Search

Abstract Background Accurate calculation of lung cancer dose using the Monte Carlo (MC) algorithm in CyberKnife (CK) is essential for precise planning. We aim to employ deep learning to directly predict the 3D dose distribution calculated by the MC algorithm, enabling rapid and accurate automatic planning. However, most current methods solely focus on conventional intensity-modulated radiation therapy and assume a consistent beam configuration across all patients. This study seeks to develop a more versatile model incorporating variable beam configurations of CK and considering the patient’s anatomy. Methods This study proposed that the AB (anatomy and beam) model be compared with the control Mask (only anatomy) model. These models are based on a 3D U-Net network to investigate the impact of CK beam encoding information on dose prediction. The study collected 86 lung cancer patients who received CK′s built-in MC algorithm plans using different beam configurations for training/validation (66 cases) and testing (20 cases). We compared the gamma passing rate, dose difference maps, and relevant dose-volume metrics to evaluate the model’s performance. In addition, the Dice similarity coefficients (DSCs) were calculated to assess the spatial correspondence of isodose volumes. Results The AB model demonstrated superior performance compared to the Mask model, particularly in the trajectory dose of the beam. The DSCs of the AB model were 20–40% higher than that of the Mask model in some dose regions. We achieved approximately 99% for the PTV and generally more than 95% for the organs at risk (OARs) referred to the clinical planning dose in the gamma passing rates (3 mm/3%). Relative to the Mask model, the AB model exhibited more than 90% improvement in small voxels (p

Abstract Background N4-Acetylcytidine (ac4C), a highly conserved post-transcriptional mechanism, plays a pivotal role in RNA modification and tumor progression. However, the molecular mechanism by which ac4C modification mediates tumor immunosuppression remains elusive in triple-negative breast cancer (TNBC). Methods NAT10 expression was analyzed in TNBC samples in the level of mRNA and protein, and compared with the corresponding normal tissues. ac4C modification levels also measured in the TNBC samples. The effects of NAT10 on immune microenvironment and tumor metabolism were investigated. NAT10-mediated ac4C and its downstream regulatory mechanisms were determined in vitro and in vivo. The combination therapy of targeting NAT10 in TNBC was further explored. Results The results revealed that the loss of NAT10 inhibited TNBC development and promoted T cell activation. Mechanistically, NAT10 upregulated JunB expression by increasing ac4C modification levels on its mRNA. Moreover, JunB further up-regulated LDHA expression and facilitated glycolysis. By deeply digging, remodelin, a NAT10 inhibitor, elevated the surface expression of CTLA-4 on T cells. The combination of remodelin and CTLA-4 mAb can further activate T cells and inhibite tumor progression. Conclusion Taken together, our study demonstrated that the NAT10-ac4C-JunB-LDHA pathway increases glycolysis levels and creates an immunosuppressive tumor microenvironment (TME). Consequently, targeting this pathway may assist in the identification of novel therapeutic strategies to improve the efficacy of cancer immunotherapy.

Abstract Background The novel anti-HER2 antibody drug conjugates (ADCs) can effectively improve the long-term survival of patients with HER2-low expression breast cancer. However, pathological responses to neoadjuvant therapy (NAT) within HER2-low expression breast cancer, the relationship between pathological response and prognosis and the transformation of HER2 status are all now poorly understood. Methods The patients with HER2-0 and HER2-low expression breast cancer receiving NAT at Harbin Medical University Cancer Hospital between Jan. 2014 and Nov. 2018 were retrospectively explored. HER2 low expression refers to the IHC 1 + or 2 + and FISH negative. The Kappa test was utilized for analyzing the consistency rate of HER2 expression. To evaluate disease-free survival (DFS) and overall survival (OS), this research employed both the Kaplan-Meier analysis and the Cox regression. Results In this study, 178 patients with HER2-0 and 344 patients with HER2-low expression breast cancer were included. In comparison with the HER2-0 group, it is shown that patients in the HER2-low group have more possibility to be younger compared to those 50 years old (P

