Image Segmentation Model Research Articles

GEA-MSNet: A Novel Model for Segmenting Remote Sensing Images of Lakes Based on the Global Efficient Attention Module and Multi-Scale Feature Extraction

The decrease in lake area has garnered significant attention within the global ecological community, prompting extensive research in remote sensing and computer vision to accurately segment lake areas from satellite images. However, existing image segmentation models suffer from poor generalization performance, the imprecise depiction of water body edges, and the inadequate inclusion of water body segmentation information. To address these limitations and improve the accuracy of water body segmentation in remote sensing images, we propose a novel GEA-MSNet segmentation model. Our model incorporates a global efficient attention module (GEA) and multi-scale feature fusion to enhance the precision of water body delineation. By emphasizing global semantic information, our GEA-MSNet effectively learns image features from remote sensing data, enabling the accurate detection and segmentation of water bodies. This study makes three key contributions: firstly, we introduce the GEA module within the encode framework to aggregate shallow feature semantics for the improved classification accuracy of lake pixels; secondly, we employ a multi-scale feature fusion structure during decoding to expand the acceptance domain for feature extraction while prioritizing water body features in images; thirdly, extensive experiments are conducted on both scene classification datasets and Tibetan Plateau lake datasets with ablation experiments validating the effectiveness of our proposed GEA module and multi-scale feature fusion structure. Ultimately, our GEA-MSNet model demonstrates exceptional performance across multiple datasets with an average intersection ratio union (mIoU) improved to 75.49%, recall enhanced to 83.79%, pixel accuracy (PA) reaching 90.21%, and the f1-score significantly elevated to 83.25%.

From CNN to Transformer: A Review of Medical Image Segmentation Models.

Medical image segmentation is an important step in medical image analysis, especially as a crucial prerequisite for efficient disease diagnosis and treatment. The use of deep learning for image segmentation has become a prevalent trend. The widely adopted approach currently is U-Net and its variants. Moreover, with the remarkable success of pre-trained models in natural language processing tasks, transformer-based models like TransUNet have achieved desirable performance on multiple medical image segmentation datasets. Recently, the Segment Anything Model (SAM) and its variants have also been attempted for medical image segmentation. In this paper, we conduct a survey of the most representative seven medical image segmentation models in recent years. We theoretically analyze the characteristics of these models and quantitatively evaluate their performance on Tuberculosis Chest X-rays, Ovarian Tumors, and Liver Segmentation datasets. Finally, we discuss the main challenges and future trends in medical image segmentation. Our work can assist researchers in the related field to quickly establish medical segmentation models tailored to specific regions.

FCSN: Global Context Aware Segmentation by Learning the Fourier Coefficients of Objects in Medical Images.

The encoder-decoder model is a commonly used Deep learning (DL) model for medical image segmentation. Encoder-decoder models make pixel-wise predictions that focus heavily on local patterns. As a result, the predicted mask often fails to preserve the object's shape and topology, which requires an understanding of the image's global context. In this work, we propose a Fourier Coefficient Segmentation Network (FCSN)---a novel global context-aware DL model that segments an object by learning the complex Fourier Coefficients of the object's masks. The Fourier coefficients are calculated by integrating over the mask's contour. Hence, FCSN is naturally motivated to incorporate a broader image context when estimating the coefficients. The global context awareness of FCSN helps produce more accurate segmentation and is more robust to local perturbations, such as additive noise or motion blur. We compare FCSN on other state-of-the-art global context-aware models (UNet++, DeepLabV3+, UNETR) on 5 medical image segmentation tasks (ISIC_2018, RIM_CUP, RIM_DISC, PROSTATE, FETAL). When compared with UNETR, FCSN attains significantly lower Hausdorff scores with 19.14 (6%), 17.42 (6%), 9.16 (14%), 11.18 (22%), and 5.98 (6%) for ISIC_2018, RIM_CUP, RIM_DISC, PROSTATE, and FETAL tasks respectively. Moreover, FCSN is lightweight by discarding the decoder module. FCSN only requires 29.7 M parameters which are 75.6 M and 9.9 M fewer than UNETR and DeepLabV3+, respectively. FCSN attains inference/training speeds of 1.6 ms/img and 6.3 ms/img, which is 8 and 3 faster than UNet and UNETR. Our work is available at https://github.com/nus-morninlab/FCSN.

