MIScnn: A Framework for Medical Image Segmentation with Convolutional Neural Networks and Deep Learning

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 09 Issue: 01 | Jan 2022 www.irjet.net p-ISSN: 2395-0072
MIScnn: A Framework for Medical Image Segmentation with
Convolutional Neural Networks and Deep Learning
Shivam Singh1, Vishesh Yadav2, Suraj Khokarale3, Devarsh Patil4
1Student at D.Y Patil Institute of Engineering and Technology
---------------------------------------------------------------------***----------------------------------------------------------------------
Abstract - The excessive ease of use and comfort of current
medical imaging has sparked a strong need for automated
medical picture segmentation. Current image segmentation
systems, however, lack the necessary functionality for a basic
medical image segmentation pipeline architecture. Pipelines
that have already been enacted are generally independent
computer programmes that are optimized for a certain
collection of public data. As a result, this article introducesthe
MIScnn Python module for ASCII text files.
MIScnn aims to provide an intuitive API that allows for the
rapid creation of a medical image segmentation pipeline that
includes data I/O, pre-processing, data augmentation, patch-
wise analysis, metrics, a library of state-of-the-art deep
learning models, and model usage suchastraining, prediction,
and fully automatic evaluation (e.g. cross-validation).
Similarly, comprehensive pipeline customization is possible
because to strong configurability and many open interfaces.
The Kidney Tumour Segmentation Challenge 2019 data set
(multiclass semantic segmentation with 300 CT images)
yielded a strong predictor based on the conventional3DU-Net
model after cross-validation using MIScnn.
We were able to demonstrate that the MIScnn framework
allows researchers to quickly put up a comprehensive medical
image segmentation pipeline using only a few lines of code in
this experiment. The MIScnn source code is accessible on
Github at https://github.com/frankkramer-lab/MIScnn.
Key Words: Medical image analysis, Computer-aided
diagnosis, Biomedical image segmentation, U-Net, Deep
learning, Open-source framework.
1. INTRODUCTION
Because of the widespreaduseofmodernmedical imagingin
technology, there is a greater demand for automatic medical
picture segmentation. Despite this crucial need, existing
medical image segmentation technologieslack thecapability
required to create a basic medical picture segmentation
pipeline.
MIScnn is an open-source framework with an intuitive API
that lets you easily build medical image segmentation
pipelines with Convolutional Neural Networks and Deep
Learning models in only a few lines of code. The goal of
MIScnn is to provide a framework API that allows for the
rapid development ofmedical imagesegmentationpipelines,
including information I/O, preprocessing, data
augmentation, patch-wise analysis, metrics, a library of
progressive deep learning models, and model usage such as
training, prediction, and automatic analysis.
Convolutional networks are commonly used for
classification tasks, with the output to an image being a
single category label. However, in a number of visual
activities, particularly in the medical specialty picture
process, the needed output should include localization, i.e.,
every element should be assigned a category name.
Furthermore, in medical specialty duties, thousands of
training photographs are oftenoutofreach.Asa result,using
an area region (patch) around each element as input, we
trained a network in a sliding-window setup to predict the
class label of each element. This network will first become
localised. Second, the number of patches in the training data
is considerably more than the number of Training Images.
© 2022, IRJET | Impact Factor value: 7.529 | ISO 9001:2008 Certified Journal | Page 42

2. Data Input
2.1 NifTi data I/O interface
For loading magnetic resonance imaging and computed
axial tomography data into the framework, MIScnnprovides
a data I/O interface for the Neuroimaging Information
Science Technology Initiative (NifTI) file format. This
structure was first devised to expedite the event while also
enhancing the usability of information science technologies
related to neuroimaging. Still, it's currently unremarkably
used for sharing public and anonymous magnetic resonance
imaging and CT knowledge sets, not only for brain imaging,
but also for any other type of human 3D imaging.The 3D
image matrix and numerous data, such as the thickness of
the magnetic resonance imaging slices,arestoredinanNIfTI
file.
2.2 Custom data I/O interface
MIScnn allows the use of bespoke knowledge I/O
interfaces for various image knowledge formats in addition
to the mandated NIfTI I/O interface. This open interface
enables MIScnn to handle certain medical specialty imaging
choices (for example, MRI slice thickness), without the loss
of this information due to a format conversion demand. A
specific I/O interface should be dedicated to the
preprocessing process, and it should return to the medical
picture as a second or 3D matrix for integration into the
process.
3. Data Preprocessing
3.1 Pixel intensity normalization
Inconsistent signal intensity ranges in photos will have a
significant impact on segmentation methods' performance.
Due to completely distinct picture formats, variable
hardware/instruments (e.g. completely separate scanners),
technical inconsistencies,andeasilya biological variance,the
signal ranges of medical specialist imaging knowledge are
very varied amongst knowledge sets. Moreover, picture
segmentation techniques that use machine learning often
perform better on alternatives that follow a standard
distribution. Scaling and standardizing imaging information
is advised to achieve dynamic signal intensity variation
uniformity.
3.2 Clipping
Similar to component intensity standardization, it's
conjointly common to clip component intensities to an
explicit vary. Outside of this range, intensity costs will be
reduced to the lowest or most variable value. Even in many
scanners, component intensity levels for similar organs or
tissue types are assumed to be equal, especially in computer
imaging images.Organ-specific componentintensityclipping
might be used to take advantage of this.
FIG 1 . Flowchart of MIScnn Pipeline