Abstract Purpose To evaluate the dosimetric characteristics of ZAP-X stereotactic radiosurgery (SRS) for single brain metastasis by comparing with two mature SRS platforms. Methods Thirteen patients with single brain metastasis treated with CyberKnife (CK) G4 were selected retrospectively. The prescription dose for the planning target volume (PTV) was 18–24 Gy for 1–3 fractions. The PTV volume ranged from 0.44 to 11.52 cc.Treatment plans of thirteen patients were replanned using the ZAP-X plan system and the Gamma Knife (GK) ICON plan system with the same prescription dose and organs at risk (OARs) constraints. The prescription dose of PTV was normalized to 70% for both ZAP-X and CK, while it was 50% for GK. The dosimetric parameters of three groups included the plan characteristics (CI, GI, GSI, beams, MUs, treatment time), PTV (D2, D95, D98, Dmin, Dmean, Coverage), brain tissue (volume of 100%-10% prescription dose irradiation V100%-V10%, Dmean) and other OARs (Dmax, Dmean),all of these were compared and evaluated. All data were read and analyzed with MIM Maestro. One-way ANOVA or a multisample Friedman rank sum test was performed, where p

Abstract The inference of cell–cell communication (CCC) is crucial for a better understanding of complex cellular dynamics and regulatory mechanisms in biological systems. However, accurately inferring spatial CCCs at single-cell resolution remains a significant challenge. To address this issue, we present a versatile method, called DeepTalk, to infer spatial CCC at single-cell resolution by integrating single-cell RNA sequencing (scRNA-seq) data and spatial transcriptomics (ST) data. DeepTalk utilizes graph attention network (GAT) to integrate scRNA-seq and ST data, which enables accurate cell-type identification for single-cell ST data and deconvolution for spot-based ST data. Then, DeepTalk can capture the connections among cells at multiple levels using subgraph-based GAT, and further achieve spatially resolved CCC inference at single-cell resolution. DeepTalk achieves excellent performance in discovering meaningful spatial CCCs on multiple cross-platform datasets, which demonstrates its superior ability to dissect cellular behavior within intricate biological processes.

PurposeTo propose and validate a comprehensive novel metric balancing the registration accuracy and imaging dose for image-guided-radiotherapy based on real patient data.Materials and methodsWith written informed consent and ethical approval, 56 patients were scanned using 6MV CBCT, 140 kV CBCT, and 100 kV CBCT on Halcyon system for three consecutive treatment fractions. Online registration was performed by various on-duty therapists under routine clinical pressure and time limitation. Offline registration was carried out by an experienced physicist without pressure. The consistency between the online and offline results was used as a surrogate of the missing ground-truth of registration accuracy, which was usually developed by introducing ‘known’ setup errors and rescan the phantoms, yet is ethnically not applicable to real patients. The registration differences (ΔD) between various imaging methods and observers were analyzed. The weighted CT dose index (CTDIw) for kV and MV CBCT was acquired using the PTW CTDI head phantom. The weighted-Dose-Accuracy-Product (DAPw) index was defined as DAPw =ΔD(mm) w1* CTDIw(mGy) w2, where w1 and w2 are the weighting factors of accuracy and dose respectively (w1+w2 = 1).ResultsThe mean and interquartile range (IQR) of ΔD decreased monotonically for MV CBCT, 100 kV CBCT, and 140 kV CBCT, supporting the registration consistency as a surrogate metric of image quality. Significant differences of ΔD were observed between the online and offline registration across three imaging methods (P

Purpose: During the radiation treatment planning process, one of the time-consuming procedures is the final high-resolution dose calculation, which obstacles the wide application of the emerging online adaptive radiotherapy techniques (OLART). There is an urgent desire for highly accurate and efficient dose calculation methods. This study aims to develop a dose super resolution-based deep learning model for fast and accurate dose prediction in clinical practice. Method: A Multi-stage Dose Super-Resolution Network (MDSR Net) architecture with sparse masks module and multi-stage progressive dose distribution restoration method were developed to predict high-resolution dose distribution using low-resolution data. A total of 340 VMAT plans from different disease sites were used, among which 240 randomly selected nasopharyngeal, lung, and cervix cases were used for model training, and the remaining 60 cases from the same sites for model benchmark testing, and additional 40 cases from the unseen site (breast and rectum) was used for model generalizability evaluation. The clinical calculated dose with a grid size of 2 mm was used as baseline dose distribution. The input included the dose distribution with 4 mm grid size and CT images. The model performance was compared with HD U-Net and cubic interpolation methods using Dose-volume histograms (DVH) metrics and global gamma analysis with 1%/1 mm and 10% low dose threshold. The correlation between the prediction error and the dose, dose gradient, and CT values was also evaluated. Results: The prediction errors of MDSR were 0.06–0.84% of Dmean indices, and the gamma passing rate was 83.1–91.0% on the benchmark testing dataset, and 0.02–1.03% and 71.3–90.3% for the generalization dataset respectively. The model performance was significantly higher than the HD U-Net and interpolation methods (p

