Automated Segmentation of Left Ventricular Myocardium on Cardiac Computed Tomography Using Deep Learning

Automated cardiac MR segmentation enables accurate and reproducible ventricular function assessment in Tetralogy of Fallot (ToF), whereas manual segmentation remains time-consuming and variable. To evaluate the deep learning (DL)-based models for automatic left ventricle (LV), right ventricle (RV), and LV myocardium segmentation in ToF, compared with manual reference standard annotations. Retrospective. 427 patients with diverse cardiac conditions (305 non-ToF, 122 ToF), with 395 for training/validation, 32 ToF for internal testing, and 12 external ToF for generalizability assessment. Steady-state free precession cine sequence at 1.5/3 T. U-Net, Deep U-Net, and MultiResUNet were trained under three regimes (non-ToF, ToF-only, mixed), using manual segmentations from one radiologist and one researcher (20 and 10 years of experience, respectively) as reference, with consensus for discrepancies. Performance for LV, RV, and LV myocardium was evaluated using Dice Similarity Coefficient (DSC), Intersection over Union (IoU), and F1-score, alongside regional (basal, middle, apical) and global ventricular function comparisons to manual results. Friedman tests were applied for architecture and regime comparisons, paired Wilcoxon tests for ED-ES differences, and Pearson's r for assessing agreement in global function. MultiResUNet model trained on a mixed dataset (TOF and non-TOF cases) achieved the best segmentation performance, with DSCs of 96.1% for LV and 93.5% for RV. In the internal test set, DSCs for LV, RV, and LV myocardium were 97.3%, 94.7%, and 90.7% at end-diastole, and 93.6%, 92.1%, and 87.8% at end-systole, with ventricular measurement correlations ranging from 0.84 to 0.99. Regional analysis showed LV DSCs of 96.3% (basal), 96.4% (middle), and 94.1% (apical), and RV DSCs of 92.8%, 94.2%, and 89.6%. External validation (n = 12) showed correlations ranging from 0.81 to 0.98. The MultiResUNet model enabled accurate automated cardiac MRI segmentation in ToF with the potential to streamline workflows and improve disease monitoring. 3. Stage 2.

Automated segmentation of the healed anterior cruciate ligament from T2 * relaxometry MRI scans.

Develop and evaluate a deep learning-based automatic meningioma segmentation method for preoperative meningioma differentiation using radiomic features. A retrospective multicentre inclusion of MR examinations (T1/T2-weighted and contrast-enhanced T1-weighted imaging) was conducted. Data from centre 1 were allocated to training (n = 307, age = 50.94 ± 11.51) and internal testing (n = 238, age = 50.70 ± 12.72) cohorts, and data from centre 2 external testing cohort (n = 64, age = 48.45 ± 13.59). A modified attention U-Net was trained for meningioma segmentation. Segmentation accuracy was evaluated by five quantitative metrics. The agreement between radiomic features from manual and automatic segmentations was assessed using intra class correlation coefficient (ICC). After univariate and minimum-redundancy-maximum-relevance feature selection, L1-regularized logistic regression models for differentiating between low-grade (I) and high-grade (II and III) meningiomas were separately constructed using manual and automatic segmentations; their performances were evaluated using ROC analysis. Dice of meningioma segmentation for the internal testing cohort were 0.94 ± 0.04 and 0.91 ± 0.05 for tumour volumes in contrast-enhanced T1-weighted and T2-weighted images, respectively; those for the external testing cohort were 0.90 ± 0.07 and 0.88 ± 0.07. Features extracted using manual and automatic segmentations agreed well, for both the internal (ICC = 0.94, interquartile range: 0.88-0.97) and external (ICC = 0.90, interquartile range: 0.78-70.96) testing cohorts. AUC of radiomic model with automatic segmentation was comparable with that of the model with manual segmentation for both the internal (0.95 vs. 0.93, p = 0.176) and external (0.88 vs. 0.91, p = 0.419) testing cohorts. The developed deep learning-based segmentation method enables automatic and accurate extraction of meningioma from multiparametric MR images and can help deploy radiomics for preoperative meningioma differentiation in clinical practice. • A deep learning-based method was developed for automatic segmentation of meningioma from multiparametric MR images. • The automatic segmentation method enabled accurate extraction of meningiomas and yielded radiomic features that were highly consistent with those that were obtained using manual segmentation. • High-grade meningiomas were preoperatively differentiated from low-grade meningiomas using a radiomic model constructed on features from automatic segmentation.