Automated segmentation of lesions and organs at risk on [68Ga]Ga-PSMA-11 PET/CT images using self-supervised learning with Swin UNETR

BackgroundProstate-specific membrane antigen (PSMA) PET/CT imaging is widely used for quantitative image analysis, especially in radioligand therapy (RLT) for metastatic castration-resistant prostate cancer (mCRPC). Unknown features influencing PSMA biodistribution can be explored by analyzing segmented organs at risk (OAR) and lesions. Manual segmentation is time-consuming and labor-intensive, so automated segmentation methods are desirable. Training deep-learning segmentation models is challenging due to the scarcity of high-quality annotated images. Addressing this, we developed shifted windows UNEt TRansformers (Swin UNETR) for fully automated segmentation. Within a self-supervised framework, the model’s encoder was pre-trained on unlabeled data. The entire model was fine-tuned, including its decoder, using labeled data.MethodsIn this work, 752 whole-body [68Ga]Ga-PSMA-11 PET/CT images were collected from two centers. For self-supervised model pre-training, 652 unlabeled images were employed. The remaining 100 images were manually labeled for supervised training. In the supervised training phase, 5-fold cross-validation was used with 64 images for model training and 16 for validation, from one center. For testing, 20 hold-out images, evenly distributed between two centers, were used. Image segmentation and quantification metrics were evaluated on the test set compared to the ground-truth segmentation conducted by a nuclear medicine physician.ResultsThe model generates high-quality OARs and lesion segmentation in lesion-positive cases, including mCRPC. The results show that self-supervised pre-training significantly improved the average dice similarity coefficient (DSC) for all classes by about 3%. Compared to nnU-Net, a well-established model in medical image segmentation, our approach outperformed with a 5% higher DSC. This improvement was attributed to our model’s combined use of self-supervised pre-training and supervised fine-tuning, specifically when applied to PET/CT input. Our best model had the lowest DSC for lesions at 0.68 and the highest for liver at 0.95.ConclusionsWe developed a state-of-the-art neural network using self-supervised pre-training on whole-body [68Ga]Ga-PSMA-11 PET/CT images, followed by fine-tuning on a limited set of annotated images. The model generates high-quality OARs and lesion segmentation for PSMA image analysis. The generalizable model holds potential for various clinical applications, including enhanced RLT and patient-specific internal dosimetry.

SC-Unext: A Lightweight Image Segmentation Model with Cellular Mechanism for Breast Ultrasound Tumor Diagnosis.

Automatic breast ultrasound image segmentation plays an important role in medical image processing. However, current methods for breast ultrasound segmentation suffer from high computational complexity and large model parameters, particularly when dealing with complex images. In this paper, we take the Unext network as a basis and utilize its encoder-decoder features. And taking inspiration from the mechanisms of cellular apoptosis and division, we design apoptosis and division algorithms to improve model performance. We propose a novel segmentation model which integrates the division and apoptosis algorithms and introduces spatial and channel convolution blocks into the model. Our proposed model not only improves the segmentation performance of breast ultrasound tumors, but also reduces the model parameters and computational resource consumption time. The model was evaluated on the breast ultrasound image dataset and our collected dataset. The experiments show that the SC-Unext model achieved Dice scores of 75.29% and accuracy of 97.09% on the BUSI dataset, and on the collected dataset, it reached Dice scores of 90.62% and accuracy of 98.37%. Meanwhile, we conducted a comparison of the model's inference speed on CPUs to verify its efficiency in resource-constrained environments. The results indicated that the SC-Unext model achieved an inference speed of 92.72ms per instance on devices equipped only with CPUs. The model's number of parameters and computational resource consumption are 1.46M and 2.13 GFlops, respectively, which are lower compared to other network models. Due to its lightweight nature, the model holds significant value for various practical applications in the medical field.

