Developing a MAMBA-Neural Radiance Fields Model for Accurate and Efficient 3D X-Ray Image Reconstruction 1 st Shiming Duan Faculty of Information Technology City University Malaysia Petaling Jaya 46100, Malaysia duan_shiming@qq.com 3 rd M. Kazem Chamran Faculty of Information Technology City University Malaysia Petaling Jaya 46100, Malaysia ORCID: 0000 0003 3836 4443 2 nd Mustafa Muwafak Alobaedy Faculty of Information Technology City University Malaysia Petaling Jaya 46100, Malaysia alobaedy@ieee.org 4 th Shyam R Department of computer application Presidency college, Hebbal Bangalore Bengaluru, Karnataka 560024, India shyam.r@presidency.edu.in Abstract—This research proposes MAMBA-Neural Radiance Fields (MAMBANeRF), a novel model for 3D reconstruction from a single X-ray image in medical applications. By integrating the efficient sequence modeling capabilities of Mamba with the powerful spatial representation of Neural Radiance Fields (NeRF), MAMBANeRF addresses key limitations of traditional 3D reconstruction techniques— such as Multi-View Stereo and Stereo Matching—which typically require multiple views and high computational resources. While X-ray imaging offers advantages like low radiation exposure, cost-effectiveness, and fast acquisition, single-image 3D reconstruction remains challenging due to limited spatial cues. MAMBANeRF overcomes this challenge by combining Mamba’s sequential reasoning with NeRF’s volumetric rendering, enabling the rapid generation of highprecision 3D models using standard computing hardware. The model supports enhanced visualization of complex anatomical structures to aid diagnostics and enables longitudinal comparisons for treatment evaluation. Furthermore, MAMBANeRF is well-suited for integration with virtual reality and digital twin technologies, paving the way for more accessible, intelligent, and personalized healthcare solutions. Keywords—X-ray image, 3D Reconstruction, Mamba, NeRF, State Space Model I. INTRODUCTION Medical 3D Reconstruction is the process of converting 2D data such as x-ray images into a 3D model [1]. Medical reconstruction can improve diagnostic accuracy, optimize treatment planning, preoperative simulation, personalized medicine, medical education, training, and research support [2]. The challenges of medical 3D reconstruction are mainly composed of several components: • Data quality and consistency The quality and consistency of image data are critical to the accuracy of 3D reconstruction. There may be discrepancies in the data from different equipment and scanning conditions, affecting the quality of the final model [3]. • Computational complexity The 3D reconstruction process involves much data processing and computation, especially high-resolution model generation, which requires powerful computational resources or efficient algorithm support. • Real-time requirements The real-time processing requirements for 3D reconstruction are very high, requiring fast generation of highquality models while maintaining low latency. II. LITERATURE REVIEW 3D reconstruction technology has been gaining popularity since its inception. From 3D reconstruction in the real world to medical 3D reconstruction, various 3D reconstruction technology systems have emerged. These studies have always focused on achieving efficient, precise, and low-consumption modeling objectives. With the evolution of DL, 3D reconstruction technology faces more opportunities and challenges. TABLE I. COMPARISON TABLE OF THE CHARACTERISTICS OF EACH MODEL Model Information Compressio n Parallel Training Training efficiency Reasoning efficiency CNN Limited longdistance dependence Parallelizable High Low Transfo rmer No compression of historical data High higher arithmetic requirements High High arithmetic requiremen ts Mamba Selecting focus Parallelizable Summarizing the history record, balancing training and reasoning Previous studies have employed Convolutional Neural Networks (CNNs) and Transformer to develop 3D reconstruction models. Recent advancements show that derivative models of Mamba exhibit comparable, and in some cases superior, efficiency in addressing tasks traditionally handled by CNNs and Transformer. The distinct characteristics and performance traits of each model are synthesized and summarized in Table I. A. Model Analysis Mamba views sequence modeling as the task of