<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Madhumita Dange on Medium]]></title>
        <description><![CDATA[Stories by Madhumita Dange on Medium]]></description>
        <link>https://medium.com/@madhumita-dange?source=rss-a62937b1239b------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*7BHEbGS0GRg7svjX</url>
            <title>Stories by Madhumita Dange on Medium</title>
            <link>https://medium.com/@madhumita-dange?source=rss-a62937b1239b------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 14:22:28 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@madhumita-dange/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Reconstruct Source Galaxy Images from Strong Gravitational Lens Images using U-Net]]></title>
            <link>https://medium.com/data-science/reconstruct-source-galaxy-images-from-strong-gravitational-lens-images-using-u-net-1d8221100601?source=rss-a62937b1239b------2</link>
            <guid isPermaLink="false">https://medium.com/p/1d8221100601</guid>
            <category><![CDATA[astronomy]]></category>
            <category><![CDATA[convolution-neural-net]]></category>
            <category><![CDATA[gravitational-lensing]]></category>
            <category><![CDATA[unet]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Madhumita Dange]]></dc:creator>
            <pubDate>Tue, 15 Jun 2021 03:03:27 GMT</pubDate>
            <atom:updated>2021-06-18T04:36:52.015Z</atom:updated>
            <content:encoded><![CDATA[<h3>Reconstruct Source Galaxy Images from Strong Gravitational Lens Images Using U-Net</h3><h4>How to generate un-lensed images of galaxies far, far away</h4><p>Gravitational pull bends the light on distant galaxies, which causes them to appear lensed. The following image shows us how light from a distant quasar bends due to strong gravitational pull of a lensing galaxy which acts similar to an optical lens in between the Quasar (source) and the observer. This creates a distorted image of quasar shown in right side of image.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2ZUvm_GmuWJAy2aCddbz5g.png" /><figcaption><strong>A simple illustration of strong gravitational lensing</strong> <strong>|</strong> <strong>Diagram by Author, Galaxy Images</strong> : Mandelbaum, Rachel, Lackner, Claire, Leauthaud, Alexie, &amp; Rowe, Barnaby. (2012). COSMOS real galaxy dataset [Data set]. Zenodo. <a href="http://doi.org/10.5281/zenodo.3242143">http://doi.org/10.5281/zenodo.3242143</a></figcaption></figure><p>This project tries to reconstruct the original image of the distant sources distorted by gravitational lenses.</p><p>Using <a href="https://en.wikipedia.org/wiki/Convolutional_neural_network">Convolutional Neural Networks</a>, <a href="https://en.wikipedia.org/wiki/U-Net">U-Net</a>, and Mean Squared Error with <a href="https://en.wikipedia.org/wiki/Structural_similarity">Structural Similarity Index</a>, we were able to generate un-lensed images of these galaxies. i.e. Undoing effect of <a href="https://en.wikipedia.org/wiki/Strong_gravitational_lensing">strong gravitational lensing</a> and generating source image . Also, we are able to predict lensing parameters of the strong gravitational lens which causes this distortion. We worked on developing an algorithm to undo this visual effect so that we can detect other possible unidentified gravitational lenses. i.e. reconstructing images of the source galaxy.