Predicting Time-to-Healing from a Digital Wound Image: A Hybrid Neural Network and Decision Tree Approach Improves Performance

Despite the societal burden of chronic wounds and despite advances in image processing, automated image-based prediction of wound prognosis is not yet in routine clinical practice. While specific tissue types are known to be positive or negative prognostic indicators, image-based wound healing prediction systems that have been demonstrated to date do not (1) use information about the proportions of tissue types within the wound and (2) predict time-to-healing (most predict categorical clinical labels). In this work, we analyzed a unique dataset of time-series images of healing wounds from a controlled study in dogs, as well as human wound images that are annotated for the tissue type composition. In the context of a hybrid-learning approach (neural network segmentation and decision tree regression) for the image-based prediction of time-to-healing, we tested whether explicitly incorporating tissue type-derived features into the model would improve the accuracy for time-to-healing prediction versus not including such features. We tested four deep convolutional encoder–decoder neural network models for wound image segmentation and identified, in the context of both original wound images and an augmented wound image-set, that a SegNet-type network trained on an augmented image set has best segmentation performance. Furthermore, using three different regression algorithms, we evaluated models for predicting wound time-to-healing using features extracted from the four best-performing segmentation models. We found that XGBoost regression using features that are (i) extracted from a SegNet-type network and (ii) reduced using principal components analysis performed the best for time-to-healing prediction. We demonstrated that a neural network model can classify the regions of a wound image as one of four tissue types, and demonstrated that adding features derived from the superpixel classifier improves the performance for healing-time prediction.

Integrated lithium niobate microwave photonic processing engine.

Integrated microwave photonics (MWP) is an intriguing technology for the generation, transmission and manipulation of microwave signals in chip-scale optical systems1,2. In particular, ultrafast processing of analogue signals in the optical domain with high fidelity and low latency could enable a variety of applications such as MWP filters3-5, microwave signal processing6-9 and image recognition10,11. An ideal integrated MWP processing platform should have both an efficient and high-speed electro-optic modulation block to faithfully perform microwave-optic conversion at low power and also a low-loss functional photonic network to implement various signal-processing tasks. Moreover, large-scale, low-cost manufacturability is required to monolithically integrate the two building blocks on the same chip. Here we demonstrate such an integrated MWP processing engine based on a 4 inch wafer-scale thin-film lithium niobate platform. It can perform multipurpose tasks with processing bandwidths of up to 67 GHz at complementary metal-oxide-semiconductor (CMOS)-compatible voltages. We achieve ultrafast analogue computation, namely temporal integration and differentiation, at sampling rates of up to 256 giga samples per second, and deploy these functions to showcase three proof-of-concept applications: solving ordinary differential equations, generating ultra-wideband signals and detecting edges in images. We further leverage the image edge detector to realize a photonic-assisted image segmentation model that can effectively outline the boundaries of melanoma lesion in medical diagnostic images. Our ultrafast lithium niobate MWP engine could provide compact, low-latency and cost-effective solutions for future wireless communications, high-resolution radar and photonic artificial intelligence.

OralImmunoAnalyser: a software tool for immunohistochemical assessment of oral leukoplakia using image segmentation and classification models.