Introduction: Peripheral artery disease (PAD) is characterized by atherosclerosis of lower extremity arteries that promotes reduced muscle perfusion and skeletal muscle atrophy, which contributes to functional impairment in PAD patients. Computed tomography (CT) imaging offers a non-invasive method for quantifying peripheral muscle characteristics; however, manual image segmentation of muscles is time-consuming, and regional muscle analysis remains understudied in PAD. Therefore, we sought to develop and validate a deep learning approach for calf muscle segmentation to assess serial regional changes in muscle density of PAD patients over an 18-month study period. Methods: Patients with PAD (n=89) were prospectively enrolled for lower extremity non-contrast CT imaging. A subset of patients (n=45) was recruited for an additional follow-up CT scan 18 months later. Calf muscle groups, including the gastrocnemius, soleus, and tibialis anterior, were manually segmented from axial CT images of each patient’s symptomatic limb, and mean muscle densities for each muscle were quantified based on CT Hounsfield units. Manual segmentations were used to train an nnU-Net deep learning model. Data augmentation techniques were performed to enhance generalization. The dataset was split into an 80/20 ratio for the training and test sets. Dice coefficients were calculated to evaluate the overlap/agreement between manual and deep learning segmentations. Paired t-tests were performed to assess the differences in muscle densities between the baseline and 18-month follow-up measures. Results: The deep learning model achieved high segmentation performance, with dice scores of 0.90 ± 0.02 for the gastrocnemius and soleus, and 0.89 ± 0.02 for the tibialis anterior. Deep learning-guided serial CT image analysis detected a significant reduction in muscle densities across all muscle groups at 18-month follow-up when compared to baseline measurements (p<0.05). Conclusion: Deep learning analysis can enable rapid, regional analysis of lower extremity skeletal muscle characteristics, reducing the time for muscle-by-muscle analysis in PAD patients from hours to seconds. Future implementation of regional muscle analysis in PAD may assist vascular medicine specialists with identifying regional muscle wasting that is related to patient walking impairment and/or symptoms as well as aid in serial monitoring of PAD patients following revascularization or supervised exercise for claudication.

Simple SummaryUterine cervical cancer (CC) is a leading cause of cancer-related deaths in women worldwide. Pelvic magnetic resonance imaging (MRI) allows the assessment of local tumor extent and guides the choice of primary treatment. MRI tumor segmentation enables whole-volume radiomic tumor profiling, which is potentially useful for prognostication and individualization of therapy in CC. Manual tumor segmentation is, however, labor intensive and thus not part of routine clinical workflow. In the current work, we trained a deep learning (DL) algorithm to automatically segment the primary tumor in CC patients. Although the achieved segmentation performance of the trained DL algorithm is slightly lower than that for human experts, it is still relatively good. This study suggests that automated MRI primary tumor segmentations by DL algorithms without any human interaction is possible in patients with CC.Uterine cervical cancer (CC) is the most common gynecologic malignancy worldwide. Whole-volume radiomic profiling from pelvic MRI may yield prognostic markers for tailoring treatment in CC. However, radiomic profiling relies on manual tumor segmentation which is unfeasible in the clinic. We present a fully automatic method for the 3D segmentation of primary CC lesions using state-of-the-art deep learning (DL) techniques. In 131 CC patients, the primary tumor was manually segmented on T2-weighted MRI by two radiologists (R1, R2). Patients were separated into a train/validation (n = 105) and a test- (n = 26) cohort. The segmentation performance of the DL algorithm compared with R1/R2 was assessed with Dice coefficients (DSCs) and Hausdorff distances (HDs) in the test cohort. The trained DL network retrieved whole-volume tumor segmentations yielding median DSCs of 0.60 and 0.58 for DL compared with R1 (DL-R1) and R2 (DL-R2), respectively, whereas DSC for R1-R2 was 0.78. Agreement for primary tumor volumes was excellent between raters (R1-R2: intraclass correlation coefficient (ICC) = 0.93), but lower for the DL algorithm and the raters (DL-R1: ICC = 0.43; DL-R2: ICC = 0.44). The developed DL algorithm enables the automated estimation of tumor size and primary CC tumor segmentation. However, segmentation agreement between raters is better than that between DL algorithm and raters.