Introduction Breast cancer is one of the most prevalent types of cancer and a leading cause of cancer-related death among females worldwide. Anoikis, a specific type of apoptosis that is triggered by the loss of anchoring between cells and the native extracellular matrix, plays a vital role in cancer invasion and metastasis. However, studies that focus on the prognostic values of anoikis-related genes (ARGs) in breast cancer are scarce. Methods Gene expression data were obtained from The Cancer Genome Atlas (TCGA), Gene Expression Omnibus (GEO), and Molecular Taxonomy of Breast Cancer International Consortium (METABRIC) databases. Five anoikis-related signatures (ARS) were selected from ARGs through univariate Cox regression analysis, LASSO regression analysis, and multivariate Cox regression analysis. Subsequently, an ARGs risk score model was established, and breast cancer patients were divided into high and low risk groups. The correlation between risk groups and overall survival (OS), tumor mutation burden (TMB), tumor microenvironment (TME), stemness, and drug sensitivity were analyzed. Moreover, RT-qPCR was performed to verify the gene expression levels of the five ARS in breast cancer tissues. Furthermore, a nomogram model was constructed based on ARGs risk score and clinicopathological factors. Results A novel ARGs risk score model was constructed based on five ARS (CEMIP, LAMB3, CD24, PTK6, and PLK1), and breast cancer patients were divided into high and low risk groups. Correlation analysis showed that the high and low risk groups had different OS, TMB, TME, stemness, and drug sensitivity. Both the ARGs risk score model and the nomogram showed promising prognosis predictive value in breast cancer. Conclusion ARS could be used as promising biomarkers for breast cancer prognosis predication and treatment options selection.

An accumulation of previous work has established organoids as good preclinical models of human tumors, facilitating translation from basic research to clinical practice. They are changing the paradigm of preclinical cancer research because they can recapitulate the heterogeneity and pathophysiology of human cancers and more closely approximate the complex tissue environment and structure found in clinical tumors than in vitro cell lines and animal models. However, the potential applications of cancer organoids remain to be comprehensively summarized. In the review, we firstly describe what is currently known about cancer organoid culture and then discuss in depth the basic mechanisms, including tumorigenesis and tumor metastasis, and describe recent advances in patient-derived tumor organoids (PDOs) for drug screening and immunological studies. Finally, the present challenges faced by organoid technology in clinical practice and its prospects are discussed. This review highlights that organoids may offer a novel therapeutic strategy for cancer research.

As the genome is organized into a three-dimensional structure in intracellular space, epigenomic information also has a complex spatial arrangement. However, most epigenetic studies describe locations of methylation marks, chromatin accessibility regions, and histone modifications in the horizontal dimension. Proper spatial epigenomic information has rarely been obtained. In this study, we designed spatial chromatin accessibility sequencing (SCA-seq) to resolve the genome conformation by capturing the epigenetic information in single-molecular resolution while simultaneously resolving the genome conformation. Using SCA-seq, we are able to examine the spatial interaction of chromatin accessibility (e.g. enhancer–promoter contacts), CpG island methylation, and spatial insulating functions of the CCCTC-binding factor. We demonstrate that SCA-seq paves the way to explore the mechanism of epigenetic interactions and extends our knowledge in 3D packaging of DNA in the nucleus.