Oral cancer ranks sixteenth amongst types of cancer by number of deaths. Many oral cancers are developed from potentially malignant disorders such as oral leukoplakia, whose most frequent predictor is the presence of epithelial dysplasia. Immunohistochemical staining using cell proliferation biomarkers such as ki67 is a complementary technique to improve the diagnosis and prognosis of oral leukoplakia. The cell counting of these images was traditionally done manually, which is time-consuming and not very reproducible due to intra- and inter-observer variability. The software presently available is not suitable for this task. This article presents the OralImmunoAnalyser software (registered by the University of Santiago de Compostela-USC), which combines automatic image processing with a friendly graphical user interface that allows investigators to oversee and easily correct the automatically recognized cells before quantification. OralImmunoAnalyser is able to count the number of cells in three staining levels and each epithelial layer. Operating in the daily work of the Odontology Faculty, it registered a sensitivity of 64.4% and specificity of 93% for automatic cell detection, with an accuracy of 79.8% for cell classification. Although expert supervision is needed before quantification, OIA reduces the expert analysis time by 56.5% compared to manual counting, avoiding mistakes because the user can check the cells counted. Hence, the SUS questionnaire reported a mean score of 80.9, which means that the system was perceived from good to excellent. OralImmunoAnalyser is accurate, trustworthy, and easy to use in daily practice in biomedical labs. The software, for Windows and Linux, with the images used in this study, can be downloaded from https://citius.usc.es/transferencia/software/oralimmunoanalyser for research purposes upon acceptance.

Bone age assessment based on three-dimensional ultrasound and artificial intelligence compared with paediatrician-read radiographic bone age: protocol for a prospective, diagnostic accuracy study

IntroductionRadiographic bone age (BA) assessment is widely used to evaluate children’s growth disorders and predict their future height. Moreover, children are more sensitive and vulnerable to X-ray radiation exposure than...

Diverter transformer-based multi-encoder-multi-decoder network model for medical retinal blood vessel image segmentation

The retinal blood vessel is an essential part of the fundus structure. It is important to accurately analyze the structure and distribution of retinal vessels, which can help make accurate medical diagnoses. However, it is still challenging to extract detailed information due to the problems of fuzzy edges, low resolution, and lots of noise in retinal blood vessel medical images. To extract the image detail information effectively, we propose a new diverter transformer-based multi-encoder-multi-decoder network model in this paper. The network model consists of a feature encoder module and a feature decoder module. Among them, the feature encoding module consists of a diverter transformer with a diverter adaptive mechanism, three encoder units with a convolution layer and max-pooling layer, and the two decoder units in the feature decoding module consist of an inverse convolution layer and an up-sampling layer, respectively. The Local Context Module (LCNet Module) in the feature encoding module learns richer local context feature information layer by layer through changing the width of the network while downsampling; the Global Encoder Module1 (G-Encoder Module1) and the Global Encoder Module2 (G-Encoder Module2) extract the global feature representation of retinal blood vessel images by performing a max-pooling operation to transform the input data into a vector of fixed dimensions, thus helping the network model to better understand and extract the global feature representation of retinal blood vessel images. The two decoder units in the feature decoding module receive local and global feature information from three encoder units, LCNet Module, G-Encoder Module1 and G-Encoder Module2, respectively. Decoder Module1 generates segmentation prediction by layer-by-layer up-sampling operation, and Decoder Module2 recovers the feature information by downsampling and decoding operations and fuses the recovered feature information to output, obtaining the final segmentation of the retinal blood vessels. The proposed diverter transformer-based multi-encoder-multi-decoder network model is validated on the DRIVE and STARE datasets with other classical and state-of-the-art network models, and its segmentation accuracy is 97.25% and 97.93%, respectively. Compared with the classical U-Net model, the improvement is 2.24% and 1.42%, respectively. Compared with the state-of-the-art SPNet model, the accuracy is increased by 0.61% on DRIVE and 1.01% on STARE. It indicates that the network model proposed in this paper has a significant competitive advantage in improving the segmentation performance of retinal blood vessel images.