3.3 Resampling
To change the width and/or height of images, the
resampling technique is used. This ends up in a brand-new
image with a changing range of pixels. resonance or laptop
pictorial representation scans will have different slice
thicknesses. However, in order to train neural network
models, the pictures must have the same slice thickness or
voxel spacing, this could be accomplished through
resampling. To begin with, downsampling images limits the
amount of GPU memoryavailablefortrainingandprediction.
3.4 One hot encoding
MIScnncanhandlebothbinary(background/cancer) and
multi-class (background/kidney/liver/lungs)segmentation
problems. Using a variable with two states, zeroandone,the
depiction of a binary segmentation may be readily
generated. Except in machine learning techniques, such as
deep learning models, where many category segmentation
labels are used, it is necessary totranslatethecategoriesinto
an extra mathematical picture. Using the One Hot coding
style, this might be accomplished by creating a single binary
variable for each segmentation category. One Hot
mechanically encodes segmentation labels with more than
two categories using MIScnn.
3.5 Patch‑wise and full image analysis
In 3D segmentation analysis,theavailableGPUhardware
plays an outsized role depending on the resolution of
medical images. Due to the largeGPUmemoryrequirements,
it is currently unable to fully integrate high-resolution MRIs
with an associated example size of 400 512 512 into
progressive convolutional neural network models. As a
result, 3D medical imaging data is either split into smaller
cuboid patches or processed slice by slice, similar to a series
of second photographs [10, 11, 23]. MIScnn cuts 3D medical
photos into patches with a customizable size (e.g. 128 128
128) by default to fully use the data from all three
dimensions. These patches will function in GPUs with RAM
capacities of 4–24 GB, which are commonly used in the
analysis, according to the model design. The slice-by-slice
second analysis is supported and may be used in MIScnn,
however, the 3D patch analysis is not. It's also possible to
combine the use of complete 3D images in the event that
you're studying unusually small medical images or have a
large GPU cluster. Second medical images are automatically
incorporated into convolutional neural networks and deep
learning models. Still, for images with high resolution, a
second patch-wise strategy is used.
4. Sampling and Batch Generation
4.1 Skipping blank patches
The well-known problem in medical images is large
unbalances between significant segments and the backdrop
results in a large number of parts that are strictly classified
as background and have no learning information. There'sno
use in multiplying these blank elements or patches,
especially when it comes to knowledge augmentation. As a
result, any patches that are completely classified as
background are omitted from the patch-wise model training
to minimize unnecessary fitting delays.
4.2 Batch Management
Sets of entire photos or patches are packaged into
batches after the data preparation and hence the optional
data augmentation for training. One batch comprises a large
number of ready images that areprocessedbythemodel and
GPU in a single phase. The neural network adjusts its
internal weights in accordance with the specified learning
rate for each batch or process step. The number of images
that may be stored in a single batch is highly dependent on
the amount of GPU memory available, and thereforemust be
appropriately planned in MIScnn. Each batch is stored in
memory so that it may be accessed at any time during the
training process. Due to the avoidance of reservecontinuous
batch preprocessing, this strategy dramatically decreases
computation time. MIScnn also permits "on-the-fly"
construction of the following batch in memory during the
runtime to overcome this limitation.
4.3 Batch Shuffling
At the end of each epoch, the order of batches, that is,the
area unit intended to be fitted andprocessedisshuffled.This
strategy lessens the risk of overfitting by reducing the
variance of the neural network duringfittingovera periodof
time. It should be emphasized, however, that just the batch
process sequence is shuffled, and the knowledge itself is not
sorted into a new batch order.
4.4 Multi-CPU and -GPU Support
In addition to GPU computing, MIScnn offers the use of
several GPUs and simultaneous central processor batch
loading. The storing ofalreadypreparedbatchesonmemory,
for example, allows for a fast and parallelizable processwith
central processors and GPU clusters by removing the risk of
batch preparation bottlenecks.
5. Deep Learning Model Creation
5.1 Model Architecture
The most important stage in a medical picture
segmentation pipeline is choosing a deep learning or
convolutional neural network model. There are many
distinct model architectures, each with its own set of
strengths and limitations. The MIScnn options provide an
open model interface that allows you to load and switch
between the several progressive convolutional neural
network models available, such as the widely used U-Net
model. Keras, an ASCII text file neural network framework
that provides a straightforward API for commonly used
neural network building blocks on top of TensorFlow, is