In the development of treatments for cardiovascular diseases, short axis cardiac cine MRI is important for the assessment of various structural and functional properties of the heart. In short axis cardiac cine MRI, Cardiac properties including the ventricle dimensions, stroke volume, and ejection fraction can be extracted based on accurate segmentation of the left ventricle (LV) myocardium. One of the most advanced segmentation methods is based on fully convolutional neural networks (FCN) and can be successfully used to do segmentation in cardiac cine MRI slices. However, the temporal dependency between slices acquired at neighboring time points is not used. Here, based on our previously proposed FCN structure, we proposed a new algorithm to segment LV myocardium in porcine short axis cardiac cine MRI by incorporating convolutional long short-term memory (Conv-LSTM) to leverage the temporal dependency. In this approach, instead of processing each slice independently in a conventional CNN-based approach, the Conv-LSTM architecture captures the dynamics of cardiac motion over time. In a leave-one-out experiment on 8 porcine specimens (3,600 slices), the proposed approach was shown to be promising by achieving average mean Dice similarity coefficient (DSC) of 0.84, Hausdorff distance (HD) of 6.35 mm, and average perpendicular distance (APD) of 1.09 mm when compared with manual segmentations, which improved the performance of our previous FCN-based approach (average mean DSCe0.84, HDe6.78 mm, and APDe1.11 mm). Qualitatively, our model showed robustness against low image quality and complications in the surrounding anatomy due to its ability to capture the dynamics of cardiac motion.

To develop and validate a deep learning-based algorithm for segmenting and quantifying the physiological and diseased aorta in computed tomography angiographies. CTA exams of the aorta of 191 patients (68.1 ± 14 years, 128 male), performed between 2015 and 2018, were retrospectively identified from our imaging archive and manually segmented by two investigators. A 3D U-Net model was trained on the data, which was divided into a training, a validation, and a test group at a ratio of 7:1:2. Cases in the test group (n = 41) were evaluated to compare manual and automatic segmentations. Dice similarity coefficient (DSC), mean surface distance (MSD), and Hausdorff surface distance (HSD) were extracted. Maximum diameter, effective diameter, and area were quantified and compared between both segmentations at eight anatomical landmarks, and at the maximum area of an aneurysms if present (n = 14). Statistics included error calculation, intraclass correlation coefficient, and Bland-Altman analysis. A DSC of 0.95 [0.94; 0.95] and an MSD of 0.76 [0.06; 0.99] indicated close agreement between segmentations. HSD was 8.00 [4.47; 10.00]. The largest absolute errors were found in the ascending aorta with 0.8 ± 1.5 mm for maximum diameter and at the coeliac trunk with - 30.0 ± 81.6 mm2 for area. Results for absolute errors in aneurysms were - 0.5 ± 2.3 mm for maximum diameter, 0.3 ± 1.6 mm for effective diameter, and 64.9 ± 114.9 mm2 for area. ICC showed excellent agreement (> 0.9; p < 0.05) between quantitative measurements. Automated segmentation of the aorta on CTA data using a deep learning algorithm is feasible and allows for accurate quantification of the aortic lumen even if the vascular architecture is altered by disease. • A deep learning-based algorithm can automatically segment the aorta, mostly within acceptable margins of error, even if the vascular architecture is altered by disease. • Quantifications performed in the segmentations were mostly within clinically acceptable limits, even in pathologically altered segments of the aorta.