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*0Ycmd9ylLQdsWK-n71j1EA.png" /><figcaption><strong>Undoing Strong Gravitational lensing Effect |</strong> <strong>Diagram by Author, Galaxy Images</strong> : Mandelbaum, Rachel, Lackner, Claire, Leauthaud, Alexie, &amp; Rowe, Barnaby. (2012). COSMOS real galaxy dataset [Data set]. Zenodo. <a href="http://doi.org/10.5281/zenodo.3242143">http://doi.org/10.5281/zenodo.3242143</a></figcaption></figure><h3>Dataset</h3><p>As not many images of strong gravitational lenses are available, We decided to simulate data using the <a href="https://zenodo.org/record/3242143">Cosmos dataset</a>. Cosmos dataset has 60K galaxy images. We used redshift properties to filter out galaxies to use for our project. After filtering galaxies, we used <a href="https://github.com/sibirrer/lenstronomy">Lenstronomy</a> and randomly generated lens parameters to simulate lensed images of the source galaxies. We also used rotation <a href="https://github.com/JelleAalbers/minilensmaker">data augmentation</a> to generate more simulated data. Our dataset is made up of generated lensed images and its source real galaxy pairs. These Images are 64x64 grayscale images. They have been distributed into test, validation, and training datasets based on galaxy’s distinct source id to avoid data leak between the sets. The lensed images are used to predict the lensing parameters of gravitational lens for a galaxy. The un-lensed images are the correct depiction of the same galaxy used to predict the un-lensed galaxy given the lensed image.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*og5NdH4Bd9lJ-Tl4OVSIvQ.png" /><figcaption><strong>Data Augmentation to create simulated strong gravitational lensing data | Diagram by Author,</strong> <strong>Galaxy Images</strong> : Mandelbaum, Rachel, Lackner, Claire, Leauthaud, Alexie, &amp; Rowe, Barnaby. (2012). COSMOS real galaxy dataset [Data set]. Zenodo. <a href="http://doi.org/10.5281/zenodo.3242143">http://doi.org/10.5281/zenodo.3242143</a></figcaption></figure><h3>Approach</h3><p>We broke our problem down into two steps.</p><p>First, predict 8 <a href="https://www.nature.com/articles/nature23463">lensing parameters</a>:</p><ul><li><a href="https://en.wikipedia.org/wiki/Einstein_radius">Einstein radius</a> (theta_E)</li><li>Exponent of the lens’s power-law mass distribution (gamma)</li><li>Ellipticity components of the lens (e1,e2)</li><li>Location of the lens in the image (center_x,center_y) in arcseconds</li><li>External shear components (gamma1,gamma2) of lensed image.</li></ul><p>Second, U-net generates the un-lensed source galaxy image from its lensed pair image.</p><h3>Architecture Part 1: Predicting lensing parameters</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*AfD-cvpl1MWYphVR0ETm0Q.png" /><figcaption><strong>Predicting lensing parameters | Diagram by Author, Galaxy Images</strong> : Mandelbaum, Rachel, Lackner, Claire, Leauthaud, Alexie, &amp; Rowe, Barnaby. (2012). COSMOS real galaxy dataset [Data set]. Zenodo. <a href="http://doi.org/10.5281/zenodo.3242143">http://doi.org/10.5281/zenodo.3242143</a></figcaption></figure><p>After we preprocessed our images, we used our training dataset in a convolutional neural network with three blocks, using the corresponding lensing parameters as the output of the last layer, which produced our model for part 1. In each block, we applied 2D Convolution, Max Pooling, and Batch Normalization, and a Relu activation function. Finally, we flattened the layers and passed the input into two dense layers, one Relu and the other a Linear function. We assigned all 8 parameters to be predicted as our labels in each dataset, then ran the function on all the images in the datasets. It is able to predict the lensing parameters with good metrics across all three datasets.</p><p><strong>Hyperparameters of Part 1</strong>: We tuned the learning rate for Adam, number of layers for our CNN, mini batch size, and epochs.