author upon reasonable request.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eRosenthal, N. \u0026amp; Brown, S. The mouse ascending: perspectives for human-disease models. \u003cem\u003eNat. Cell. Biol.\u003c/em\u003e \u003cb\u003e9\u003c/b\u003e, 993\u0026ndash;999 (2007).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eOsuchowski, M. F. et al. Abandon the mouse research ship? Not just yet! \u003cem\u003eShock Augusta Ga.\u003c/em\u003e \u003cb\u003e41\u003c/b\u003e, 463\u0026ndash;475 (2014).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSchoppe, O. et al. Deep learning-enabled multi-organ segmentation in whole-body mouse scans. \u003cem\u003eNat. Commun.\u003c/em\u003e \u003cb\u003e11\u003c/b\u003e, 5626 (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHackam, D. G. \u0026amp; Redelmeier, D. A. Translation of research evidence from animals to humans. \u003cem\u003eJAMA\u003c/em\u003e \u003cb\u003e296\u003c/b\u003e, 1731\u0026ndash;1732 (2006).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePerrin, S. Preclinical research: Make mouse studies work. \u003cem\u003eNature\u003c/em\u003e \u003cb\u003e507\u003c/b\u003e, 423\u0026ndash;425 (2014).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKahn, J., Tofilon, P. J. \u0026amp; Camphausen, K. Preclinical models in radiation oncology. \u003cem\u003eRadiat. Oncol. Lond. Engl.\u003c/em\u003e \u003cb\u003e7\u003c/b\u003e, 223 (2012).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVan Dyke, T. \u0026amp; Jacks, T. Cancer modeling in the modern era: progress and challenges. \u003cem\u003eCell\u003c/em\u003e \u003cb\u003e108\u003c/b\u003e, 135\u0026ndash;144 (2002).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWoods, K., Neph, R., Nguyen, D. \u0026amp; Sheng, K. A sparse orthogonal collimator for small animal intensity-modulated radiation therapy. Part II: hardware development and commissioning. \u003cem\u003eMed. Phys.\u003c/em\u003e \u003cb\u003e46\u003c/b\u003e, 5733\u0026ndash;5747 (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJiang, L., Lyu, Q., Abdelhamid, A. M. H., Hui, S. \u0026amp; Sheng, K. An efficient rectangular optimization method for sparse orthogonal collimator based small animal irradiation. \u003cem\u003ePhys. Med. Biol.\u003c/em\u003e \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1088/1361-6560/ac910b\u003c/span\u003e\u003cspan address=\"10.1088/1361-6560/ac910b\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRedler, G. et al. Small Animal IMRT Using 3D-Printed Compensators. \u003cem\u003eInt. J. Radiat. Oncol. Biol. Phys.\u003c/em\u003e \u003cb\u003e110\u003c/b\u003e, 551\u0026ndash;565 (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiu, X. et al. Improving the efficiency of small animal 3D-printed compensator IMRT with beamlet intensity total variation regularization. \u003cem\u003eMed. Phys.\u003c/em\u003e \u003cb\u003e49\u003c/b\u003e, 5400\u0026ndash;5408 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWang, H., Stout, D. B. \u0026amp; Chatziioannou, A. F. Estimation of mouse organ locations through registration of a statistical mouse atlas with micro-CT images. \u003cem\u003eIEEE Trans. Med. Imaging\u003c/em\u003e. \u003cb\u003e31\u003c/b\u003e, 88\u0026ndash;102 (2012).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003evan der Heyden, B. et al. Automatic multiatlas based organ at risk segmentation in mice. \u003cem\u003eBr. J. Radiol.\u003c/em\u003e \u003cb\u003e92\u003c/b\u003e, 20180364 (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLappas, G. et al. Automatic contouring of normal tissues with deep learning for preclinical radiation studies. \u003cem\u003ePhys. Med. Biol.\u003c/em\u003e \u003cb\u003e67\u003c/b\u003e, 044001 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVan Der Heyden, B. et al. Automated CT-derived skeletal muscle mass determination in lower hind limbs of mice using a 3D U-Net deep learning network. \u003cem\u003eJ. Appl. Physiol.\u003c/em\u003e \u003cb\u003e128\u003c/b\u003e, 42\u0026ndash;49 (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSchoppe, O. et al. Deep learning-enabled multi-organ segmentation in whole-body mouse scans. \u003cem\u003eNat. Commun.\u003c/em\u003e \u003cb\u003e11\u003c/b\u003e, 5626 (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLappas, G. et al. Inter-observer variability of organ contouring for preclinical studies with cone beam Computed Tomography imaging. \u003cem\u003ePhys. Imaging Radiat. Oncol.\u003c/em\u003e \u003cb\u003e21\u003c/b\u003e, 11\u0026ndash;17 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003evan de Worp, W. R. P. H. et al. Deep Learning Based Automated Orthotopic Lung Tumor Segmentation in Whole-Body Mouse CT-Scans. \u003cem\u003eCancers\u003c/em\u003e \u003cb\u003e13\u003c/b\u003e, 4585 (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRosenhain, S. et al. A preclinical micro-computed tomography database including 3D whole body organ segmentations. \u003cem\u003eSci. Data\u003c/em\u003e. \u003cb\u003e5\u003c/b\u003e, 180294 (2018).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDobiasch, S. et al. Histopathological Tumor and Normal Tissue Responses after 3D-Planned Arc Radiotherapy in an Orthotopic Xenograft Mouse Model of Human Pancreatic Cancer. \u003cem\u003eCancers\u003c/em\u003e \u003cb\u003e13\u003c/b\u003e, 5656 (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNeupane, B., Aryal, J. \u0026amp; Rajabifard, A. Dual skip connections in U-Net, ResUnet and U-Net3\u0026thinsp;+\u0026thinsp;for remote extraction of buildings. Preprint at (2023). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttp://arxiv.org/abs/2303.09064\u003c/span\u003e\u003cspan address=\"http://arxiv.org/abs/2303.09064\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGaldran, A., Carneiro, G. \u0026amp; Ballester, M. \u0026Aacute;. G. On the Optimal Combination of Cross-Entropy and Soft Dice Losses for Lesion Segmentation with Out-of-Distribution Robustness. Preprint at (2022). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttp://arxiv.org/abs/2209.06078\u003c/span\u003e\u003cspan address=\"http://arxiv.org/abs/2209.06078\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKingma, D. P., Ba, J. \u0026amp; Adam A Method for Stochastic Optimization. Preprint at (2017). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttp://arxiv.org/abs/1412.6980\u003c/span\u003e\u003cspan address=\"http://arxiv.org/abs/1412.6980\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTaha, A. A. \u0026amp; Hanbury, A. Metrics for evaluating 3D medical image segmentation: analysis, selection, and tool. \u003cem\u003eBMC Med. Imaging\u003c/em\u003e. \u003cb\u003e15\u003c/b\u003e, 29 (2015).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHuttenlocher, D. P., Klanderman, G. A. \u0026amp; Rucklidge, W. J. Comparing images using the Hausdorff distance. \u003cem\u003eIEEE Trans. Pattern Anal. Mach. Intell.\u003c/em\u003e \u003cb\u003e15\u003c/b\u003e, 850\u0026ndash;863 (1993).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eM\u0026uuml;ller, D. et al. MISeval: a Metric Library for Medical Image Segmentation Evaluation. Preprint at (2022). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttp://arxiv.org/abs/2201.09395\u003c/span\u003e\u003cspan address=\"http://arxiv.org/abs/2201.09395\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"","lastPublishedDoi":"10.21203/rs.3.rs-6891084/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6891084/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eModern preclinical irradiators have evolved to mimic their clinical Linac counterparts in terms of 360-degree beam delivery and on-board imaging capabilities with CBCT. The primary factor preventing widespread 3D conformal small-animal RT is the necessity of manually segmentation, as this task is time-consuming and impractical for large-scale studies. Although DL-based auto-contouring methods have been explored for preclinical irradiator CBCT, these methods have been limited to high-contrast, minimally anatomically complex structures. Thus, DL-based segmentation for low-contrast abdominal structures has yet to be addressed. In combining DL with iodine-based contrast-agent, precise full-body auto-contouring was achieved. A U-net-like architecture was trained to contour kidneys, spinal cord, stomach, liver, bowels, heart, lungs, and bones in small-animal irradiator CBCT mouse scans. Post contrast-enhancement, 41 mice were manually contoured, establishing ground truths. The model was trained with 26 mice, 2 for validation, and 15 for testing. Performance was evaluated using dice, precision, HD, and MSD. The proposed model predicted high-quality contours within a second, with the median for all organs reported: dice\u0026thinsp;\u0026gt;\u0026thinsp;97%, precision\u0026thinsp;\u0026gt;\u0026thinsp;98%, HD\u003csub\u003e95\u003c/sub\u003e\u0026thinsp;\u0026lt;\u0026thinsp;2.15 mm, and MSD\u0026thinsp;\u0026lt;\u0026thinsp;0.55 mm. The proposed combination of a DL and contrast-enhanced model is a viable method to vastly improve efficiency of small-animal IGRT.\u003c/p\u003e","manuscriptTitle":"Murine Organ Auto-Contouring in Small-Animal Precision Irradiation: A Comprehensive Approach Integrating Deep Learning and Contrast Enhancement for Onboard CBCT","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-12-22 10:24:10","doi":"10.21203/rs.3.rs-6891084/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"editorInvitedReview","content":"","date":"2025-10-19T02:24:09+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-10-09T02:18:02+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"118258446979886622865321511373130157782","date":"2025-10-06T19:11:14+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"279236520315971510677103744915750329020","date":"2025-09-29T08:09:03+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-09-18T15:47:47+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-09-17T09:44:53+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2025-06-17T10:46:01+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-06-17T04:28:33+00:00","index":"","fulltext":""},{"type":"submitted","content":"Scientific Reports","date":"2025-06-14T00:16:51+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"b8bf5f26-abb2-43d1-b907-afd589a579b0","owner":[],"postedDate":"December 22nd, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[{"id":58416373,"name":"Biological sciences/Biophysics/Computational biophysics"},{"id":58416374,"name":"Biological sciences/Computational biology and bioinformatics/Machine learning"},{"id":58416375,"name":"Health sciences/Medical research/Pre clinical studies"},{"id":58416376,"name":"Biological sciences/Cancer/Cancer therapy/Radiotherapy"}],"tags":[],"updatedAt":"2025-12-22T10:24:10+00:00","versionOfRecord":[],"versionCreatedAt":"2025-12-22 10:24:10","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-6891084","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6891084","identity":"rs-6891084","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}