![](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F9711370%2F55fdd37178b63612afb111403208297d%2FOptimizer.png?generation=1688521271597503&alt=media\")
| \n | \n |
|---|---|
| 🔴 | \nHigh Level Recommended | \n
| 🟡 | \nMedium Level Recommended | \n
| 🟢 | \nLow Level Recommended | \n
| Concept | \n\n | Usage | \n
|---|---|---|
| Convolution | \n🔴 | \nNetwork Layer | \n
| Convolution Transpose | \n🔴 | \nNetwork Layer | \n
| Concatenation | \n🔴 | \nNetwork Layer | \n
| Adam | \n🔴 | \nOptimizer | \n
| Padding | \n🔴 | \nRuntime Processing | \n
| Strides | \n🔴 | \nRuntime Processing | \n
| ReLU | \n🟡 | \nActivation Function | \n
| Sigmoid | \n🟡 | \nActivation Function | \n
| Binary Cross Entropy | \n🟡 | \nLoss | \n
| Dropout | \n🟢 | \nNetwork Layer | \n
| Pooling | \n🟢 | \nNetwork Layer | \n
We know that CNNs are showing great results in the field of Vision related tasks
achieve accuracies of over 99% on image classification tasks.achieve accuracies of over 90% on object detection tasksachieve accuracies of over 80% on semantic segmentation tasksWhile CNNs have already existed for a long time, their success was limited due to the
The normal classification task was actually pretty easy for the network to learn. But as the field of Image Segmentation, increased, it became difficult for the model to learn both the class label and position/localization .Localization is a challenging task, because it requires the model to learn both the appearance of the objects/features of interest, and their spatial location in the image.
The major use of Image Semantic Segmentation is in Bio-Medical'. InBio-Medicalimage processing, thedesired outputis not just aclassification of the image, but also thelocation of the objects` or features of interest in the image
For example, in Bio-Medical image processing, it is often important to be able to localize tumors or other abnormalities in images of tissues or organs. This is because the location of these abnormalities can be used to diagnose diseases or to plan treatments.
Deep Neural Networks Segment Neuronal Membranes in Electron Microscopy Images used the method of masking to predict the correct labels of the image
Ciresan model had many advantages such as
number of patches that was extracted from a training image was much larger than the number of training images themselves.This was because a patch is a small, rectangular section of an image. For example, a patch of size 64x64 can be extracted from any image. If we have a training set of 100 images, we can extract a total of 100x64x64 = 32,768,000 patches.
But the technique used by Ciresan had some drawbacks
quite slow because the network must be run separately for each patch, which means that the model must be run multiple times for each image. This can be very slow, especially for large images.image is divided into patches, there is a lot of overlap between the patches. This is because the patches are typically small, and they are often arranged in a regular grid.balancing the accuracy of localization with the amount of context .Larger patches allow for more accurate localization, but they also reduce the amount of context that is available to the network. This is because max-pooling layers are typically used to down-sample the image.Smaller patches, allow for more context, but they also reduce the accuracy of localization. This is because smaller patches contain less information about the location of the object.Less Training Data - The model is modified such that it works with very few training images and yields more precise segmentations.
Touching Objects - weighted loss is applied, where the separating background labels between touching cells obtain a large weight in the loss function.
The weights are intialized such that each feature map in the network has approximately unit variance. This was achieved by intializing the weights by
\n a Gaussian Distribution(1 , sqrt(2/N))
The Adam optimizer is used to train the $UNet$ model, and the momentum parameter is set to 0.99.
The Adam optimizer is a Stochastic Gradient Descent (SGD) optimizer that was introduced in 2014. The momentum parameter controls how much the optimizer remembers its past gradients. Though, a momentum value of 0.99 is a relatively high value. This means that the Adam optimizer will be more likely to follow its past gradients, which can help to improve the stability of training and prevent the model from overfitting.
The energy function is computed by a pixel-wise soft-max over the final feature map
![](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F9711370%2Fd5ba65b1f7d31619da7823a977567643%2FPredictioon.png?generation=1688521293363603&alt=media\")
![]()
\n![](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F9711370%2Faa1e84d19521a53ab934ba4562eb0946%2FAssumpetions.png?generation=1688521344912976&alt=media\")
The cross entropy then penalizes at each position the deviation of p^`_x (x) from 1
\n![](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F9711370%2Fd7ce2dae61705f5d1c91310a3e7844bd%2FLos.png?generation=1688521366011103&alt=media\")
The network is tested on 2 different segmentation tasks
| \n | Error | \nAccuracy | \n
|---|---|---|
| Neural Structures in Electron Microscopic Recordings | \n0.0003529 | \n0.9203 | \n
| Light Microscopic Images | \n\n | 92% | \n
This study was supported by the Excellence Initiative of the German Federal and State governments (EXC 294) and by the BMBF (Fkz 0316185B).