used to define models. The already established models, like
the Optimized High-Resolution Dense-U-Net model, are
significantly flexible by a determinable range of neurons,
custom input sizes, ex gratia dropout, and batch
normalization layers, or enhanced design versions. MIScnn
additionally provides 3D architectures, as well as 2D
architectures for medical image segmentation. The open
model interface enables for bespoke deep learning model
implementations and simple integration of these custom
models into the MIScnn pipeline, in addition to the freedom
in changing between previously imposed models.
5.2 Metrics
MIScnn has a large number of different metrics that may
be utilized for loss performance training,figureanalysis,and
manual performance analysis. One of the most often used
metrics for medical picture segmentation is the Dice
constant, also known as the Dice similarity index. It
calculates the degree of similarity between the expected
segmentation and the ground truth. False positives, on the
other hand, are penalized. There's a simple and class-wise
Dice constant implementation to enjoy the precision metric
reckoning on the segmentation categories (binary or multi-
class). Unlike the simple approach, which merely counts the
number of right and erroneous predictions, the category-
wise method accounts for prediction performance for each
segmentation class, which is strongly recommended for
typicallyclass-unbalancedmedical images.TheJaccardIndex
is another widely accepted statistic. Even though it's similar
to the Dice constant, it doesn't focus entirely on exact
segmentation.Itdoes,however,punishover-segmentationof
understanding. MIScnn, on the other hand, employs the
Tversky loss for training. The Tversky loss performance
solves knowledge imbalance and is admired by the Dice
constant. Even so, it achievesa farbetter balanceofprecision
and recall. As a result, Tversky loss performance provides
smart binary as well as multi-class segmentation
performance. In addition, many common metrics used in
Keras, such as accuracy and cross-entropy, are used in
MIScnn. MIScnn allows you to blend additional metrics for
training and analysis in addition to the already imposed
metrics or loss functions. As described in Keras, a custom
metric is enforced and readily supplied to the deep learning
model.
5.3 Model Utilization
The model will now be utilized for training on the
information to suit model weights or for prediction by
leveraging an already fitted model with the initialized deep
learning model and hence the totally preprocessed
information. Instead, the model will do AN analysis by using
cross-validationwithnumeroustrainingandpredictioncalls,
as an example. The model API allows you to save and load
models so that you may reprocess previously fitted models
for prediction or share pre-trained models.
5.4 Training
Various parameters should be establishedwhiletraining
a convolutional neural network or deep learning model. The
information augmentation options of the data set, which
have a significant impact on medical picture segmentation
training, should already be described in the pipeline.
Following that, the batch management configuration listed
the batch size settings as well as the batch shuffling at the
end of each epoch. As a result, simply the learning rate and
hence the number of epochs must be altered before the
training technique can be used. The training rate of a neural
network model is determined by the amount to which the
neural network model's prior weights are modified in each
iteration or epoch. The number of epochs, ontheotherhand,
determines the proportion of times the full knowledge set
will be fitted into the model. The resulting fitted model can
then be stored in memory. Because of the remaining fitting
time, the underlying Keras framework providesinsightsinto
the current model performance using specified metrics
throughout training. MIScnn also providestheuseofa fitting
evaluation questioning practicality, in which the fitting
scores and metrics are saved in a tab-separated file or
immediately plotted as a graph.
5.5 Prediction
Once trained,anassociatedegreepreviouslyfittedneural
network model will be utilized directly or it will beimported
from a file for segmentation prediction. For each panel, the
algorithm estimates a Sigmoid price for each category. This
panel's probability assessment for the related label is
represented by the Sigmoid price. TheargmaxoftheOne Hot
encoded category is then born-again to one result variable
holding the category with the best Sigmoid pricing, which is
known for multi-class segmentation challenges. MIScnn
provides two approaches for patches inside the prediction
after using the overlappingpatch-wiseanalysismethodology
during the training. Either the prediction approachdevelops
discrete patches and regards overlapping patches as solely
knowledge enhancement during the training,oroverlapping
patches are produced for prediction. Due to a lack of
prediction capacity at patch edges, generating a second
prediction for edge pixels in patches byvictimizingassociate
degree overlap may be an underutilized technique. Within
the overlapping section of two patches and with numerous
predictions, a merging method for the pixels is needed
within the subsequent merge of patches back to the basic
medical picture form. MIScnn estimates the mean of the
predicted Sigmoid values for each category in each
overlapping panel by default. The resulting picture matrix
with segmentation prediction, which has the sameformat as
the original medical image, is stored into a file structure
using the knowledge I/O interface supplied. The predicted
segmentation matrix is stored in NIfTI format with no
further information by default when using the NIfTI
knowledge I/O interface.