Abstract Background Early detection is critical for improving the survival of breast cancer (BC) patients. Exhaled breath testing as a non-invasive technique might help to improve BC detection. However, the breath test accuracy for BC diagnosis is unclear. Methods This multi-center cohort study consecutively recruited 5047 women from four areas of China who underwent BC screening. Breath samples were collected through standardized breath collection procedures. Volatile organic compound (VOC) markers were identified from a high-throughput breathomics analysis by the high-pressure photon ionization–time-of-flight mass spectrometry (HPPI-TOFMS). Diagnostic models were constructed using the random forest algorithm in the discovery cohort and tested in three external validation cohorts. Results A total of 465 (9.21%) participants were identified with BC. Ten optimal VOC markers were identified to distinguish the breath samples of BC patients from those of non-cancer women. A diagnostic model (BreathBC) consisting of 10 optimal VOC markers showed an area under the curve (AUC) of 0.87 in external validation cohorts. BreathBC-Plus, which combined 10 VOC markers with risk factors, achieved better performance (AUC = 0.94 in the external validation cohorts), superior to that of mammography and ultrasound. Overall, the BreathBC-Plus detection rates were 96.97% for ductal carcinoma in situ, 85.06%, 90.00%, 88.24%, and 100% for stages I, II, III, and IV BC, respectively, with a specificity of 87.70% in the external validation cohorts. Conclusions This is the largest study on breath tests to date. Considering the easy-to-perform procedure and high accuracy, these findings exemplify the potential applicability of breath tests in BC screening.

Objective: Neoadjuvant therapy (NAT) has been widely implemented as an essential treatment to improve therapeutic efficacy in patients with locally-advanced cancer to reduce tumor burden and prolong survival, particularly for human epidermal growth receptor 2-positive and triple-negative breast cancer. The role of peripheral immune components in predicting therapeutic responses has received limited attention. Herein we determined the relationship between dynamic changes in peripheral immune indices and therapeutic responses during NAT administration. Methods: Peripheral immune index data were collected from 134 patients before and after NAT. Logistic regression and machine learning algorithms were applied to the feature selection and model construction processes, respectively. Results: Peripheral immune status with a greater number of CD3+ T cells before and after NAT, and a greater number of CD8+ T cells, fewer CD4+ T cells, and fewer NK cells after NAT was significantly related to a pathological complete response (P < 0.05). The post-NAT NK cell-to-pre-NAT NK cell ratio was negatively correlated with the response to NAT (HR = 0.13, P = 0.008). Based on the results of logistic regression, 14 reliable features (P < 0.05) were selected to construct the machine learning model. The random forest model exhibited the best power to predict efficacy of NAT among 10 machine learning model approaches (AUC = 0.733). Conclusions: Statistically significant relationships between several specific immune indices and the efficacy of NAT were revealed. A random forest model based on dynamic changes in peripheral immune indices showed robust performance in predicting NAT efficacy.

Abstract Although programmed death-ligand 1 (PD-L1) inhibitors have achieved some therapeutic success in breast cancer, their efficacy is limited by low therapeutic response rates, which is closely related to the immune escape of breast cancer cells. Tissue differentiation inducing non-protein coding RNA (TINCR), a long non-coding RNA, as an oncogenic gene associated with the progression of various malignant tumors, including breast cancer; however, the role of TINCR in tumor immunity, especially in breast cancer, remains unclear. We confirmed that TINCR upregulated PD-L1 expression in vivo and in vitro, and promoted the progression of breast cancer. Next, we revealed that TINCR knockdown can significantly improve the therapeutic effect of PD-L1 inhibitors in breast cancer in vivo. Mechanistically, TINCR recruits DNMT1 to promote the methylation of miR-199a-5p loci and inhibit its transcription. Furthermore, in the cytoplasm, TINCR potentially acts as a molecular sponge of miR-199a-5p and upregulates the stability of USP20 mRNA through a competing endogenous RNA (ceRNA) regulatory mechanism, thus promoting PD-L1 expression by decreasing its ubiquitination level. IFN-γ stimulation activates STAT1 by phosphorylation, which migrates into the nucleus to promote TINCR transcription. This is the first study to describe the regulatory role of TINCR in breast cancer tumor immunity, broadening the current paradigm of the functional diversity of TINCR in tumor biology. In addition, our study provides new research directions and potential therapeutic targets for PD-L1 inhibitors in breast cancer.