\n| Paper | \nAuthors | \nPlace | \nYear | \n
|---|---|---|---|
| Deep neural net-works segment neuronal membranes in electron microscopy images | \nCiresan D.C. Gambardella L.M. Giusti A. Schmidhuber J. | \nNIPS. pp. 2852 2860 | \n2012 | \n
| Discriminative un-supervised feature learning with convolutional neural networks | \nDosovitskiy A. Springenberg J.T. Riedmiller M. Brox T. | \nNIPS | \n2014 | \n
| Rich feature hierarchies for ac-curate object detection and semantic segmentation. | \nGirshick R. Donahue J. Darrell T. Malik J. | \nProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) | \n2014 | \n
| Hyper-columns for object segmentation and ne-grained localization | \nHariharan B. Arbelez P. Girshick R. Malik J. | \n\n | 2014 | \n
| Delving deep into rectifiers: Surpassing human-level performance on image-net classification | \nHe, K., Zhang, X., Ren, S., Sun, J. | \n\n | 2015 | \n
| Convolutional architecture for fast feature embedding | \nJia Y. Shelhamer E. Donahue J. Karayev S. Long J. Girshick R. Guadar-rama S. Darrell T. | \n\n | 2014 | \n
| Image-net classification with deep convolutional neural networks | \nKrizhevsky A. Sutskever I. Hinton G.E. | \nNIPS. pp. 1106 1114 | \n2012 | \n
| Backpropagation applied to handwritten zip code recognition. Neural Computation | \nLeCun Y. Boser B. Denker J.S. Henderson D. Howard R.E. Hubbard W. Jackel L.D. | \n1(4) 541 551 | \n1989 | \n
| Fully convolutional networks for semantic segmentation | \nLong J. Shelhamer E. Darrell T. | \n\n | 2014 | \n
| A benchmark for comparison of cell tracking algorithms | \nMaska M. (...) de Solorzano C.O. | \nBioinformatics 30, 1609 1617 | \n2014 | \n
| Very deep convolutional networks for large-scale image recognition | \nSimonyan K. Zisserman A. | \n\n | 2014 | \n
| Web page of the cell tracking challenge | \n\n | \n | \n |
| Web page of the em segmentation challenge | \n\n | \n | \n |
| \n | \n |
|---|---|
| 🔴 | \nHigh Level Recommended | \n
| 🟡 | \nMedium Level Recommended | \n
| 🟢 | \nLow Level Recommended | \n
| Concept | \n\n | Usage | \n
|---|---|---|
| Convolution | \n🔴 | \nNetwork Layer | \n
| Convolution Transpose | \n🔴 | \nNetwork Layer | \n
| Concatenation | \n🔴 | \nNetwork Layer | \n
| Adam | \n🔴 | \nOptimizer | \n
| Padding | \n🔴 | \nRuntime Processing | \n
| Strides | \n🔴 | \nRuntime Processing | \n
| ReLU | \n🟡 | \nActivation Function | \n
| Sigmoid | \n🟡 | \nActivation Function | \n
| Binary Cross Entropy | \n🟡 | \nLoss | \n
| Dropout | \n🟢 | \nNetwork Layer | \n
| Pooling | \n🟢 | \nNetwork Layer | \n
We know that CNNs are showing great results in the field of Vision related tasks
achieve accuracies of over 99% on image classification tasks.achieve accuracies of over 90% on object detection tasksachieve accuracies of over 80% on semantic segmentation tasksWhile CNNs have already existed for a long time, their success was limited due to the
The normal classification task was actually pretty easy for the network to learn. But as the field of Image Segmentation, increased, it became difficult for the model to learn both the class label and position/localization .Localization is a challenging task, because it requires the model to learn both the appearance of the objects/features of interest, and their spatial location in the image.