compressing contextual information into compact states. From this perspective, attention mechanisms in Transformers appear inefficient—they retain the full context rather than compressing it. This leads to high memory usage, linear-time inference, and multi-dimensional training complexity, ultimately limiting the scalability and efficiency of Transformer-based models [4]. The trade-off between the efficiency and effectiveness of sequential models is characterized by how well they compress their state. Typically, efficient models must have a small, compressed state, while effective models must have a state that contains all the necessary information from the context. Mamba proposes that the basic principle for constructing sequence models is selectivity or the context-aware ability to attend to or filter out inputs to sequence states. In particular, the selection mechanism controls how information propagates or interacts along the sequence dimension. The core of Mamba is defined by two equations: the state equation (1) and Output equation (2) [5]. h(t) = Ah(t − 1) + Bx(t) (1) y(t) = Ch(t) + Dx(t) (2) In other words, mapping an input sequence x(t) to a potential state representation h(t) and deriving the predicted output sequence y(t). The Mamba model utilizes a circular selection mechanism, replaces convolutional algorithms, and recursion. After weighing, Mamba uses the hardware-aware algorithm as the solution to improve parallel computation issues. Hardware-aware algorithms mainly utilize the GPU to realize states only at more efficient levels in the memory hierarchy. Most operations (except matrix multiplication) are memory bandwidth limited [4]. Scanning operations are limited by memory bandwidth, where Mamba uses kernel fusion to reduce the number of memories IO's, which leads to significant speedups compared to standard implementations. Core features of the Mamba model: B. Efficient Self-Attention Mechanism. The Mamba model uses State Space Model to replace the self-attention mechanism to improve computational efficiency and reduce memory usage. C. Enhanced Context Modeling. The Mamba model employs improved context modeling methods that allow the model to capture long-range dependencies better [6]. D. Efficient Training and Inference. Mamba models employ more efficient techniques in the training and inference process, such as mixed-precision training and model pruning, to speed up training and reduce the consumption of computational resources. E. Task Adaptability. The Mamba model is highly task-adaptable and can perform well in a wide range of natural language processing tasks. Through different pre-training strategies and taskspecific fine-tuning techniques, Mamba applies to various application scenarios [2]. From 2019, most 3D reconstruction models (e.g. Zero-1to-3 [7]， Point-E [3]) adhere to solving 3D reconstruction problems. Some exploratory studies use 3D reconstruction algorithms for image reconstruction, where the basic approach is to reconstruct the topological structure and features of objects in a picture using a 3D reconstruction model. The image is then reconstructed based on the features and topological structure. We have statistically analyzed the past decade research contributions of each study. Fig. 1. Model-Counts Distribution in MAMBANeRF-Related Research. As shown in Fig. 1, according to the above chart, this study analyzed 65 research papers that studied various models for 3D reconstruction, including Neural Radiance Field (NeRF), GAN, U-Net, and other categories such as datasets and reviews. Firstly, we can confirm that there are many methods for 3D reconstruction but NeRF has led the pack with an impressive 86% share of the papers collected in the past decade. So, the NeRF has potential for future development and application. III. MAMBANERF: A NOVEL FUSION OF MAMBA AND NERF FOR MEDICAL 3D RECONSTRUCTION X-ray 3D reconstruction is a method of obtaining organ structural information about an object from X-ray images. It typically combines different X-ray imaging techniques and algorithms to create a detailed 3D model of an object [8]. Commonly, 3D reconstruction schemes are Mesh, Point Cloud, Voxel, and Implicit [9]. NeRF essentially provides a new inverse rendering approach to the neural network to simulate continuous field and body rendering [10]. The NeRF model represents