Automatic Detection of Collapsed Buildings after the 6 February 2023 Türkiye Earthquakes Using Post-Disaster Satellite Images with Deep Learning-Based Semantic Segmentation Models

This study focuses on the identification of collapsed buildings in satellite images after earthquakes through deep learning-based image segmentation models. The performance of four different architectures, namely U-Net, LinkNet, FPN, and PSPNet, was evaluated using various performance metrics, such as accuracy, precision, recall, F1 score, specificity, AUC, and IoU. The study used satellite images taken from the area located in the south and southeast of Türkiye covering the eleven provinces which are most affected by the Mw 7.7 Pazarcık (Kahramanmaraş) and Mw 7.6 Elbistan (Kahramanmaraş) earthquakes. The results indicated that FPN and U-Net were the best-performing models depending on the performance metric of interest. FPN achieved the highest accuracy and specificity scores, as well as the best precision score, while U-Net achieved the best recall and F1 score values, as well as the best AUC and IoU scores. The training and validation accuracy and loss curves were analyzed, and the results indicated that all four models achieved an accuracy value of over 96%. The FPN model outperformed the others in terms of accurately segmenting images while maintaining a low loss value. This study provides insights into the potential of deep learning-based image segmentation models in disaster management and can be useful for future research in this field.

Attention guided approach for food type and state recognition

Advancements in computer vision have resulted in significant breakthroughs across various applications, and one notable area of progress is in the recognition of food ingredient types and states. The identification of food items, distinguishing between types like oranges or apples, and assessing their states, whether whole, peeled, sliced, or juiced, is a pivotal task with far-reaching implications for fields such as food safety, recipe analysis, and restaurant quality control. This paper introduces an innovative approach to food type and state recognition that capitalizes on attention mechanisms and incorporates mask fusion to improve the accuracy and robustness of the recognition process. We evaluate the proposed approach through quantitative and qualitative analyses and comparisons to previous methods. The results consistently demonstrate that our proposed approach, integrating attention mechanisms, outperforms baseline and state-of-the-art methods, achieving an accuracy of 87.11%. This achievement signifies a step forward in refining food image segmentation models and reinforces the applicability of advanced techniques in real-world scenarios.

An Algorithm for Measuring the Similarity of Histograms for Texture Image Segmentation

A simple algorithm for measuring the similarity between multi-column histograms is presented. The proposed algorithm is intended for texture segmentation of images using histograms as texture features. The purpose of developing such a specialized algorithm is to more accurately determine the boundaries between neighboring texture segments. The algorithm is specially designed so that to express the similarity value as a percentage. The main peculiarity of the proposed algorithm is that when calculating the similarity value, it considers not only the corresponding histogram columns but also takes into account their neighboring components. Due to this, the algorithm more adequately evaluates the similarity of histograms. The proposed algorithm was implemented as a computer program as an integral part of the image segmentation model. The efficiency of the histogram comparison algorithm is indirectly confirmed by the texture segmentation results of the image segmentation model in image processing experiments.

An Improved Variational-Based Model for Denoising and Segmentation of Vector-Valued Images

Preserving important features such as edges is one of the main concerns in models for denoising and segmenting vector-valued (colour) images. The Rudin-Osher-Fatemi (ROF) model is a well-known variational-based image denoising model that is capable of reducing image noise while preserving image edges. However, the ROF model is not formulated for denoising colour images and is less effective in preserving corners and weak edges. On the other hand, a variational-based selective segmentation model for colour images called the selective distance segmentation (DSS2) model has recently been proposed, which can effectively partition or extract a specific object in an image. However, the DSS2 model has problems in segmenting colour images with noise, which may result in poor segmentation. Therefore, in this research, we first modify the ROF model to denoise vector-valued images by including the edge detector and extending the formulation into a vector-valued framework. Second, we reformulate the DSS2 model by incorporating the modified ROF model as a new fitting term in the DSS2 model. Peak signal-to-noise ratio (PSNR) is used to measure the image quality, while Jaccard and Dice similarity index are used to evaluate the segmentation quality. The comparison between our proposed model and existing model shows that our model is more effective as indicated by higher PSNR, Jaccard and Dice similarity index values.