The major use of Image Semantic Segmentation is in Bio-Medical'. InBio-Medicalimage processing, thedesired outputis not just aclassification of the image, but also thelocation of the objects` or features of interest in the image
For example, in Bio-Medical image processing, it is often important to be able to localize tumors or other abnormalities in images of tissues or organs. This is because the location of these abnormalities can be used to diagnose diseases or to plan treatments.
Deep Neural Networks Segment Neuronal Membranes in Electron Microscopy Images used the method of masking to predict the correct labels of the image
Ciresan model had many advantages such as
number of patches that was extracted from a training image was much larger than the number of training images themselves.This was because a patch is a small, rectangular section of an image. For example, a patch of size 64x64 can be extracted from any image. If we have a training set of 100 images, we can extract a total of 100x64x64 = 32,768,000 patches.
But the technique used by Ciresan had some drawbacks
quite slow because the network must be run separately for each patch, which means that the model must be run multiple times for each image. This can be very slow, especially for large images.image is divided into patches, there is a lot of overlap between the patches. This is because the patches are typically small, and they are often arranged in a regular grid.balancing the accuracy of localization with the amount of context .Larger patches allow for more accurate localization, but they also reduce the amount of context that is available to the network. This is because max-pooling layers are typically used to down-sample the image.Smaller patches, allow for more context, but they also reduce the accuracy of localization. This is because smaller patches contain less information about the location of the object.Less Training Data - The model is modified such that it works with very few training images and yields more precise segmentations.
Touching Objects - weighted loss is applied, where the separating background labels between touching cells obtain a large weight in the loss function.
The weights are intialized such that each feature map in the network has approximately unit variance. This was achieved by intializing the weights by
\n a Gaussian Distribution(1 , sqrt(2/N))
The Adam optimizer is used to train the $UNet$ model, and the momentum parameter is set to 0.99.
The Adam optimizer is a Stochastic Gradient Descent (SGD) optimizer that was introduced in 2014. The momentum parameter controls how much the optimizer remembers its past gradients. Though, a momentum value of 0.99 is a relatively high value. This means that the Adam optimizer will be more likely to follow its past gradients, which can help to improve the stability of training and prevent the model from overfitting.
The energy function is computed by a pixel-wise soft-max over the final feature map
![]()
\n
The cross entropy then penalizes at each position the deviation of p^`_x (x) from 1
\n
The network is tested on 2 different segmentation tasks
| \n | Error | \nAccuracy | \n
|---|---|---|
| Neural Structures in Electron Microscopic Recordings | \n0.0003529 | \n0.9203 | \n
| Light Microscopic Images | \n\n | 92% | \n
This study was supported by the Excellence Initiative of the German Federal and State governments (EXC 294) and by the BMBF (Fkz 0316185B).
\n| Paper | \nAuthors | \nPlace | \nYear | \n
|---|---|---|---|
| Deep neural net-works segment neuronal membranes in electron microscopy images | \nCiresan D.C. Gambardella L.M. Giusti A. Schmidhuber J. | \nNIPS. pp. 2852 2860 | \n2012 | \n
| Discriminative un-supervised feature learning with convolutional neural networks | \nDosovitskiy A. Springenberg J.T. Riedmiller M. Brox T. | \nNIPS | \n2014 | \n
| Rich feature hierarchies for ac-curate object detection and semantic segmentation. | \nGirshick R. Donahue J. Darrell T. Malik J. | \nProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) | \n2014 | \n