a continuous scene as a function whose inputs are 5D vectors, including the 3D coordinate position of a spatial point, roll, and pitch [11]. NeRF 86% GAN 5% U-Net 1% other 8% Model-Counts Distribution NeRF GAN U-Net other X-ray 3D reconstruction techniques have made significant milestones in the medical imaging field. Deep Learning (DL) algorithms play a critical role in 3D reconstruction [12]. Convolutional Neural Networks (CNN) and Generative Adversarial Networks (GAN) have improved image quality, reduced noise, and enhanced reconstruction under low-dose conditions [13]. The Mamba from the State Space Model (SSM) is a novel innovation in the Deep Learning domain [14]. The Mamba has achieved state-of-the-art performance in solving modalities such as speech, audio, and genomics [15]. In the field of 3D reconstruction, the Mamba model is still a newcomer. CNN, GAN, and Transformer based models can all make a mark in the 3D reconstruction field [16]. Therefore, this research proposes to combine the Mamba with the NeRF model (MAMBANeRF) to reconstruct a 3D model from x-ray images. Theoretical validation that MAMBANeRF can combine the advantages of Mamba and NeRF models. MAMBANeRF model presents the first deep integration of Mamba and NeRF. MAMBANeRF innovation solve the bottlenecks of traditional 3D reconstruction techniques and paves the way for high-precision and high-efficiency 3D medical modeling. The research scope as follow: F. 3D reconstruction from a Single X-ray Image 3D reconstruction from a single X-ray image has long been a focal and challenging problem in the field of 3D reconstruction [17]. The core challenge lies in efficiently and accurately reconstructing 3D structures from limited 2D information [18]. G. Limitations of NeRF and Its Variants The existing research primarily focuses on reconstruction tasks under multi-view or dense input conditions, leaving a significant knowledge gap in single-image input scenarios [19]. Additionally, NeRF models still have considerable room for optimization for computational efficiency, generalization ability, and detail reconstruction. H. Advantages and Potential of the Mamba Model Compared to traditional Transformer models, Mamba demonstrates higher efficiency and accuracy in processing long-sequence data, especially showing significant advantages when handling high-dimensional data [15]. These characteristics make it an ideal candidate for integration with NeRF, potentially bringing improvements to the 3D reconstruction process. Significance of MAMBANeRF Model This study proposes a novel MAMBANeRF model, which combines Mamba's efficient sequence modeling capabilities with NeRF's neural radiance field representation to address key challenges in 3D reconstruction from a single X-ray image. The architecture of Mamba enables it to run efficiently on standard hardware, thereby reducing reliance on highperformance computing resources. By integrating Mamba's linear-complexity attention mechanism into NeRF, MAMBANeRF can substantially improve rendering efficiency, shorten reconstruction time, and enhance reconstruction resolution. This innovative integration addresses the computational bottlenecks of traditional NeRF models. The MAMBANeRF model is highly significance because traditional medical image reconstruction takes a long time, with high hardware costs, and massive computational resources. Theoretically, the results of this study can achieve accurate models with less computing power, lower hardware requirements. IV. M ETHODOLOGY This study analyzes the three research directions of X-ray, the 3D reconstruction, and the DL model based on X-ray technology, the 3D reconstruction model and Mamba model. The study finds some potential intersection points between these three research directions. As shown in Fig. 2, MAMBANeRF model combines the features of Mamba and NeRF to construct a new 3D reconstruction model. The connection between the three research fields in the figure is that NeRF has many applications in the X-ray 3D reconstruction field. Meanwhile, the improvement of X-ray denoising technology will affect the accuracy of the NeRF model. The 3D reconstruction model focuses on the medical X-ray reconstruction field. The improvement of image denoising by the Mamba model will affect the efficiency of X-ray denoising technology. The advantage of Mamba