Learning Adaptive Tensorial Density Fields for Clean Cryo-ET Reconstruction
Authors: YUANHAO WANG, Ramzi Idoughi, Wolfgang Heidrich
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate the superiority of our framework over existing methods using synthetic and real data. Thus, our framework boosts the quality of the reconstruction while reducing the computation time and the memory footprint. The code is available at https://github.com/yuanhaowang1213/adaptivetensordf. |
| Researcher Affiliation | Academia | Yuanhao Wang Ramzi Idoughi Wolfgang Heidrich King Abdullah University of Science and Technology (KAUST) {yuanhao.wang,ramzi.idoughi,wolfgang.heidrich}@kaust.edu.sa |
| Pseudocode | No | The paper describes the method in text and uses equations and figures, but it does not include any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | The code is available at https://github.com/yuanhaowang1213/adaptivetensordf. |
| Open Datasets | Yes | We evaluated our method on tilt-series datasets from EMPIAR (Electron Microscopy Pilot Image Archive): EMPIAR 10643 [38], and 10761 [12]. The EMPIAR 10643 dataset is a cryo-ET acquisition of the HIV-1 Gagdelta MASP1T8I assemblies with an angular range of [ 60 , 60 ] and an increment of 3 . From this dataset, we reconstructed two different series (40 and 51) independently. The EMPIAR 10751 dataset corresponds to the cryo-ET acquisition of a HEK cell, with an angular range of [ 60 , 60 ]. |
| Dataset Splits | No | The paper discusses using a synthetic dataset for “parameter tuning” and “robustness evaluation,” which implies a form of validation, but it does not specify explicit train/validation/test splits or their proportions. |
| Hardware Specification | No | The paper does not provide specific details about the hardware used for the experiments, such as GPU models, CPU types, or memory. |
| Software Dependencies | No | The paper mentions implementing the framework in C++ but does not specify versions for any libraries, frameworks, or other software dependencies. |
| Experiment Setup | Yes | After several experiments on all the hyperparameters of our approach, it appears that the most important parameters are the tensor dimensions (the dimension of the matrix-vector factors) and the feature size. The choice of tensor dimensions affects the speed and quality of the training and reconstruction. Smaller dimensions lead to faster training and smoother reconstruction, but they may miss some fine details of the sample. Larger dimensions capture more details, but they may also introduce overfitting or noise, as the network tends to learn the noise after learning the structures (See the Supplement for visual comparison). The feature size also influences the training speed and the reconstruction quality. Smaller features can speed up the training, but they may lose some details. Larger features can preserve more details, but they may increase the computation cost. A balance between these trade-offs should be sought for a better recovery. To study the impact of these two parameters, we measured the 3D Peak-Signal-to-Noise Ratio (PSNR) and 3D Structural Similarity Index Measure (SSIM) of the reconstructed volume using different values of these parameters. The results are shown in Table 1 and Table 2. From the analysis of Table 1 and Table 2, the best performances are obtained with a tensor dimensions and feature size equals to 22 and 16, respectively. |