Three-dimensional (3D) prostate image segmentation is useful for cancer diagnosis and therapy guidance, but can be time-consuming to perform manually and involves varying levels of difficulty and interoperator variability within the prostatic base, midgland (MG), and apex. In this study, the authors measured accuracy and interobserver variability in the segmentation of the prostate on T2-weighted endorectal magnetic resonance (MR) imaging within the whole gland (WG), and separately within the apex, midgland, and base regions. The authors collected MR images from 42 prostate cancer patients. Prostate border delineation was performed manually by one observer on all images and by two other observers on a subset of ten images. The authors used complementary boundary-, region-, and volume-based metrics [mean absolute distance (MAD), Dice similarity coefficient (DSC), recall rate, precision rate, and volume difference (ΔV)] to elucidate the different types of segmentation errors that they observed. Evaluation for expert manual and semiautomatic segmentation approaches was carried out. Compared to manual segmentation, the authors' semiautomatic approach reduces the necessary user interaction by only requiring an indication of the anteroposterior orientation of the prostate and the selection of prostate center points on the apex, base, and midgland slices. Based on these inputs, the algorithm identifies candidate prostate boundary points using learned boundary appearance characteristics and performs regularization based on learned prostate shape information. The semiautomated algorithm required an average of 30 s of user interaction time (measured for nine operators) for each 3D prostate segmentation. The authors compared the segmentations from this method to manual segmentations in a single-operator (mean whole gland MAD = 2.0 mm, DSC = 82%, recall = 77%, precision = 88%, and ΔV = - 4.6 cm(3)) and multioperator study (mean whole gland MAD = 2.2 mm, DSC = 77%, recall = 72%, precision = 86%, and ΔV = - 4.0 cm(3)). These results compared favorably with observed differences between manual segmentations and a simultaneous truth and performance level estimation reference for this data set (whole gland differences as high as MAD = 3.1 mm, DSC = 78%, recall = 66%, precision = 77%, and ΔV = 15.5 cm(3)). The authors found that overall, midgland segmentation was more accurate and repeatable than the segmentation of the apex and base, with the base posing the greatest challenge. The main conclusions of this study were that (1) the semiautomated approach reduced interobserver segmentation variability; (2) the segmentation accuracy of the semiautomated approach, as well as the accuracies of recently published methods from other groups, were within the range of observed expert variability in manual prostate segmentation; and (3) further efforts in the development of computer-assisted segmentation would be most productive if focused on improvement of segmentation accuracy and reduction of variability within the prostatic apex and base.

BackgroundThe evaluation of automatic segmentation algorithms is commonly performed using geometric metrics. An analysis based on dosimetric parameters might be more relevant in clinical practice but is often lacking in the literature. The aim of this study was to investigate the impact of state-of-the-art 3D U-Net-generated organ delineations on dose optimization in radiation therapy (RT) for prostate cancer patients.MethodsA database of 69 computed tomography images with prostate, bladder, and rectum delineations was used for single-label 3D U-Net training with dice similarity coefficient (DSC)-based loss. Volumetric modulated arc therapy (VMAT) plans have been generated for both manual and automatic segmentations with the same optimization settings. These were chosen to give consistent plans when applying perturbations to the manual segmentations. Contours were evaluated in terms of DSC, average and 95% Hausdorff distance (HD). Dose distributions were evaluated with the manual segmentation as reference using dose volume histogram (DVH) parameters and a 3%/3 mm gamma-criterion with 10% dose cut-off. A Pearson correlation coefficient between DSC and dosimetric metrics, i.e. gamma index and DVH parameters, has been calculated.Results3D U-Net-based segmentation achieved a DSC of 0.87 (0.03) for prostate, 0.97 (0.01) for bladder and 0.89 (0.04) for rectum. The mean and 95% HD were below 1.6 (0.4) and below 5 (4) mm, respectively. The DVH parameters, V_{60/65/70,{mathrm{Gy}}} for the bladder and V_{50/65/70,{mathrm{Gy}}} for the rectum, showed agreement between dose distributions within pm , 5% and pm ,2%, respectively. The D_{98/2%} and V_{95%}, for prostate and its 3 mm expansion (surrogate clinical target volume) showed agreement with the reference dose distribution within 2% and 3 Gy with the exception of one case. The average gamma pass-rate was 85%. The comparison between geometric and dosimetric metrics showed no strong statistically significant correlation.ConclusionsThe 3D U-Net developed for this work achieved state-of-the-art geometrical performance. Analysis based on clinically relevant DVH parameters of VMAT plans demonstrated neither excessive dose increase to OARs nor substantial under/over-dosage of the target in all but one case. Yet the gamma analysis indicated several cases with low pass rates. The study highlighted the importance of adding dosimetric analysis to the standard geometric evaluation.