in model acceleration is that it also has the potential to accelerate NeRF. Fig. 2. Conceptual Framework of MAMBANeRF: Integration of X-ray Imaging, Deep Learning, and NeRF for Enhanced Medical and Material Scene Understanding. The NeRF model has an abnormal tolerance mechanism, and the original noise in the X-ray scan data has less impact on the NeRF model. At the same time, the NeRF model can reconstruct data from a single photo. Therefore, NeRF has the potential and value in handling medical X-ray data. However, the difficulties of NeRF include the fact that NeRF requires more hardware computing power, and the NeRF 3D reconstruction process requires a longer reconstruction time [8]. This study reconstructs the NeRF model using the model characteristics of the Mamba model in the time series domain. This innovation is a reconstruction of the basic NeRF model. As a radiance field model, the NeRF model consists of discrete points between points. By reconstructing the encoding structure of the NeRF model using Mamba, it is possible to serialize the discrete points. This change can achieve the characteristics of optimizing the encoding and processing speed of the NeRF model. This study refines its research content into three core components: X-ray data acquisition, 3D reconstruction models, and acceleration models. The detailed analysis is as follows: I. X-ray Data Acquisition and Challenges Data acquisition is achieved through X-ray scanning. However, X-ray images are essentially projections of 3D organs from a specific angle, where data perpendicular to the projection direction is superimposed on the same pixel. J. Requirements and Selection of 3D Reconstruction Models Based on comprehensive considerations of acquisition speed, cost, noise tolerance, and practical feasibility, the 3D reconstruction model needs to meet the following characteristics: high noise tolerance, high reconstruction accuracy, fast reconstruction speed, and low hardware requirements. After careful consideration, the NeRF model is selected as the foundational model for this study due to its high reconstruction accuracy and high noise tolerance. K. Design and Selection of Acceleration Models After determining the data acquisition approach and 3D reconstruction model, an acceleration model or module is needed to address the shortcomings of the NeRF model. The Mamba model is selected due to its high accuracy in handling potential relationships between data points and its hardware-friendly design. The Mamba model can efficiently capture correlations between data points while significantly improving computational efficiency through its hardwareoptimized design. L. Construction of the Solution By integrating X-ray data acquisition, the NeRF 3D reconstruction model, and the Mamba acceleration model, this study develops an efficient, accurate, and hardware-friendly solution for X-ray 3D reconstruction. V. P ROPOSED MAMBANERF MODEL After literature review and background analysis, we can find that the Mamba model is suitable for compression of arithmetic power, Vision Mamba has a great advantage in image pre-processing, and the NeRF model is more efficient in the process of 3D reconstruction. The Vision Mamba is a model centered around the Mamba model and focused on computer vision [15]. The Vision Mamba model has high accuracy and efficient computing capabilities in both image classification and segmentation. Vision Mamba is a series of derivative models centered around Mamba, and it has various variants suitable for different application fields. In the actual model, corresponding derivative models will be matched according to different applications. Although the derivative models may change, the core part of the model has not affected the MAMBANeRF model of this study. In this study, the most basic Vision Mamba model will be used for construction. Therefore, the initial design logic is to input the data into Vision Mamba to cut the image and make a dataset of these data. Vision Mamba then make a dataset of image classification. And use the original data, image cutting dataset, image classification dataset input to the NeRF model for training. The NeRF model training result is a dot matrix, the output dot matrix and the real clock may have errors. Therefore, the NeRF model data is input to the Mamba model for compression and accuracy