| Hyper-columns for object segmentation and ne-grained localization | \nHariharan B. Arbelez P. Girshick R. Malik J. | \n\n | 2014 | \n
| Delving deep into rectifiers: Surpassing human-level performance on image-net classification | \nHe, K., Zhang, X., Ren, S., Sun, J. | \n\n | 2015 | \n
| Convolutional architecture for fast feature embedding | \nJia Y. Shelhamer E. Donahue J. Karayev S. Long J. Girshick R. Guadar-rama S. Darrell T. | \n\n | 2014 | \n
| Image-net classification with deep convolutional neural networks | \nKrizhevsky A. Sutskever I. Hinton G.E. | \nNIPS. pp. 1106 1114 | \n2012 | \n
| Backpropagation applied to handwritten zip code recognition. Neural Computation | \nLeCun Y. Boser B. Denker J.S. Henderson D. Howard R.E. Hubbard W. Jackel L.D. | \n1(4) 541 551 | \n1989 | \n
| Fully convolutional networks for semantic segmentation | \nLong J. Shelhamer E. Darrell T. | \n\n | 2014 | \n
| A benchmark for comparison of cell tracking algorithms | \nMaska M. (...) de Solorzano C.O. | \nBioinformatics 30, 1609 1617 | \n2014 | \n
| Very deep convolutional networks for large-scale image recognition | \nSimonyan K. Zisserman A. | \n\n | 2014 | \n
| Web page of the cell tracking challenge | \n\n | \n | \n |
| Web page of the em segmentation challenge | \n\n | \n | \n |
\n\n## 5.2 | Prediction\n\nThe `energy function` is computed by a `pixel-wise soft-max` over the final feature map\n\n\n\n![]()\n### Where\n\n* a_k(x) = Feature Channel k\n* X 𝛜 Ω = Pixel Position\n* K = Number Of Labels\n\n### Assumptions\n\n\n\n## 5.3 | Loss\n\nThe cross entropy then penalizes at each position the deviation of p^`_x (x) from 1\n\n\n\n# 6 | Experiments 🥼\n\nThe network is tested on 2 different `segmentation tasks`\n\n||Error|Accuracy\n|---|---|---\n|Neural Structures in Electron Microscopic Recordings|0.0003529|0.9203\n|Light Microscopic Images||92%\n\n# 8 | Acknowledgements 🤝 \n\nThis study was supported by the Excellence Initiative of the German Federal and State governments (EXC 294) and by the BMBF (Fkz 0316185B).\n\n# 9 | Refrences 🔗\n\n|Paper|Authors|Place|Year|\n|---|---|---|\n|Deep neural net-works segment neuronal membranes in electron microscopy images|`Ciresan` `D.C.` `Gambardella` `L.M.` `Giusti A.` `Schmidhuber J.`|NIPS. pp. 2852 2860|2012\n|Discriminative un-supervised feature learning with convolutional neural networks|`Dosovitskiy` `A.` `Springenberg` `J.T.` `Riedmiller` `M.` `Brox` `T.`|NIPS|2014\n|Rich feature hierarchies for ac-curate object detection and semantic segmentation.|`Girshick` `R.` `Donahue` `J.` `Darrell` `T.` `Malik` `J.`| Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)|2014\n|Hyper-columns for object segmentation and ne-grained localization |`Hariharan` `B.` `Arbelez` `P.` `Girshick` `R.` `Malik` `J.`||2014\n|Delving deep into rectifiers: Surpassing human-level performance on image-net classification|`He`, `K.`, `Zhang`, `X.`, `Ren`, `S.`, `Sun`, `J.`||2015\n|Convolutional architecture for fast feature embedding |`Jia` `Y.` `Shelhamer` `E.` `Donahue` `J.` `Karayev` `S.` `Long` `J.` `Girshick` `R.` `Guadar-rama` `S.` `Darrell` `T.`||2014\n|Image-net classification with deep convolutional neural networks|`Krizhevsky` `A.` `Sutskever` `I.` `Hinton` `G.E.`|NIPS. pp. 1106 1114|2012\n|Backpropagation applied to handwritten zip code recognition. Neural Computation|`LeCun` `Y.` `Boser` `B.` `Denker` `J.S.` `Henderson` `D.` `Howard` `R.E.` `Hubbard` `W.` `Jackel` `L.D.`|1(4) 541 551|1989\n|Fully convolutional networks for semantic segmentation |`Long` `J.` `Shelhamer` `E.` `Darrell` `T.`||2014\n|A benchmark for comparison of cell tracking algorithms |`Maska` `M.` `(...)` `de Solorzano` `C.O.`|Bioinformatics 30, 1609 1617|2014|Image segmentation with cascaded hierarchical models and logistic disjunctive normal networks.|`Seyedhosseini` `M.` `Sajjadi` `M.` `Tasdizen` `T.`|Computer Vision (ICCV), 2013 IEEE International Conference on. pp.|2013\n|Very deep convolutional networks for large-scale image recognition|`Simonyan` `K.` `Zisserman` `A.`||2014\n|[Web page of the cell tracking challenge](http://www.codesolorzano.com/celltrackingchallenge/Cell_Tracking_Challenge/Welcome.html)\n|[Web page of the em segmentation challenge](http://brainiac2.mit.edu/isbi_challenge/)",