Cardiac MRI has experienced a crescent relevance in clinical investigations. The segmentation of myocardial walls is a prerequisite for assessment of cardiac viability. Manual or semi-automatic segmentation of all the images of a subject is tedious, as well as consuming much time from cardiologists. In this study, we selected 23 slices of simulated cardiac MR by MRXCAT and 30 real slices of CINE-MR from 15 patients with Chagas Disease. The proposed pipeline of the fully automatic segmentation consists of three steps: 1. Preprocessing; 2. Automatic Seeds Definition; and 3. Segmentation by Geodesic Active Contour. An experienced cardiologist provided the gold standard annotations of apical, mid-ventricular and basal LV myocardium. We use the following three metrics to validate the proposed pipeline with different signal to noise ratio: Dice similarity (DS), Precision (Pr) and Volumetric Similarity (VS). DS show good agreement between manual segmentation and the automatic segmentation in simulated images with SNR 200, 25, 15 and 5, i.e., 0.98, 0.93, 0.9 and 0.72, respectively. We found moderate agreements between manual segmentation and Snake segmentation in simulated images with SNR 200, 25, 15 and 5, i.e., 0.38, 0.42, 0.34 and 0.39, respectively. The DS, VS, and Pr obtained suggest substantial agreements between the manual and our proposed method segmentation in images of Chagas’s Disease, i.e., 0.8 [0.69–0.87], 0.89 [0.72–0.99] and 0.9 [0.76–0.98] (mean [min–max]), respectively. Our findings suggest that one can use the proposed method in the automatic myocardium segmentation with reliability similar to manual tracing, although completely free of human interaction.

Manual segmentation in lumbar paraspinal muscle studies is time-consuming, subject to variability, and limits comprehensive evaluation. To solve the problems, we aim to develop a deep learning algorithm for the automatic segmentation and multi-parameter quantification of eight paraspinal muscles in lumbar CT. We collected CT images covering the range from the superior endplate of the L1 vertebra to the inferior endplate of the S1 vertebra. The dataset was partitioned into training, validation, and test sets in a ratio of 7:1:2. Six convolutional neural networks were employed to automatically segment the bilateral psoas major, quadratus lumborum, erector spinae, and multifidus across all slices. Model performance was evaluated using the Dice Similarity Coefficient (DSC), Hausdorff Distance (HD), and mean Intersection over Union (mIoU). Muscle parameters including cross-sectional area, muscle volume, fat infiltration, CT density, and paraspinal muscle index were calculated. The Intraclass Correlation Coefficient [ICC(3,1)] was used to assess agreement between model's segmentations and manual ground truth. 100 lumbar CT scans (age, 55.02 ± 16.2; 62 female) were used for model development. The mean DSC across all six models ranged from 0.754 to 0.903. TransUNet demonstrated the best overall (DSC, 0.903; 95% CI: 0.902, 0.913) (HD, 9.934mm; 95% CI: 9.815, 11.902) (mIoU, 0.841; 95% CI: 0.839, 0.850). For individual muscles segmented by TransUNet, the highest DSC was 0.936 (95% CI: 0.934-0.938) for the right multifidus. The mean ICC for all parameters was 0.931. The highest ICC value was 0.999 (p < 0.001). The deep learning tool enables automatic and accurate segmentation of eight lumbar paraspinal muscles. Muscle parameter measurements derived from model's segmentations and manual ground truth exhibit high agreement. Accurate segmentation and quantification of paraspinal muscle parameters facilitates large-scale epidemiological studies of paraspinal muscles and spine-related diseases, overcoming the limitations in manual methods.