enhancement. As shown in Fig. 3 Input data into NeRF module and Vision Mamba module respectively. Image segmentation and classification are performed in the Vision Mamba module. Different organs and tissues require different constraints. The segmentation and classification results are used to provide 3D reconstruction constraints to the NeRF module. The data are combined with constraints for 3D reconstruction in the NeRF module. The Mamba module reorganizes the 3D reconstruction results into organs -the state space of the model. Fig. 3. Structural Overview of the MAMBANeRF Model for Classification and Segmentation-Driven 3D Reconstruction. As shown in Fig. 3, the MAMBANeRF model achieves 3D reconstruction through the following steps, with significant optimization effects and research highlights: Data Preprocessing Application of Vision Mamba: MAMBANeRF first utilizes Vision Mamba to segment and classify the raw input data. This step addresses the weak generalization capability of the NeRF model, especially when handling complex data. The data encoding after image segmentation and image classification based on Vision Mamba has connections among states. That is, the corresponding results can be described by the initial state plus the changes. Data Integration: After processing by Vision Mamba, the generated feature data, classification data, and raw data are fed into the NeRF model. The core of 3D reconstruction using the NeRF model can be understood as the relationship of states between the light points emitted by the camera and the target. In other words, using SSM to interpret NeRF is equivalent to summing up the states of the points. The X-ray data itself carries the density changes of the coordinate points. Therefore, the classification labels and segmented images obtained in the Vision Mamba section will constrain the reconstruction results of the NeRF model. These data add the virtual viewpoints and constraints of NeRF to train the NeRF model. 3D reconstruction Initial Reconstruction: In the preliminary training phase, the block of MAMBANeRF can convert 2D images into 3D models. During the initial reconstruction process, the reconstruction task is mainly accomplished by the NeRF model. The results output by Vision Mamba can already be represented as the superimposition of image states. By combining the NeRF model to superimpose the position states of each pixel point, a potential three-dimensional model can be established in this way. However, the reconstruction efficiency of such a three-dimensional model is relatively low, and it requires high hardware equipment and computing power. Therefore, further processing is needed to complete the optimization. However, this study optimizes the process to improve efficiency and hardware friendliness. Data Representation Optimization: Mamba combines the feature and classification data to build a compact representation of the raw data. The data is encoded into 3D models composed of point superimpositions in NeRF and into image state superimpositions in Vision Mamba. The corresponding 3D models are reconstructed from the image state superimpositions directly through Mamba. This process can significantly reduce the performance loss and computational requirements in the NeRF model. This representation significantly reduces the required data for NeRF model computations, enhancing computational efficiency. Model Optimization Mapping Relationship Reconstruction: Using the data optimized by Mamba, MAMBANeRF redefines the mapping relationship between images and 3D models. This step improves the model's computational speed and maximizes the utilization of hardware performance. Enhanced Generalization Capability: Through the preprocessing by Vision Mamba, the MAMBANeRF model has the potential of generalization capability when handling most medical X-ray images of tissue structures. Research Highlights Optimized Computational Efficiency: By combining feature data and classification data using Mamba to represent raw data, the computational burden on the NeRF model is greatly reduced, thereby accelerating the model's computational speed and making better use of hardware resources. Through the preprocessing and data representation optimization by Vision Mamba, the MAMBANeRF model improves the efficiency and accuracy of 3D reconstruction. VI. RESULTS AND DISCUSSION