Abstract Immunotherapy, especially immune checkpoint inhibitors (ICIs), has revolutionized the treatment of many types of cancer, particularly advanced-stage cancers. Nevertheless, although a subset of patients experiences dramatic and long-term disease regression in response to ICIs, most patients do not benefit from these treatments. Some may even experience cancer progression. Immune escape by tumor cells may be a key reason for this low response rate. N 6-methyladenosine (m6A) is the most common type of RNA methylation and has been recognized as a critical regulator of tumors and the immune system. Therefore, m6A modification and related regulators are promising targets for improving the efficacy of tumor immunotherapy. However, the association between m6A modification and tumor immune escape (TIE) has not been comprehensively summarized. Therefore, this review summarizes the existing knowledge regarding m6A modifications involved in TIE and their potential mechanisms of action. Moreover, we provide an overview of currently available agents targeting m6A regulators that have been tested for their elevated effects on TIE. This review establishes the association between m6A modifications and TIE and provides new insights and strategies for maximizing the efficacy of immunotherapy by specifically targeting m6A modifications involved in TIE.

Abstract Background Robotic linac is ideally suited to deliver hypo-fractionated radiotherapy due to its compact head and flexible positioning. The non-coplanar treatment space improves the delivery versatility but the complexity also leads to prolonged optimization and treatment time. Methods In this study, we attempted to use the deep learning (pytorch) framework for the plan optimization of circular cone based robotic radiotherapy. The optimization problem was topologized into a simple feedforward neural network, thus the treatment plan optimization was transformed into network training. With this transformation, the pytorch toolkit with high-efficiency automatic differentiation (AD) for gradient calculation was used as the optimization solver. To improve the treatment efficiency, plans with fewer nodes and beams were sought. The least absolute shrinkage and selection operator (lasso) and the group lasso were employed to address the “sparsity” issue. Results The AD-S (AD sparse) approach was validated on 6 brain and 6 liver cancer cases and the results were compared with the commercial MultiPlan (MLP) system. It was found that the AD-S plans achieved rapid dose fall-off and satisfactory sparing of organs at risk (OARs). Treatment efficiency was improved by the reduction in the number of nodes (28%) and beams (18%), and monitor unit (MU, 24%), respectively. The computational time was shortened to 47.3 s on average. Conclusions In summary, this first attempt of applying deep learning framework to the robotic radiotherapy plan optimization is promising and has the potential to be used clinically.

Summary: Neoadjuvant therapy (NAT) is currently recommended to patients with human epidermal growth factor receptor 2-positive breast cancer (HER2+ BC) that typically exhibit a poor prognosis. The tumor immune microenvironment profoundly affects the efficacy of NAT. However, the correlation between tumor-infiltrating lymphocytes or their specific subpopulations and the response to NAT in HER2+ BC remains largely unknown. In our study, the immune infiltration status of 295 patients was classified as “immune-rich” or “immune-poor” phenotypes. The “immune-rich” phenotype was significantly positively related to pathological complete response (pCR). Ten genes were correlated with both pCR and the immune phenotype based on the results of spline and logistic regression. We constructed a generalized non-linear model combining linear and non-linear gene effects and successfully validated its predictive power using an internal and external validation set (AUC = 0.819, 0.797; respectively) and a clinical set (accuracy = 0.75).

Purpose: To investigate and compare 2 cone-beam computed tomography (CBCT) correction methods for CBCT-based dose calculation. Materials and Methods: Routine CBCT image sets of 12 head and neck cancer patients who received volumetric modulated arc therapy (VMAT) treatment were retrospectively analyzed. The CBCT images obtained using an on-board imager (OBI) at the first treatment fraction were firstly deformable registered and padded with the kVCT images to provide enough anatomical information about the tissues for dose calculation. Then, 2 CBCT correction methods were developed and applied to correct CBCT Hounsfield unit (HU) values. One method (HD method) is based on protocol-specific CBCT HU to physical density (HD) curve, and the other method (HM method) is based on histogram matching (HM) of HU value. The corrected CBCT images (CBCT HD and CBCT HM for HD and HM methods) were imported into the original planning system for dose calculation based on the HD curve of kVCT (the planning CT). The dose computation result was analyzed and discussed to compare these 2 CBCT-correction methods. Results: Dosimetric parameters, such as the D mean , D max and D 5% of the target volume in CBCT plan doses, were higher than those in the kVCT plan doses; however, the deviations were less than 2%. The D 2% , in parallel organs such as the parotid glands, the deviations from the CBCT HM plan dose were less than those of the CBCT HD plan dose. The differences were statistically significant ( P