EU-net: An automated CNN based ebola U-net model for efficient medical image segmentation

EU-net: An automated CNN based ebola U-net model for efficient medical image segmentation

STRATEGIES AND EXPERIMENTS FOR MASSIVE 3D DIGITALIZATION OF THE REMAINS AFTER THE NOTRE DAME DE PARIS’ FIRE

Abstract. After the catastrophic fire at Notre Dame de Paris, a significant challenge was presented by the numerous lead-contaminated remnants. To address this, a detailed digitization strategy was devised and executed, tailored to the unique needs of this extensive and diverse corpus. This strategy involved the development and customization of both hardware and software tools, ensuring their effectiveness throughout the digitization process – from initial data acquisition to data dissemination.Central to our approach was the alignment of our methods with the distinct characteristics of each artifact, facilitating their effective preservation and future utility. Our strategy's adaptability was key, allowing us to incorporate advanced deep learning techniques into various aspects of our workflow. Notably, this included the implementation of the Segment Anything Model for automatic image segmentation, enhancing our image-based modeling capabilities. We also ventured into pioneering methods like 3D Gaussian Splatting and the exploration of radiance field methods for visualization.Moreover, the project has been mindful of data responsibility, aiming to make all digital data openly accessible beyond 2025. We have placed a strong emphasis on harmonizing and managing data, minimizing redundancies, and ensuring efficient storage, all while maintaining transparency about the limitations and errors in our methodologies. This holistic approach to digitization, balancing technological innovation with responsible data management, aims to preserve and make accessible the digital heritage of Notre Dame de Paris for future generations.

Detection and mapping of Amaranthus spinosus L. in bermudagrass pastures using drone imagery and deep learning for a site‐specific weed management

AbstractWeed encroachment negatively affects pasture productivity by reducing herbage allowance, stocking rates, and livestock performance. Amaranthus spinosus L. is a weed species widely found in pastures worldwide and is considered challenging for ranchers due to its great potential for invasion, making it difficult to control. The high costs of chemical application and the global concern about environmental impacts restrict indiscriminate herbicide spraying in pastures. Site‐specific weed management (SSWM) is a weed management strategy based on weed spot‐spraying that has the potential to overcome these issues. Images from unmanned aerial vehicles (UAVs) can provide valuable information for weed mapping to drive the herbicide application in pastures. Deep learning techniques have been highlighted in image classification tasks. We developed a deep convolutional neural network (CNN)‐based image segmentation model based on the U‐Net architecture to detect and map Amaranthus spinosus in bermudagrass pastures using red–green–blue images acquired through UAV flying in moderate‐high altitude. The images were acquired from twelve paddocks under three treatments (weed‐free, weed‐strips, or weed‐infested) during the summer (2021‐2022). The CNN model was able to detect around 80% of the A. spinosus with an average prediction accuracy of 94%. Our weed mapping showed the potential of using the U‐Net model to generate a herbicide application map to be inserted into the sprayer system, reducing up to 76% of the amount of herbicide applied. Further studies are encouraged to increase the robustness of the model across species and development stages and develop sprayer systems to implement the spot‐spraying in field conditions.

A Regional Brightness Control Method for a Beam Projector to Avoid Human Glare

In this study, we proposed a system to reduce the speaker’s suffering from the strong light of a beam projector by applying regional brightness control over the screen. Since the original image and the projected one on the screen are quite different in area, brightness, and color, the proposed system first transforms them so that they have the same area and similar color tone. Then, to accurately determine the difference between those images, we have introduced a SSIM map, which is a perception-based method of measuring image similarity. Accordingly, an image segmentation model is used to determine the speaker’s silhouette from the SSIM map. We applied a couple of well-trained segmentation models, such as Selfie and DeepLab-v3, provided with MediaPipe. The experimental results showed the operability of the proposed system and that it determines most of a lecturer’s body area on the screen. To closely evaluate the system’s effectiveness, we have measured error rates consisting of false-positive and false-negative errors in the confusion matrix. With the measured results, the error rates appeared so insignificant and stable that the proposed system provides a practical effect for the speakers, especially in the case of applying DeepLab-v3. With the results, it is implied that an accurate segmentation model can considerably elevate the effectiveness of the system.