Automatic segmentation of the choroid on optical coherence tomography (OCT) images helps ophthalmologists in diagnosing eye pathologies. Compared to manual segmentations, it is faster and is not affected by human errors. The presence of the large speckle noise in the OCT images limits the automatic segmentation and interpretation of them. To solve this problem, a new curvelet transform-based K-SVD method is proposed in this study. Furthermore, the dataset was manually segmented by a retinal ophthalmologist to draw a comparison with the proposed automatic segmentation technique. In this study, curvelet transform-based K-SVD dictionary learning and Lucy-Richardson algorithm were used to remove the speckle noise from OCT images. The Outer/Inner Choroidal Boundaries (O/ICB) were determined utilizing graph theory. The area between ICB and outer choroidal boundary was considered as the choroidal region. The proposed method was evaluated on our dataset and the average dice similarity coefficient (DSC) was calculated to be 92.14% ± 3.30% between automatic and manual segmented regions. Moreover, by applying the latest presented open-source algorithm by Mazzaferri et al. on our dataset, the mean DSC was calculated to be 55.75% ± 14.54%. A significant similarity was observed between automatic and manual segmentations. Automatic segmentation of the choroidal layer could be also utilized in large-scale quantitative studies of the choroid.

PurposeAdaptive optics scanning laser ophthalmoscope (AOSLO) capillary perfusion images can possess large variations in contrast, intensity, and background signal, thereby limiting the use of global or adaptive thresholding techniques for automatic segmentation. We sought to develop an automated approach to segment perfused capillaries in AOSLO images.Methods12,979 image patches were extracted from manually segmented AOSLO montages from 14 eyes and used to train a convolutional neural network (CNN) that classified pixels as capillaries, large vessels, background, or image canvas. 1764 patches were extracted from AOSLO montages of four separate subjects, and were segmented manually by two raters (ground truth) and automatically by the CNN, an Otsu's approach, and a Frangi approach. A modified Dice coefficient was created to account for slight spatial differences between the same manually and CNN-segmented capillaries.ResultsCNN capillary segmentation had an accuracy (0.94), a Dice coefficient (0.67), and a modified Dice coefficient (0.90) that were significantly higher than other automated approaches (P < 0.05). There were no significant differences in capillary density and mean segment length between manual ground-truth and CNN segmentations (P > 0.05).ConclusionsClose agreement between the CNN and manual segmentations enables robust and objective quantification of perfused capillary metrics. The developed CNN is time and computationally efficient, and distinguishes capillaries from areas containing diffuse background signal and larger underlying vessels.Translational RelevanceThis automatic segmentation algorithm greatly increases the efficiency of quantifying AOSLO capillary perfusion images.