BackgroundT1–2 breast cancer patients with only one sentinel lymph node (SLN) metastasis have an extremely low non-SLN (NSLN) metastatic rate and are favorable for axillary lymph node dissection (ALND) exemption. This study aimed to construct a nomogram-based preoperative prediction model of NSLN metastasis for such patients, thereby assisting in preoperatively selecting proper surgical procedures.MethodsA total of 729 T1–2 breast cancer patients with only one SLN metastasis undergoing sentinel lymph node biopsy and ALND were retrospectively selected from Harbin Medical University Cancer Hospital between January 2013 and December 2020, followed by random assignment into training (n=467) and validation cohorts (n=262). A nomogram-based prediction model for NSLN metastasis risk was constructed by incorporating the independent predictors of NSLN metastasis identified from multivariate logistic regression analysis in the training cohort. The performance of the nomogram was evaluated by the calibration curve and the receiver operating characteristic (ROC) curve. Finally, decision curve analysis (DCA) was used to determine the clinical utility of the nomogram.ResultsOverall, 160 (21.9%) patients had NSLN metastases. Multivariate analysis in the training cohort revealed that the number of negative SLNs (OR: 0.98), location of primary tumor (OR: 2.34), tumor size (OR: 3.15), and lymph-vascular invasion (OR: 1.61) were independent predictors of NSLN metastasis. The incorporation of four independent predictors into a nomogram-based preoperative estimation of NSLN metastasis demonstrated a satisfactory discriminative capacity, with a C-index and area under the ROC curve of 0.740 and 0.689 in the training and validation cohorts, respectively. The calibration curve showed good agreement between actual and predicted NSLN metastasis risks. Finally, DCA revealed the clinical utility of the nomogram.ConclusionThe nomogram showed a satisfactory discriminative capacity of NSLN metastasis risk in T1–2 breast cancer patients with only one SLN metastasis, and it could be used to preoperatively estimate NSLN metastasis risk, thereby facilitating in precise clinical decision-making on the selective exemption of ALND in such patients.

Background and ObjectiveSentinel lymph node biopsy (SLNB) is used to assess the status of axillary lymph node (ALN), but it causes many adverse reactions. Considering the low rate of sentinel lymph node (SLN) metastasis in T1 breast cancer, this study aims to identify the characteristics of T1 breast cancer without SLN metastasis and to select T1 breast cancer patients who avoid SLNB through constructing a nomogram.MethodsA total of 1,619 T1 breast cancer patients with SLNB in our hospital were enrolled in this study. Through univariate and multivariate logistic regression analysis, we analyzed the tumor anatomical and clinicopathological factors and constructed the Heilongjiang Medical University (HMU) nomogram. We selected the patients exempt from SLNB by using the nomogram.ResultsIn the training cohort of 1,000 cases, the SLN metastasis rate was 23.8%. Tumor volume, swollen axillary lymph nodes, pathological types, and molecular subtypes were found to be independent predictors for SLN metastasis in multivariate regression analysis. Distance from nipple or surface and position of tumor have no effect on SLN metastasis. A regression model based on the results of the multivariate analysis was developed to predict the risk of SLN metastasis, indicating an AUC of 0.798. It showed excellent diagnostic performance (AUC = 0.773) in the validation cohort.ConclusionThe HMU nomogram for predicting SLN metastasis incorporates four variables, including tumor volume, swollen axillary lymph nodes, pathological types, and molecular subtypes. The SLN metastasis rates of intraductal carcinoma and HER2 enriched are 2.05% and 6.67%. These patients could be included in trials investigating the SLNB exemption.