Study on the Influence of Label Image Accuracy on the Performance of Concrete Crack Segmentation Network Models.

A high-quality dataset is a basic requirement to ensure the training quality and prediction accuracy of a deep learning network model (DLNM). To explore the influence of label image accuracy on the performance of a concrete crack segmentation network model in a semantic segmentation dataset, this study uses three labelling strategies, namely pixel-level fine labelling, outer contour widening labelling and topological structure widening labelling, respectively, to generate crack label images and construct three sets of crack semantic segmentation datasets with different accuracy. Four semantic segmentation network models (SSNMs), U-Net, High-Resolution Net (HRNet)V2, Pyramid Scene Parsing Network (PSPNet) and DeepLabV3+, were used for learning and training. The results show that the datasets constructed from the crack label images with pix-el-level fine labelling are more conducive to improving the accuracy of the network model for crack image segmentation. The U-Net had the best performance among the four SSNMs. The Mean Intersection over Union (MIoU), Mean Pixel Accuracy (MPA) and Accuracy reached 85.47%, 90.86% and 98.66%, respectively. The average difference between the quantized width of the crack image segmentation obtained by U-Net and the real crack width was 0.734 pixels, the maximum difference was 1.997 pixels, and the minimum difference was 0.141 pixels. Therefore, to improve the segmentation accuracy of crack images, the pixel-level fine labelling strategy and U-Net are the best choices.

Technical note: Generalizable and promptable artificial intelligence model to augment clinical delineation in radiation oncology

Efficient and accurate delineation of organs at risk (OARs) is a critical procedure for treatment planning and dose evaluation. Deep learning-based auto-segmentation of OARs has shown promising results and is increasingly being used in radiation therapy. However, existing deep learning-based auto-segmentation approaches face two challenges in clinical practice: generalizability and human-AI interaction. A generalizable and promptable auto-segmentation model, which segments OARs of multiple disease sites simultaneously and supports on-the-fly human-AI interaction, can significantly enhance the efficiency of radiation therapy treatment planning. Meta's segment anything model (SAM) was proposed as a generalizable and promptable model for next-generation natural image segmentation. We further evaluated the performance of SAM in radiotherapy segmentation. Computed tomography (CT) images of clinical cases from four disease sites at our institute were collected: prostate, lung, gastrointestinal, and head & neck. For each case, we selected the OARs important in radiotherapy treatment planning. We then compared both the Dice coefficients and Jaccard indices derived from three distinct methods: manual delineation (ground truth), automatic segmentation using SAM's 'segment anything' mode, and automatic segmentation using SAM's 'box prompt' mode that implements manual interaction via live prompts during segmentation. Our results indicate that SAM's segment anything mode can achieve clinically acceptable segmentation results in most OARs with Dice scores higher than 0.7. SAM's box prompt mode further improves Dice scores by 0.1∼0.5. Similar results were observed for Jaccard indices. The results show that SAM performs better for prostate and lung, but worse for gastrointestinal and head & neck. When considering the size of organs and the distinctiveness of their boundaries, SAM shows better performance for large organs with distinct boundaries, such as lung and liver, and worse for smaller organs with less distinct boundaries, like parotid and cochlea. Our results demonstrate SAM's robust generalizability with consistent accuracy in automatic segmentation for radiotherapy. Furthermore, the advanced box-prompt method enables the users to augment auto-segmentation interactively and dynamically, leading to patient-specific auto-segmentation in radiation therapy. SAM's generalizability across different disease sites and different modalities makes it feasible to develop a generic auto-segmentation model in radiotherapy.