5.6 Evaluation
Multiple automatic analysis techniquesaresupportedby
MIScnn to investigate medical image segmentation
performance: k-fold cross-validation, leave-one-out cross-
validation, percentage-split validation, hold-out sets for
checking (data set split into test and plaything with a given
percentage), and elaborate validation in which it is
frequently given which pictures should be used for training
and testing. Apart from extensive validation, sampling is
used to create training and testing data sets in all analytical
methodologies. The specified measurements and loss
performance for the model are automatically planned in
figures during the study and recorded in tab-separated files
for easy examination. In addition to performance measures,
the constituent worth changes, and the frequency of
segmentation categories in medical images are frequently
examined as part of the MIScnn study. By producing picture
representations with segmentation overlays, the resulting
forecast is frequently compared directly to thebottomtruth.
The slices with the segmentation overlays are mechanically
depicted inside the Graphics Interchange Format (GIF) for
3D images, such as MRIs.
5.7 Convolutional Neural Network (U-net)
A contracting path (left side) and an expansive path
(right side) make up the network design (right side). The
convolutional network's contracting route follows the
standard architecture. It is made up of two 3x3 convolutions
(unpadded convolutions) that are applied repeatedly, each
followed by a rectified linear unit (ReLU) and a 2x2 max
pooling operation with stride 2 for downsampling. We
quadruple the number of feature channels with each
downsampling step. An upsampling of the feature map is
followed by a 2x2convolution("up-convolution")thathalves
the number of feature channels, a concatenation with the
proportionally cropped feature map from the contracting
path, and two 3x3 convolutions, each followed by a ReLU in
the expanding path. Due to the loss of boundary pixels in
every convolution, cropping isrequired.A1x1convolutionis
employed at the final layer to convert each 64-component
feature vector to the appropriate number of classes. The
network comprises a total of 23 convolutional layers. To
ensure that the output segmentation map tiles seamlessly
(see Figure 2), the input tile size should be chosen so that all
2x2 max-pooling operations are applied to a layer with an
even x- and y-dimension.
FIG 2 U-Net Architecture
6. Real-Life Use Cases
6.1 Kidney Tumor Segmentation
Computed Tomography scans of urinary organ tumors
from the urinary organ tumor Segmentation Challenge2019
information set showing the urinary organ (red)
and tumor (blue) segmentation as overlays. The
photographs show the segmentation variationsbetween the
bottom truth provided by the KiTS19
challenge and therefore the prediction from the quality 3D
U-Net models of our three-fold cross-validation.

