In our study, TensorFlow was used to implement U-Net, while U-NeXt, DeepLabV3+, and ConResNet were implemented using PyTorch. The models were tested on a computer with an AMD Ryzen 5 2600X six-core 3.60 GHz GPU and NVIDIA Quadro P6000 GPU. The models are uploaded in the GitHub page, listed in the Supplementary Materials section. Table 2 shows the model architecture parameters—including the number of epochs, the activation function, the loss function, the optimizer, and the number of trainable parameters. The number of epochs for training for each model was chosen to avoid overfitting. The number was determined empirically for each model through trial and error so that the difference between training accuracy and test accuracy was small.
Quadro p6000 gpu
Manufactured by NVIDIA
The NVIDIA Quadro P6000 is a professional-grade graphics processing unit (GPU) designed for high-performance computing applications. It features 3,840 CUDA cores, a 12GB GDDR5X frame buffer, and a memory bandwidth of up to 547.7 GB/s. The Quadro P6000 is capable of delivering exceptional graphics performance and compute power for a variety of professional workloads.
Automatically generated - may contain errors
Lab products found in correlation
14 protocols using quadro p6000 gpu
1
Comparative Evaluation of Deep Learning Models
2
High-Resolution 3D Imaging Workflow
Check if the same lab product or an alternative is used in the 5 most similar protocols
+ Expand
Image processing was performed on a Windows 10-based workstation equipped with two Intel Xeon Gold 5120 CPUs, 1 TB of RAM, an NVIDIA Quadro P6000 GPU. To stitch the sub-volumes, the Fiji-based plugin BigStitcher30 (link) was used. Image analysis was performed with Fiji31 (link) and MATLAB (Mathworks), and 3D renderings were produced with ChimeraX or Arivis.
3
U-Net-based Network Architecture for Image Reconstruction
Check if the same lab product or an alternative is used in the 5 most similar protocols
+ Expand
The network architecture we employed in this work was derived from the U-net54 . More details of the network structure are provided in Supplementary Fig. S5 . We adopted the Adam optimizer with a learning rate of , , , and to update the weights in the neural network. We also used an exponential decay with a decay rate of 0.9 and decay steps of 100. The momentum and epsilon parameters in the batch normalization were 0.99 and 0.001, respectively. The leak parameter of Leaky ReLU was 0.2. The regularization parameter of the TV was 10−10. The code was run on a computer with an Intel Xeon CPU E5-2696 V3, 64 GB RAM, and an NVIDIA Quadro P6000 GPU. The main progress is illustrated in Algorithm 1 . For the sake of comparison, we use the same network model for GIDC and GIDL. We also released our code at https://github.com/FeiWang0824/GIDC .![]()
4
Optimized Deep Learning for Segmentation
Check if the same lab product or an alternative is used in the 5 most similar protocols
+ Expand
The network was trained using a Tensorflow distributed machine learning system [18 ]. Stochastic gradient descent was used to update the parameters and batch size 1, to contribute to regularization. The networks at each fold were trained for 33K epochs to prevent overfitting, starting with a learning rate of 0.05, with 0.9 decay rate every 1000 epochs, in conjunction with a momentum of 0.9 [19 ]. The dropout ratio was set to 0.5. The experiments were performed on a NVidia Quadro P6000 GPU. Training was carried out with the goal of optimising the inverse of generalized dice coefficient as the objective function, defined at pixel level as:
where are the reference pixel values, are the probabilistic predicted values and provides invariance to imbalanced label set distributions, by correcting the contribution of each label, , by the inverse of its size [20 ].
where are the reference pixel values, are the probabilistic predicted values and provides invariance to imbalanced label set distributions, by correcting the contribution of each label, , by the inverse of its size [20 ].
5
Deep Learning Image Processing Workflow
Check if the same lab product or an alternative is used in the 5 most similar protocols
+ Expand
Python (version 3.8.8) and TensorFlow (version 2.3.0) framework were used in Jupyter Notebook to run all the experiments in this study. The image pre-processing and models training were done on a Windows 10 workstation which equipped with dual Intel XEON E5-2630cv3 processors, NVIDIA Quadro P6000 GPU, 120 GB of RAM, and CUDA version of 11.4.
6
High-Resolution 3D Imaging Workflow
7
Deep Learning Model Optimization
Check if the same lab product or an alternative is used in the 5 most similar protocols
+ Expand
We evaluated several models with two, three and four fully connected layer models. We selected the number of units for each layer from {250, 500, …, 4000} with dropout rate from {0.2, 0.5} and learning rate from {0.01, 0.001, 0.0001} for the Adam optimizer [68 ]. We trained the models with mini-batch size of 32. We performed 50 trials of random search for best parameters for each type of the models and selected the best model based on validation loss. We use the TensorFlow 2.0 [69 ] machine learning system with Keras API and tune our parameters with Keras Tuner.