</p><pre>model = models.Sequential()<br>model.add(layers.Conv2D(32, (3, 3),input_shape=(64, 64, 1),use_bias=<strong>False</strong>)) <br>model.add(layers.MaxPooling2D((2, 2))) model.add(layers.BatchNormalization()) model.add(layers.Activation(&quot;relu&quot;))<br>model.add(layers.Conv2D(64, (3, 3) ,use_bias=<strong>False</strong>)) model.add(layers.MaxPooling2D((2, 2))) model.add(layers.BatchNormalization()) model.add(layers.Activation(&quot;relu&quot;))<br>model.add(layers.Conv2D(64, (3, 3),use_bias=<strong>False</strong>)) model.add(layers.MaxPooling2D((2, 2))) model.add(layers.BatchNormalization()) model.add(layers.Activation(&quot;relu&quot;)) model.add(layers.Flatten()) model.add(layers.Dense(64, activation=&#39;relu&#39;)) model.add(layers.Dense(8,activation=&#39;linear&#39;))</pre><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*AQkwrLZiOULOoYoRB5z9Xw.png" /><figcaption>CNN to Predict Lensing parameters</figcaption></figure><p>Metric of “Root Mean squared error” Output per Lensed Parameter Class for Test Set:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*gHbNTS56JtUilHyh" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/884/0*rxCn57sEEe5pBPGX" /><figcaption>RMS per Output Class Parameter, <strong>Diagram by Author,</strong></figcaption></figure><p>This shows that theta_E (Einstein’s radius) is easier to predict than lensed center coordinates and gamma, but harder than the lens ellipticity components (e1, e2).</p><h3>Architecture Part 2: Generating source galaxy image</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*H6JPchdXOGtGpzSbjLW-Iw.png" /><figcaption><strong>lensed image to an un-lensed source galaxy image using U-Net</strong> <strong>| Diagram by Author,</strong> <strong>Galaxy Images</strong> : Mandelbaum, Rachel, Lackner, Claire, Leauthaud, Alexie, &amp; Rowe, Barnaby. (2012). COSMOS real galaxy dataset [Data set]. Zenodo. <a href="http://doi.org/10.5281/zenodo.3242143">http://doi.org/10.5281/zenodo.3242143</a></figcaption></figure><p>In part 2, we converted a lensed image to an un-lensed source galaxy image. For this we used our U-net encoder decoder network and image similarity for our error/loss function.</p><h4>U-Net Experiments and hyperparameter tuning</h4><p>We started a simple <a href="https://en.wikipedia.org/wiki/U-Net">U-Net </a>with multiple Conv2D, max pooling,<a href="https://naokishibuya.medium.com/up-sampling-with-transposed-convolution-9ae4f2df52d0"> transpose convolution layers</a>, and linear activation for the last layer with MSE loss and Adam optimizer. After that, we added more hidden blocks in U-Net that improved its performance. Then we experimented with different activation functions: relu, elu, etc. Relu gave us better results for intermediate layers. Then we added dropouts which did not improve results much so removed it. Also, adding batch normalization helped U-Net. Still it was not generating better images for spiral galaxies from manual error analysis, so we decided to use different image similarity loss functions.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/749/0*YHtFJhHfGvErliWD" /><figcaption>Initial Model</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/743/0*yynIsoaDhYdxobv_" /><figcaption><strong>Intermediate model with mse generating correct shapes</strong> <strong>|Galaxy Images</strong> : Mandelbaum, Rachel, Lackner, Claire, Leauthaud, Alexie, &amp; Rowe, Barnaby. (2012). COSMOS real galaxy dataset [Data set]. Zenodo. <a href="http://doi.org/10.5281/zenodo.3242143">http://doi.org/10.5281/zenodo.3242143</a></figcaption></figure><p><strong>Metrics and Custom Loss functions :</strong></p><p>We experimented with image structural similarity SSIM index, Peak signal to noise ratio (PSNR), and mean squared error (MSE) as loss functions. SSIM gave us better results. We further improved our results by writing a custom loss function which is a combination of SSIM and MSE.