To compare choroidal neovascularization (CNV) area and vessel density (VD) measurements between two different swept-source optical coherence tomography angiography (SS-OCTA) devices. En face optical coherence tomography angiography (OCTA) images of patients affected by neovascular age-related macular degeneration were collected prospectively from two devices: Zeiss PLEX Elite 9000 (Carl Zeiss Meditec, Dublin, CA) and Topcon DRI OCT Triton SS-OCTA (Topcon, Tokyo, Japan). Choroidal neovascularization area and VD of images were measured and analyzed with ImageJ software by two readers to evaluate the agreement between two devices, with respect to different image size (3 × 3 and 6 × 6 mm) and different image segmentation (automatic vs. manual), and a Topcon equivalent Zeiss segmentation as control (i.e., the equivalent anatomical slab given by Topcon device on the Zeiss device). A total of 30 eyes (30 patients) were analyzed. There was an excellent agreement between the two readers in CNV area measurements intraclass correlation coefficient >0.9 in all analyses. We found excellent agreement in CNV area measurements (manual and automatic segmentations) when comparing 3 × 3-mm or 6 × 6-mm images both for each single device and between the two devices (overall intraclass correlation coefficient > 0.9). Vessel density measurements between manual to automatic segmentation within the same device and same image size had a high intraclass correlation coefficient value, but there was a poor agreement in VD between different image sizes (3 × 3 mm vs. 6 × 6 mm) in the same device and also comparing the two devices (3 × 3 Topcon vs. 3 × 3 Zeiss; 6 × 6 Topcon vs. 6 × 6 Zeiss). There was a poor agreement between the Topcon equivalent Zeiss segmentation and all other segmentations. There was an excellent agreement in CNV area measurements for both swept-source optical coherence tomography angiography devices in automatic and manual segmentations. However, the Topcon equivalent Zeiss segmentation was not comparable with any of the preset segmentations of Topcon and Zeiss devices. There was a poor agreement in CNV VD between different image size and different devices. For these reasons, it seems that, for accurate longitudinal analysis of VD, it is better to use the same device for each individual, even if both devices can be used interchangeably for CNV area measurements using automatic or manual segmentations.

BackgroundNasopharyngeal carcinoma (NPC), an endemic malignancy in Southeast Asia and southern China, necessitates precise delineation of the gross tumor volume (GTV) for radiotherapy to optimize patient prognosis. Positron emission tomography (PET) imaging offers valuable metabolic insights to guide radiotherapy planning but is hampered by manual GTV segmentation, which is labor-intensive and prone to inter- and intra-observer variability. Conventional segmentation methods are inherently limited, and comparative evaluations of deep learning (DL) frameworks for PET-based primary NPC lesion segmentation remain scarce. Thus, this study aimed to optimize the delineation of primary NPC radiotherapy target volumes by leveraging PET-image-based DL segmentation technology, with the goal of enhancing both the precision and efficiency of lesion delineation.MethodsResearchers retrospectively collected PET imaging data from 212 NPC patients at the Affiliated Hospital of Guangdong Medical University. The patients were randomly divided into training (170 patients) and testing (42 patients) datasets. A radiation oncologist and a nuclear medicine physician collaboratively delineated the cancer lesion boundaries through consensus. After data preprocessing, three DL models (Res-Unet, Nn-Unet, and Nn-Former) were used to automatically segment the lesions. Training was based on the training dataset. During evaluation, the models’ segmentation performance was comprehensively assessed using Dice similarity coefficient (DSC) and Hausdorff distance (HD). A visual analysis of the results was also conducted to intuitively understand the models’ segmentation capabilities.ResultsIn the training set, Nn-Unet achieved the highest DSC of 0.869, whereas in the testing set, its DSC was 0.833. The DSC for Nn-Former hovered around 0.8 in both the training and testing sets. In contrast, Res-Unet demonstrated the lowest DSC values among the three models, particularly in the testing set (DSC =0.794). Statistical analysis revealed that the DSC of Nn-Unet was significantly higher than that of both Res-Unet and Nn-Former in the training set (P<0.01 for both comparisons). Regarding the 95% HD and average surface distance (ASD) values, Nn-Unet outperformed the other two models in both the training and testing sets. However, Res-Unet exhibited a much higher HD value in the testing set compared to the training set. The Loss curves for all models gradually decreased as training progressed, indicating that the models were learning relevant features of NPC lesions. The final Loss value for Res-Unet was approximately 0.7, whereas Nn-Unet and Nn-Former had final Loss values below 0.8.ConclusionsThe PET-image-based automatic segmentation model for the primary tumor of NPC established in this study demonstrates the clinical potential of Nn-Unet in primary lesion of NPC segmentation tasks.