PurposeThe purpose of this study was to evaluate and explore the difference between an atlas-based and deep learning (DL)-based auto-segmentation scheme for organs at risk (OARs) of nasopharyngeal carcinoma cases to provide valuable help for clinical practice.Methods120 nasopharyngeal carcinoma cases were established in the MIM Maestro (atlas) database and trained by a DL-based model (AccuContour®), and another 20 nasopharyngeal carcinoma cases were randomly selected outside the atlas database. The experienced physicians contoured 14 OARs from 20 patients based on the published consensus guidelines, and these were defined as the reference volumes (Vref). Meanwhile, these OARs were auto-contoured using an atlas-based model, a pre-built DL-based model, and an on-site trained DL-based model. These volumes were named Vatlas, VDL-pre-built, and VDL-trained, respectively. The similarities between Vatlas, VDL-pre-built, VDL-trained, and Vref were assessed using the Dice similarity coefficient (DSC), Jaccard coefficient (JAC), maximum Hausdorff distance (HDmax), and deviation of centroid (DC) methods. A one-way ANOVA test was carried out to show the differences (between each two of them).ResultsThe results of the three methods were almost similar for the brainstem and eyes. For inner ears and temporomandibular joints, the results of the pre-built DL-based model are the worst, as well as the results of atlas-based auto-segmentation for the lens. For the segmentation of optic nerves, the trained DL-based model shows the best performance (p < 0.05). For the contouring of the oral cavity, the DSC value of VDL-pre-built is the smallest, and VDL-trained is the most significant (p < 0.05). For the parotid glands, the DSC of Vatlas is the minimum (about 0.80 or so), and VDL-pre-built and VDL-trained are slightly larger (about 0.82 or so). In addition to the oral cavity, parotid glands, and the brainstem, the maximum Hausdorff distances of the other organs are below 0.5 cm using the trained DL-based segmentation model. The trained DL-based segmentation method behaves well in the contouring of all the organs that the maximum average deviation of the centroid is no more than 0.3 cm.ConclusionThe trained DL-based segmentation performs significantly better than atlas-based segmentation for nasopharyngeal carcinoma, especially for the OARs with small volumes. Although some delineation results still need further modification, auto-segmentation methods improve the work efficiency and provide a level of help for clinical work.

PurposeConsistent training and testing datasets can lead to good performance for deep learning (DL) models. However, a large high-quality training dataset for unusual clinical scenarios is usually not easy to collect. The work aims to find optimal training data collection strategies for DL-based dose prediction models.Materials and MethodsA total of 325 clinically approved cervical IMRT plans were utilized. We designed comparison experiments to investigate the impact of (1) beam angles, (2) the number of beams, and (3) patient position for DL dose prediction models. In addition, a novel geometry-based beam mask generation method was proposed to provide beam setting information in the model training process. What is more, we proposed a new training strategy named “full-database pre-trained strategy”.ResultsThe model trained with a homogeneous dataset with the same beam settings achieved the best performance [mean prediction errors of planning target volume (PTV), bladder, and rectum: 0.29 ± 0.15%, 3.1 ± 2.55%, and 3.15 ± 1.69%] compared with that trained with large mixed beam setting plans (mean errors of PTV, bladder, and rectum: 0.8 ± 0.14%, 5.03 ± 2.2%, and 4.45 ± 1.4%). A homogeneous dataset is more accessible to train an accurate dose prediction model (mean errors of PTV, bladder and rectum: 2.2 ± 0.15%, 5 ± 2.1%, and 3.23 ± 1.53%) than a non-homogeneous one (mean errors of PTV, bladder and rectum: 2.55 ± 0.12%, 6.33 ± 2.46%, and 4.76 ± 2.91%) without other processing approaches. The added beam mask can constantly improve the model performance, especially for datasets with different beam settings (mean errors of PTV, bladder, and rectum improved from 0.8 ± 0.14%, 5.03 ± 2.2%, and 4.45 ± 1.4% to 0.29 ± 0.15%, 3.1 ± 2.55%, and 3.15 ± 1.69%).ConclusionsA consistent dataset is recommended to form a patient-specific IMRT dose prediction model. When a consistent dataset is not accessible to collect, a large dataset with different beam angles and a training model with beam information can also get a relatively good model. The full-database pre-trained strategies can rapidly form an accuracy model from a pre-trained model. The proposed beam mask can effectively improve the model performance. Our study may be helpful for further dose prediction studies in terms of training strategies or database establishment.