6.2 Coronavirus disease 2019(Covid-19)
COVID-19 might be a global pandemic virus that spreads
quickly over the planet. RT-PCR (Reverse Transcription
enzyme Chain Reaction) is a commonly used test for
detecting COVID-19 infection. RT-PCRtestingisthatthegold
standard for COVID-19 testing, RT-PCR is an incredibly
complicated, long, and labor-intensive method, sparse
availability, and not correct. Chest X-ray may be used for the
initial screening of the COVID-19 in places having a shortage
of RT-PCR kits and is additional correct at diagnosing. Many
researchers have used deep learning to classify if the chest
infection is thanks to COVID-19 or different ailments.
6.2.1 PDCOVIDNET
The use of dilation to sight dominant options withinthe
image was explored. The authors planned a parallel
expanded CNN model. The expanded module concerned the
skipping of pixels throughout the convolution method.
Parallel CNN branches square measure proposed with
completely different dilation rates. The results of parallel
branches were combined and sent into the next convolution
layer. The concatenation-convolution process was used to
investigate feature relationships in enlarged convolutions,
resulting in visually prominent classification alternatives.
The model conjointly used Grad-CAM and Grad-CAM++ to
highlight theregionsofclass-discriminative noticeablemaps.
The performance metrics used were accuracy, precision,
recall, F1-score with ROC/AUC, and are 96.58%, 95%, 91%,
93%, and 99.1% severally.
7. Future Scope
Multiple essential aspects are now the focus of active
MIScnn development: Adding more knowledge I/O
interfaces for the most widely used medical image formats,
such as DICOM, expanding preprocessing and knowledge
augmentation procedures, and implementing a variety of
cost-effective patch skipping approaches rather than
ignoring each blank patch (e.g. denoising patch skipping)
Implementation of an open interface for bespoke
preprocessing algorithms for certain picture types, such as
MRIs. The MIScnn road map also contains a model library
expansion with a large number of progressive deep learning
models for medical image segmentation, in addition to the
scheduled feature implementations. Furthermore, an
objective comparison of the U-Net model version selection
has been made available in order to encourage a lot of
insights on alternative model performances using the same
pipeline. Contributions from the community in terms of
implementations or critique are encouraged and might be
included in the assessment. MIScnn currently has a strong
pipeline for medical image segmentation; however, it will
continue to be upgraded and extended in the future.
8. Results
CNN models are delineated as black boxes and there is a
great deal of analysis happening in terms of analyzing and
understanding output at each layer. Since medical images
square measure concerned, we'd like an associate degree
responsible and economical prediction system that ought to
even be able to articulate about a call taken. Image
captioning is being done by researchers (textual
representations of the image). This can change physicians to
grasp the perception of the network at eachoutputlayerand
intermediate levels. Researchers have tried theorem deep
learning models that calculate the uncertainty estimates.
This would facilitate physicians to assess the model. All of
these might help clinicians analyze medical images faster by
employing CNNs. Here, we tend to analyze and measure
information from the urinary organ tumor Segmentation
Challenge 2019 exploitation MIScnn. The main goal of this
experiment is to demonstrate MIScnn's 'out-of-the-box'
performance without doing extensive and lengthy
optimizations on the data set or the medical anomaly. The
scripts in the Appendix were used to acquire all of the
findings.
9. Conclusion
The ASCII text file Python package MIScnn: A framework
for medical image segmentation with convolutional neural
networks and deep learning was introduced in this
publication. The user-friendly API enables rapid creation of
medical image segmentation pipeline, as well as knowledge
I/O, preprocessing, knowledge augmentation, patch-wise
analysis, metrics, a library of progressive deep learning
models, and model usages such as training, prediction, and
completely automated analysis (e.g. cross-validation).Users
may entirely personalize the pipeline because of its high
FIG 3 Kidney Tumor Segmentation