This research systematically analyzes the three areas in this study, the X-ray field, the application field of Mamba and Mamba derivative models, and the three-dimensional reconstruction field. The X-ray field is the basis for the study, and the current status of X-ray three-dimensional reconstruction is analyzed to determine the research gap. This study proposes a potential model based on analysis. In 3D reconstruction modeling, this study first determines the technical principles and partial reconstruction methods of 3D reconstruction. After conducting a comprehensive analysis in the 3D reconstruction field since the release of the NeRF model in 2020, the application of NeRF in the field of 3D reconstruction has experienced rapid development. However, these studies still face the challenges of the inherent NeRF and the difficulties of medical images. Therefore, the application and solutions of NeRF in medical X-ray reconstruction are not mature. This study fills the knowledge gap between these two models by combining Mamba and NeRF and utilizes Mamba to solve the inherent challenges of NeRF. This study reviews the technical features of Mamba and NeRF and designs the theoretical model of a potential MAMBANeRF model. The extension model of Mamba can complement the defects of the NeRF. Mamba in the direction of image processing has some advantages. Mamba in arithmetic power compression is lightweight, which has some potential. Therefore, the Mamba and NeRF models have the potential to be combined. VII. EXPECTING RESULTS The MAMBANeRF model is currently in the theoretical design phase, and its potential benefits are not yet clear. We plan to evaluate the entire model by assessing the proportion of existing modules within the overall model. We have made the following observations: According to the workflow analysis of the MAMBANeRF model, NeRF demonstrates significant advantages over other 3D reconstruction models when processing noisy image data. The Mamba model outperforms the Transformer and CNN models regarding computational efficiency and hardwarefriendliness design. Based on these observations, we make the following inferences: Expected Model Architecture Integration Deep Integration of Mamba and NeRF Enhance deployment capabilities on mobile/edge devices. Maintain the original robustness in handling noisy data. Multi-Task Enhancement of Vision Mamba Potentially develop a unified framework supporting joint training for 3D reconstruction and semantic understanding. Enable end-to-end scene understanding (3D geometry + semantic segmentation + object detection). VIII. CONCLUSION This study focuses on key issues in medical image 3D reconstruction, systematically analyzing the current state of X-ray imaging and 3D reconstruction technologies. It identifies that 3D reconstruction based on single-view X-ray images remains a critical research direction requiring breakthroughs. To address the technical challenges in this field, we have established two core research areas: (1) Improving the accuracy of single-view 3D reconstruction. (2) Enhancing the computational efficiency of NeRF models. The innovation of this study lies in the first introduction of Mamba into the field of medical 3D reconstruction. Theoretical analysis demonstrates that Mamba's unique sequence modeling capabilities and efficient computational architecture provide a new paradigm for addressing these dual challenges. Based on this, we propose the MAMBANeRF innovative architecture: by integrating Mamba's time-varying state modeling mechanism with NeRF's implicit representation, we successfully construct a solution for 3D reconstruction from single-view X-ray images. MAMBANeRF has completed its full theoretical framework design and uses simulations to verify its feasibility. Although the physical model hasn't implementation, theoretical derivations confirm its advantages. (1) overcoming the dimensional limitations of single-view reconstruction, (2) effectively reducing memory usage. This work expands the application boundaries of the Mamba model in the field of medical imaging. MAMBANeRF provides a new technical pathway for clinical needs. REFERENCES [1] E. Ahishakiye, M. B. Van Gijzen, J. Tumwiine, R. Wario, and J. Obungoloch, “A survey on deep learning in medical image reconstruction,” Intell. Med., vol. 1, no. 3, pp. 118–127, Sep. 2021, doi: 