Deep Networks
Discussion For all training set sizes in Figure 4, the model trained
using data augmentation achieved a lower dice loss and a higher accuracy than the model trained without it. Tis sug- gests that augmentation is playing a significant role in forcing the model to learn generalizable information about the differ- ence between cell cytoplasm and the nuclear envelope. Figure 5 shows that even for the smallest number of train-
ing examples the deep network can either provide a crude localization of the nuclei or a skeletonized outline, assuming strong contrast for the nuclear membrane over the cytoplasm. Te segmentation for the 150 training examples and above is strikingly similar to the hand segmented masks, demonstrat- ing the power of deep learning for complex image processing tasks. Te result from such fully trained networks requires no additional processing before use, allowing the model to operate as a “one-stop-shop” for end-to-end image process- ing. Te data processing and machine learning strategy in this article may accelerate future work on more complicated image segmentation tasks, where even fewer training examples are available. One complicating factor in this work is the heterogeneous
nature of the cheek cell data-set, which is composed of many shapes of cells that present morphologically distinct nuclei, as seen in Figure 3. It is expected that for less complicated image processing tasks, the number of training images may be con- siderably fewer, since each image contains much more infor- mation regarding the population of example nuclei. In such cases where there exist several distinct “classes” of
images
within one dataset, it may be advantageous to split the data along such lines to simplify the training process. Future work will examine the relative tradeoffs between
training a single model on a large heterogeneous dataset ver- sus several smaller homogenous classes. Overall, modern encoder-decoder architectures appear to be robust models for image processing tasks. Since the prediction takes place in milliseconds, these models make attractive solutions for handling the deluge of data currently emerging from modern microscopes. It is expected that as awareness of deep learning methods spread within the microscopy community, standards for recording and processing data will further leverage their scalability.
Conclusion Tis article further establishes that automated segmen-
tation of micrograph image features is feasible. Methods are shown for improving deep learning as an image processing framework for biological imaging. Te use of image augmenta- tion, in which a small number of hand-labeled images is trans- formed into a larger set through image transformations, allows strong performance even with the small numbers of images common to research environments. Such methods are not lim- ited to TEM imaging; they should be equally applicable to X-ray microscopy and light microscopy. Tis work may inspire the application of deep learning methods in the imaging commu- nity in situations where conventional methods struggle.
Acknowledgments Te author would like to acknowledge the help of Karl
Hujsak and Yue Li of Northwestern University for their men- torship and guidance with the preparation of this manuscript.
References [1] F Garma et al., Journal of Clinical Engineering 38(2) (2013) 79–83.
[2] Y Zheng et al., IEEE T Med Imaging 27(11) (2008) 1668–81. [3] A Coates and A Ng, “Learning Feature Representations with K-Means” in Neural Networks: Tricks of the Trade, eds. Montavon et al., Springer, Berlin Heidelberg, 2012, 561–80.
[4] N Pinto et al., PLoS Comput Biol 4(1) (2008) e27. [5] M Everingham et al., Int J Comput Vision 88(2) (2010) 303–38.
[6] J Deng et al. Computer Vision and Pattern Recognition 2009 (2009) 248–55.
[7] O Ronneberg et al., International Conference on Medi- cal image computing and computer-assisted intervention, Springer International Publishing, Basel, Switzerland, 2015, 234–41.
[8] K He et al., Proceedings of the IEEE conference on com- puter vision and pattern recognition, IEEE, Piscataway, NJ, 2016, 770–78.
[9] Z Zhang et al., IEEE Geosci Remote S 15(5) (2018) 749–53. [10] D Kingma and J Ba, arXiv preprint arXiv:1412.6980 (2014).
Precision, Speed, Stability NANO-POSITIONING FOR MICROSCOPY
NEW
PInano®
II, XYZ piezo stage w/ advanced controller
Piezo Focus package w/ advanced controller
Miniature piezo motors
PI (Physik Instrumente) LP ·
www.pi.ws/mi ·
info@pi-usa.us · 508-832-3456 2019 January •
www.microscopy-today.com 23
FSM fast beam steering
Compact 6-axis positioner
Page 1 |
Page 2 |
Page 3 |
Page 4 |
Page 5 |
Page 6 |
Page 7 |
Page 8 |
Page 9 |
Page 10 |
Page 11 |
Page 12 |
Page 13 |
Page 14 |
Page 15 |
Page 16 |
Page 17 |
Page 18 |
Page 19 |
Page 20 |
Page 21 |
Page 22 |
Page 23 |
Page 24 |
Page 25 |
Page 26 |
Page 27 |
Page 28 |
Page 29 |
Page 30 |
Page 31 |
Page 32 |
Page 33 |
Page 34 |
Page 35 |
Page 36 |
Page 37 |
Page 38 |
Page 39 |
Page 40 |
Page 41 |
Page 42 |
Page 43 |
Page 44 |
Page 45 |
Page 46 |
Page 47 |
Page 48 |
Page 49 |
Page 50 |
Page 51 |
Page 52