Our model is trained and tuned in less than 1 hour on a single Nvidia Quadro P6000 GPU. In average, it annotates more than 100 samples per second.
Our model is trained and tuned in less than 1 hour on a single Nvidia Quadro P6000 GPU. In average, it annotates more than 100 samples per second.
8
Evaluating AlphaFold Protein Structure Predictions
Check if the same lab product or an alternative is used in the 5 most similar protocols
+ Expand
AlphaFold initial release and v2.0.0 were downloaded from github and installed as described (https://github.com/deepmind/alphafold ) under Linux (Debian 10, 96 GB RAM, NVidia Quadro P6000 GPU with 24 GB RAM or NVidia RTX A6000 GPU with 48GB RAM GPU). We introduced minor modifications into the code to overcome memory usage problems in case of large multiple sequence alignment files and to be able to run multimer predictions with the initial release (http://alphafold.hegelab.org ). Our runs used all genetic databases (--db_preset=full_dbs). Generated structures were evaluated based on PAE, pLDDT, and ipTM+pTM scores [21 (link),36 (link)]. In addition, all top-scored structures were inspected visually.
9
CycleGAN for Image-to-Image Translation
Check if the same lab product or an alternative is used in the 5 most similar protocols
+ Expand
The CycleGAN model was composed of two discriminators and two generators that achieve translation between two domains, A and B. In addition, the networks were trained to minimize cycle consistency losses, 11 which aim to guarantee consistency when forward and backward translations are applied successively on an image. A detailed depiction of the network can be seen in Fig. 1.
The model was implemented using TensorFlow 15 and trained for approximately 2.5 epochs with the following parameters: batch size = 24, generator with six residual blocks, 16 Adam optimizer, 17 and learning rate = 0.0002. Residual blocks were composed of two convolutional layers followed by instance normalization, 18 and Rectified Linear Unit (ReLU) activation.
Training was performed on a workstation with a 3.6-GHz, six-core processor with 64-GB RAM, NVIDIA Quadro P6000 GPU.
The model was implemented using TensorFlow 15 and trained for approximately 2.5 epochs with the following parameters: batch size = 24, generator with six residual blocks, 16 Adam optimizer, 17 and learning rate = 0.0002. Residual blocks were composed of two convolutional layers followed by instance normalization, 18 and Rectified Linear Unit (ReLU) activation.
Training was performed on a workstation with a 3.6-GHz, six-core processor with 64-GB RAM, NVIDIA Quadro P6000 GPU.
10
Evaluating Artificial Data Augmentation for Surgical Tool Segmentation
Check if the same lab product or an alternative is used in the 5 most similar protocols
+ Expand
In order to quantitatively assess the informative content of the generated images, we evaluated whether the proposed method can serve as a data augmentation method for surgical-tool segmentation tasks. A total of 1600 images were generated by means of the proposed approach and used to train a standard U-Net architecture [27] . The network was trained for 50 epochs using binary cross-entropy as loss function and Adam optimizer (learning rate = 0.0001, batch size = 16). The model was implemented in Keras and trained using a NVIDIA Quadro P6000 GPU. After training, the best model was chosen as the one that minimized the loss on the validation set (30% of the whole dataset).
The model was finally tested on 40 images from the original MIS dataset. We calculate five evaluation indexes respectively: Søresen Dice Coefficient (Dice), Jaccard Similarity (Jaccard), Precision (Precision), Recall (Recall) and F1-score (F1):
Jaccard = T P T P + FN + FP (9)
It is important to point out that the main goal of the study was not to achieve high segmentation performance; rather, we wanted to evaluate the informative content of the artificially generated images. For this reason, no further parameter tuning was performed to improve segmentation results.
The model was finally tested on 40 images from the original MIS dataset. We calculate five evaluation indexes respectively: Søresen Dice Coefficient (Dice), Jaccard Similarity (Jaccard), Precision (Precision), Recall (Recall) and F1-score (F1):
Jaccard = T P T P + FN + FP (9)
It is important to point out that the main goal of the study was not to achieve high segmentation performance; rather, we wanted to evaluate the informative content of the artificially generated images. For this reason, no further parameter tuning was performed to improve segmentation results.
About PubCompare
Our mission is to provide scientists with the largest repository of trustworthy protocols and intelligent analytical tools, thereby offering them extensive information to design robust protocols aimed at minimizing the risk of failures.
We believe that the most crucial aspect is to grant scientists access to a wide range of reliable sources and new useful tools that surpass human capabilities.
However, we trust in allowing scientists to determine how to construct their own protocols based on this information, as they are the experts in their field.
Ready to get started?
Sign up for free.
Registration takes 20 seconds.
Available from any computer
No download required
Revolutionizing how scientists
search and build protocols!