"votes": 3
}
],
"comments": [],
"raw_markdown_by_id": {
"2330418": "# 1 | Pre-Reqeuistics 🎓\n\n|||\n|---|---\n|🔴|`High Level Recommended`\n|🟡|`Medium Level Recommended`\n|🟢|`Low Level Recommended`\n\n|Concept||Usage|\n|---|---|---\n|Convolution|🔴|`Network Layer`\n|Convolution Transpose|🔴|`Network Layer`\n|Concatenation|🔴|`Network Layer`\n|Adam|🔴|`Optimizer`\n|Padding|🔴|`Runtime Processing`\n|Strides|🔴|`Runtime Processing`\n|ReLU|🟡|`Activation Function`\n|Sigmoid|🟡|`Activation Function`\n|Binary Cross Entropy|🟡|`Loss`\n|Dropout|🟢|`Network Layer`\n|Pooling|🟢|`Network Layer`\n\n# 2 | Starting 🌱\n\nWe know that `CNNs` are showing great results in the field of `Vision` related tasks\n\n* **Image Classification** - CNNs have been used to `achieve accuracies` of over 99% on `image classification` tasks.\n* **Object Detection** - CNNs have been used to `achieve accuracies` of over 90% on `object detection` tasks\n* **Semantic Segmentation** - CNNs have been used to `achieve accuracies` of over 80% on `semantic segmentation` tasks\n\nWhile `CNNs` have already `existed for a long time`, their `success was limited` due to the\n* size of the available training sets\n* size of the considered networks.\n\nThe `normal classification` task was actually `pretty easy` for the network to learn. But as the field of `Image Segmentation`, increased, it became `difficult` for the model to `learn both the class label and position/localization` .`Localization` is a `challenging task`, because it requires the model to `learn both the appearance` of the `objects/features` of interest, and their `spatial location` in the image.\n\nThe major use of `Image Semantic Segmentation` is in `Bio-Medical'. In `Bio-Medical` image processing, the `desired output` is not just a `classification of the image`, but also the `location of the objects` or features of interest in the image\n\nFor example, in `Bio-Medical` image processing, it is often `important` to be able to `localize tumors` or `other abnormalities` in `images of tissues` or `organs`. This is because the `location of these abnormalities` can be used to `diagnose diseases` or to `plan treatments`.\n\n# 3 | Inspiration 💡\n\n**[Deep Neural Networks Segment Neuronal Membranes in Electron Microscopy Images](https://papers.nips.cc/paper_files/paper/2012/file/459a4ddcb586f24efd9395aa7662bc7c-Paper.pdf)** used the `method of masking to predict the correct labels` of the image\n\n`Ciresan` model had many advantages such as\n* Localization - The network was able to identify the location of an object in an image.\n* Larger Training Data - the `number of patches` that was extracted from a `training image` was `much larger than the number of training images themselves`.\n\nThis was because a `patch` is a `small`, `rectangular section` of an image. For example, a patch of size 64x64 can be extracted from any image. If we have a training set of 100 images, we can extract a total of 100x64x64 = 32,768,000 `patches`.\n\nBut the technique used by `Ciresan` had some `drawbacks`\n* Long Training Time - The model was `quite slow` because the `network must be run separately for each patch`, which means that the model `must be run multiple times for each image`. This can be very slow, especially for large images.\n* Redundancy - When an `image` is `divided into patches`, there is a lot of `overlap between the patches`. This is because the `patches are typically small`, and they are `often arranged in a regular grid`.\n* Localization-Context Trade-Off - The problem of `balancing the accuracy of localization` with the `amount of context` .\n* * `Larger patches` allow for more `accurate localization`, but they `also reduce the amount of context` that is available to the network. This is because `max-pooling` layers are typically used to `down-sample the image`.\n* * `Smaller patches`, allow for `more context`, but they also `reduce the accuracy of localization`. This is because `smaller patches contain less information` about the location of the object.