configurability and many open interfaces. Researchers may
use this framework to replace a complete medical image
segmentation process with just a few lines of code. We're
going to test the MIScnn practicality by conducting
automatic cross-validation on the excretoryorgan neoplasm
Segmentation Challenge 2019 CT knowledge set, which will
result in a robust predictor. We expect that it will make it
easier to move medical picture segmentationfromanalytical
laboratories to practical applications.
10. Acknowledgement
This Paper and solution it provide would not be possible
without the help our professor at D.Y Patil Institute of
Engineering and Technology, ambi and the owner of the
reviewed technology. We want to thank Profs. Rohini
Hanchate for sharing their resourceswithuswhichwasused
for this work. We also would like to thank Dominik Müller
and Frank Kramer for their technology.
11. References
1. Aggarwal P, Vig R, Bhadoria S, Dethe CG. Role of
segmentation in medical imaging: a comparative study. Int J
Comput Appl. 2011;29:54–61.
2. Gibelli D, Cellina M, Gibelli S, Oliva AG, Termine G,
Pucciarelli V, et al. Assessing symmetry of zygomatic bone
through three-dimensional segmentation on computed
tomography scan and “mirroring” procedure: a contribution
for reconstructive maxillofacial surgery. J Cranio-Maxillofac
Surg. 2018;46:600–4. https
://doi.org/10.1016/j.jcms.2018.02.012.
3. Cellina M, Gibelli D, Cappanella A, Toluian T, Pittino CV,
Carlo M, et al. Segmentation procedures for the assessment
of paranasal sinuses volumes. Neuroradiol J. 2020. https
://doi.org/10.1177/19714 00920 94663 5.
4. Hu X, Luo W, Hu J, Guo S, Huang W, Scott MR, et al. Brain
SegNet: 3D local refinement network for brain lesion
segmentation. BMC Med Imaging. 2020;20:17. https
://doi.org/10.1186/s1288 0-020-0409-2.
5. Sun R, Wang K, Guo L, Yang C, Chen J, Ti Y, et al. A potential
field segmentation basedmethodfortumorsegmentation on
multi-parametric MRI of glioma cancer patients. BMC Med
Imaging. 2019;19:48. https ://doi. org/10.1186/s1288 0-
019-0348-y.
6. Claudia C, Farida C, Guy G, Marie-Claude M, Carl-Eric A.
Quantitative evaluation of an automatic segmentation
method for 3D reconstruction of intervertebral scoliotic
disks from MR images. BMC Med Imaging. 2012;12:26.https
://doi.org/10.1186/1471-2342-12-26.
7. Guo Y, Liu Y, Georgiou T, Lew MS. A review of semantic
segmentation using deep neural networks.Int JMultimedInf
Retr. 2018;7:87–93. https ://doi.org/10.1007/s13735-017-
0141-z.
8. Anwar SM, Majid M, Qayyum A, Awais M, Alnowami M,
Khan MK. Medical image analysis usingconvolutional neural
networks: a review. J Med Syst. 2018;42:226. https
://doi.org/10.1007/s1091 6-018-1088-1.
9. Wang G. A perspective on deep imaging. IEEE Access.
2016;4:8914–24.
10. Litjens G, Kooi T, Bejnordi BE, Setio AAA, Ciompi F,
Ghafoorian M, et al. A survey on deep learning in medical
image analysis. Med Image Anal. 2012;2017(42):60–88.
11. Shen D, Wu G, Suk H-I. Deep learning in medical image
analysis. Annu Rev Biomed Eng. 2017;19:221–48. https
://doi.org/10.1146/annur ev-bioen g-07151 6-04444 2.
12. Ronneberger O, Fischer P, Brox T. U-Net: convolutional
networks for biomedical image segmentation. Lect Notes
Comput Sci (includingSubserLectNotesArtifIntellLectNotes
Bioinformatics). 2015;9351:234–41.
13. Zhou Z, Siddiquee MMR, Tajbakhsh N, Liang J. UNet++: a
nested U-Net architecture for medical image segmentation.
2018. http://arxiv .org/ abs/1807.10165 . Accessed 19 Jul
2019.
14. Isensee F, Petersen J, Klein A, Zimmerer D, Jaeger PF,
Kohl S, et al. nnUNet: self-adapting framework for U-Net-
based medical image segmentation. 2018. http://arxiv
.org/abs/1809.10486 . Accessed 19 Jul 2019.
15. Kolařik M, Burget R, Uher V, Řiha K, Dutta M. Optimized
high resolution 3D dense-U-Net network for brainandspine
segmentation. Appl Sci. 2019;9:404.
16. Cicek O, Abdulkadir A, LienkampSS,BroxT,Ronneberger
O. 3D U-net: learning dense volumetric segmentation from
sparse annotation. Lect Notes Comput Sci. 2016;9901:424–
32.
17. Lee K, Zung J, Li P, Jain V, Seung HS. Superhuman
accuracy on the SNEMI3D connectomics challenge. 2017;
Nips:1–11. http://arxiv .org/ abs/1706.00120 .

MIScnn: A Framework for Medical Image Segmentation with Convolutional Neural Networks and Deep Learning

Recommended

More Related Content

Similar to MIScnn: A Framework for Medical Image Segmentation with Convolutional Neural Networks and Deep Learning (20)

More from IRJET Journal (20)

Recently uploaded (20)

MIScnn: A Framework for Medical Image Segmentation with Convolutional Neural Networks and Deep Learning