10.1016/j.imed.2021.03.003. [2] K. Sohan and M. A. Yousuf, “3D Bone Shape Reconstruction from 2D X-ray Images Using MED Generative Adversarial Network,” in 2020 2nd International Conference on Advanced Information and Communication Technology (ICAICT), Dhaka, Bangladesh: IEEE, Nov. 2020, pp. 53–58. doi: 10.1109/ICAICT51780.2020.9333477. [3] P. Maken and A. Gupta, “2D-to-3D: A Review for Computational 3D Image Reconstruction from X-ray Images,” Arch. Comput. Methods Eng., vol. 30, no. 1, pp. 85–114, Jan. 2023, doi: 10.1007/s11831-022- 09790-z. [4] A. Gu and T. Dao, “Mamba: Linear-Time Sequence Modeling with Selective State Spaces,” May 31, 2024, arXiv: arXiv:2312.00752. Accessed: Aug. 07, 2024. [Online]. Available: http://arxiv.org/abs/2312.00752 [5] A. Gu et al., “Combining Recurrent, Convolutional, and Continuoustime Models with Linear State-Space Layers”. [6] S. Çimen, A. Gooya, M. Grass, and A. F. Frangi, “Reconstruction of coronary arteries from X-ray angiography: A review,” Med. Image Anal., vol. 32, pp. 46–68, 2016. [7] Y. Kasten, D. Doktofsky, and I. Kovler, “End-To-End Convolutional Neural Network for 3D Reconstruction of Knee Bones from Bi-planar X-Ray Images,” in Machine Learning for Medical Image Reconstruction, vol. 12450, F. Deeba, P. Johnson, T. Würfl, and J. C. Ye, Eds., in Lecture Notes in Computer Science, vol. 12450. , Cham: Springer International Publishing, 2020, pp. 123–133. doi: 10.1007/978-3-030-61598-7_12. [8] I. A. Elbakri and J. A. Fessler, “Segmentation-free statistical image reconstruction for polyenergetic x-ray computed tomography with experimental validation,” Phys. Med. Biol., vol. 48, no. 15, pp. 2453– 2477, Aug. 2003, doi: 10.1088/0031-9155/48/15/314. [9] A. Farshian et al., “Deep-learning-based 3-d surface reconstruction— a survey,” Proc. IEEE, vol. 111, no. 11, pp. 1464–1501, 2023. [10] K. Gao, Y. Gao, H. He, D. Lu, L. Xu, and J. Li, “NeRF: Neural Radiance Field in 3D Vision, A Comprehensive Review,” Nov. 30, 2023, arXiv: arXiv:2210.00379. Accessed: Oct. 26, 2024. [Online]. Available: http://arxiv.org/abs/2210.00379 [11] F. Dellaert and L. Yen-Chen, “Neural Volume Rendering: NeRF And Beyond,” Jan. 14, 2021, arXiv: arXiv:2101.05204. Accessed: Oct. 26, 2024. [Online]. Available: http://arxiv.org/abs/2101.05204 [12] S. Hosseinian and H. Arefi, “3D RECONSTRUCTION FROM MULTI-VIEW MEDICAL X-RAY IMAGES – REVIEW AND EVALUATION OF EXISTING METHODS,” Int. Arch. Photogramm. Remote Sens. Spat. Inf. Sci., vol. XL-1/W5, pp. 319–326, Dec. 2015, doi: 10.5194/isprsarchives-XL-1-W5-319-2015. [13] T. Kattenborn, J. Leitloff, F. Schiefer, and S. Hinz, “Review on Convolutional Neural Networks (CNN) in vegetation remote sensing,” ISPRS J. Photogramm. Remote Sens., vol. 173, pp. 24–49, Mar. 2021, doi: 10.1016/j.isprsjprs.2020.12.010. [14] D. Han et al., “Demystify Mamba in Vision: A Linear Attention Perspective,” May 26, 2024, arXiv: arXiv:2405.16605. Accessed: Aug. 07, 2024. [Online]. Available: http://arxiv.org/abs/2405.16605 [15] X. Liu, C. Zhang, and L. Zhang, “Vision Mamba: A Comprehensive Survey and Taxonomy,” May 07, 2024, arXiv: arXiv:2405.04404. Accessed: Aug. 07, 2024. [Online]. Available: http://arxiv.org/abs/2405.04404 [16] Z. Shao, X. Yao, Z. Wang, and J. Gao, “ST-MambaSync: The Complement of Mamba and Transformers for Spatial-Temporal in Traffic Flow Prediction,” May 09, 2024, arXiv: arXiv:2404.15899. Accessed: Aug. 07, 2024. [Online]. Available: http://arxiv.org/abs/2404.15899 [17] A. Corona-Figueroa, J. Frawley, S. Bond-Taylor, S. Bethapudi, H. P. H. Shum, and C. G. Willcocks, “MedNeRF: Medical Neural Radiance Fields for Reconstructing 3D-aware CT-Projections from a Single Xray,” Apr. 08, 2022, arXiv: arXiv:2202.01020. Accessed: Sep. 23, 2024. [Online]. Available: http://arxiv.org/abs/2202.01020 [18] Y.-L. Liao, Y.-C. Yang, and Y.-C. F. Wang, “3D Shape Reconstruction from a Single 2D Image via 2D-3D Self-Consistency,” Nov. 29, 2018, arXiv: arXiv:1811.12016. doi: 10.48550/arXiv.1811.12016. [19] C. Deng et al., “Nerdi: Single-view nerf synthesis with languageguided diffusion as general image priors,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 20637–20647. Accessed: Oct. 26, 2024. [Online]. Available: http://openaccess.thecvf.com/content/CVPR2023/html/Deng_NeRDi_ Single-View_NeRF_Synthesis_With_Language- Guided_Diffusion_As_General_Image_CVPR_2023_paper.html