\n\n# 4 | Advantages ✅ \n\n* Less Training Data - The model is `modified` such that it works with `very few training images` and yields `more precise segmentations`.\n\n* Touching Objects - `weighted loss` is applied, where the `separating background labels between touching cells` obtain a `large weight` in the loss function.\n\n# 5 | Parameters and Formulations 🧬 \n\n## 5.1 | Weights Intialization\n\nThe weights are intialized such that each feature map in the network has approximately unit variance. This was achieved by intializing the weights by\n a `Gaussian Distribution(1 , sqrt(2/N))`\n\n## 5.1 | Optimizer\n\nThe `Adam optimizer` is used to train the $UNet$ model, and the `momentum parameter` is set to 0.99.\n\nThe `Adam optimizer` is a `Stochastic` `Gradient` `Descent` (SGD) `optimizer` that was introduced in 2014. The `momentum parameter` controls `how much the optimizer remembers its past gradients`. Though, a momentum value of 0.99 is a `relatively high value`. This means that the Adam `optimizer` will be more likely to `follow its past gradients`, which can `help to improve the stability of training` and `prevent the model from overfitting`.\n\n\n\n## 5.2 | Prediction\n\nThe `energy function` is computed by a `pixel-wise soft-max` over the final feature map\n\n\n\n![]()\n### Where\n\n* a_k(x) = Feature Channel k\n* X 𝛜 Ω = Pixel Position\n* K = Number Of Labels\n\n### Assumptions\n\n\n\n## 5.3 | Loss\n\nThe cross entropy then penalizes at each position the deviation of p^`_x (x) from 1\n\n\n\n# 6 | Experiments 🥼\n\nThe network is tested on 2 different `segmentation tasks`\n\n||Error|Accuracy\n|---|---|---\n|Neural Structures in Electron Microscopic Recordings|0.0003529|0.9203\n|Light Microscopic Images||92%\n\n# 8 | Acknowledgements 🤝 \n\nThis study was supported by the Excellence Initiative of the German Federal and State governments (EXC 294) and by the BMBF (Fkz 0316185B).\n\n# 9 | Refrences 🔗\n\n|Paper|Authors|Place|Year|\n|---|---|---|\n|Deep neural net-works segment neuronal membranes in electron microscopy images|`Ciresan` `D.C.` `Gambardella` `L.M.` `Giusti A.` `Schmidhuber J.`|NIPS. pp. 2852 2860|2012\n|Discriminative un-supervised feature learning with convolutional neural networks|`Dosovitskiy` `A.` `Springenberg` `J.T.` `Riedmiller` `M.` `Brox` `T.`|NIPS|2014\n|Rich feature hierarchies for ac-curate object detection and semantic segmentation.|`Girshick` `R.` `Donahue` `J.` `Darrell` `T.` `Malik` `J.`| Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)|2014\n|Hyper-columns for object segmentation and ne-grained localization |`Hariharan` `B.` `Arbelez` `P.` `Girshick` `R.` `Malik` `J.`||2014\n|Delving deep into rectifiers: Surpassing human-level performance on image-net classification|`He`, `K.`, `Zhang`, `X.`, `Ren`, `S.`, `Sun`, `J.`||2015\n|Convolutional architecture for fast feature embedding |`Jia` `Y.` `Shelhamer` `E.` `Donahue` `J.` `Karayev` `S.` `Long` `J.` `Girshick` `R.` `Guadar-rama` `S.` `Darrell` `T.`||2014\n|Image-net classification with deep convolutional neural networks|`Krizhevsky` `A.` `Sutskever` `I.` `Hinton` `G.E.`|NIPS. pp. 1106 1114|2012\n|Backpropagation applied to handwritten zip code recognition. Neural Computation|`LeCun` `Y.` `Boser` `B.` `Denker` `J.S.` `Henderson` `D.` `Howard` `R.E.` `Hubbard` `W.` `Jackel` `L.D.`|1(4) 541 551|1989\n|Fully convolutional networks for semantic segmentation |`Long` `J.` `Shelhamer` `E.` `Darrell` `T.`||2014\n|A benchmark for comparison of cell tracking algorithms |`Maska` `M.` `(...)` `de Solorzano` `C.O.`|Bioinformatics 30, 1609 1617|2014|Image segmentation with cascaded hierarchical models and logistic disjunctive normal networks.|`Seyedhosseini` `M.` `Sajjadi` `M.` `Tasdizen` `T.`|Computer Vision (ICCV), 2013 IEEE International Conference on. pp.|2013\n|Very deep convolutional networks for large-scale image recognition|`Simonyan` `K.` `Zisserman` `A.`||2014\n|[Web page of the cell tracking challenge](http://www.codesolorzano.com/celltrackingchallenge/Cell_Tracking_Challenge/Welcome.html)\n|[Web page of the em segmentation challenge](http://brainiac2.mit.edu/isbi_challenge/)"
}
}