</p><h4>U-Net with Custom loss function</h4><pre>inputs = Input((64, 64, 1))</pre><pre>s = Lambda(lambda x: x / 255) (inputs)<br>c1 = Conv2D(16, (3, 3),  padding=&#39;same&#39;,use_bias=False) (s)</pre><pre>c1 = BatchNormalization()(c1)<br>c1 = Activation(&#39;relu&#39;)(c1)<br>c1 = Conv2D(16, (3, 3), padding=&#39;same&#39;,use_bias=False) (c1)<br>c1 = BatchNormalization()(c1)<br>c1 = Activation(&#39;relu&#39;)(c1)<br>p1 = MaxPooling2D((2, 2)) (c1)</pre><pre>c2 = Conv2D(32, (3, 3),  padding=&#39;same&#39;,use_bias=False) (p1)<br>c2 = BatchNormalization()(c2)<br>c2 = Activation(&#39;relu&#39;)(c2)<br>c2 = Conv2D(32, (3, 3), padding=&#39;same&#39;,use_bias=False) (c2)<br>c2 = BatchNormalization()(c2)<br>c2 = Activation(&#39;relu&#39;)(c2)<br>p2 = MaxPooling2D((2, 2)) (c2)</pre><pre>c3 = Conv2D(64, (3, 3), padding=&#39;same&#39;,use_bias=False) (p2)<br>c3 = BatchNormalization()(c3)<br>c3 = Activation(&#39;relu&#39;)(c3)<br>c3 = Conv2D(64, (3, 3), padding=&#39;same&#39;,use_bias=False) (c3)<br>c3 = BatchNormalization()(c3)<br>c3 = Activation(&#39;relu&#39;)(c3)<br>c3 = Conv2D(64, (3, 3), padding=&#39;same&#39;,use_bias=False) (c3)<br>c3 = BatchNormalization()(c3)<br>c3 = Activation(&#39;relu&#39;)(c3)<br>p3 = MaxPooling2D((2, 2)) (c3)</pre><pre>c4 = Conv2D(128, (3, 3), padding=&#39;same&#39;, use_bias=False) (p3)<br>c4 = BatchNormalization()(c4)<br>c4 = Activation(&#39;relu&#39;)(c4)<br>c4 = Conv2D(128, (3, 3), padding=&#39;same&#39;,use_bias=False) (c4)<br>c4 = BatchNormalization()(c4)<br>c4 = Activation(&#39;relu&#39;)(c4)<br>p4 = MaxPooling2D(pool_size=(2, 2)) (c4)</pre><pre>c5 = Conv2D(256, (3, 3), padding=&#39;same&#39;,use_bias=False) (p4)<br>c5 = BatchNormalization()(c5)<br>c5 = Activation(&#39;relu&#39;)(c5)<br>c5 = Conv2D(256, (3, 3),  padding=&#39;same&#39;,use_bias=False) (c5)<br>c5 = BatchNormalization()(c5)<br>c5 = Activation(&#39;relu&#39;)(c5)</pre><pre>u6 = Conv2DTranspose(128, (2, 2), strides=(2, 2), padding=&#39;same&#39;) (c5)<br>u6 = concatenate([u6, c4])<br>c6 = Conv2D(128, (3, 3),  padding=&#39;same&#39;,use_bias=False) (u6)<br>c6 = BatchNormalization()(c6)<br>c6 = Activation(&#39;relu&#39;)(c6)<br>c6 = Conv2D(128, (3, 3),  padding=&#39;same&#39;,use_bias=False) (c6)<br>c6 = BatchNormalization()(c6)<br>c6 = Activation(&#39;relu&#39;)(c6)</pre><pre>u7 = Conv2DTranspose(64, (2, 2), strides=(2, 2), padding=&#39;same&#39;) (c6)<br>u7 = concatenate([u7, c3])<br>c7 = Conv2D(64, (3, 3),  padding=&#39;same&#39;,use_bias=False) (u7)<br>c7 = BatchNormalization()(c7)<br>c7 = Activation(&#39;relu&#39;)(c7)<br>c7 = Conv2D(64, (3, 3),  padding=&#39;same&#39;,use_bias=False) (c7)<br>c7 = BatchNormalization()(c7)<br>c7 = Activation(&#39;relu&#39;)(c7)<br>c7 = Conv2D(64, (3, 3),  padding=&#39;same&#39;,use_bias=False) (c7)<br>c7 = BatchNormalization()(c7)<br>c7 = Activation(&#39;relu&#39;)(c7)</pre><pre>u8 = Conv2DTranspose(32, (2, 2), strides=(2, 2), padding=&#39;same&#39;) (c7)<br>u8 = concatenate([u8, c2])<br>c8 = Conv2D(32, (3, 3), padding=&#39;same&#39;,use_bias=False) (u8)<br>c8 = BatchNormalization()(c8)<br>c8 = Activation(&#39;relu&#39;)(c8)<br>c8 = Conv2D(32, (3, 3), padding=&#39;same&#39;,use_bias=False) (c8)<br>c8 = BatchNormalization()(c8)<br>c8 = Activation(&#39;relu&#39;)(c8)</pre><pre>u9 = Conv2DTranspose(16, (2, 2), strides=(2, 2), padding=&#39;same&#39;) (c8)<br>u9 = concatenate([u9, c1], axis=3)<br>c9 = Conv2D(16, (3, 3),  padding=&#39;same&#39;,use_bias=False) (u9)<br>c9 = BatchNormalization()(c9)<br>c9 = Activation(&#39;relu&#39;)(c9)<br>c9 = Conv2D(16, (3, 3), padding=&#39;same&#39;,use_bias=False) (c9)<br>c9 = BatchNormalization()(c9)<br>c9 = Activation(&#39;relu&#39;)(c9)</pre><pre>out = Conv2D(1, (1, 1), use_bias=False) (c9)<br>out = BatchNormalization()(out)<br>outputs = Activation(&#39;linear&#39;)(out)</pre><pre># Optimizer<br>optimizer = tf.keras.optimizers.Adam(learning_rate=0.001)</pre><pre>model = Model(inputs=[inputs], outputs=[outputs])</pre><pre>model.compile(loss=mse_ssim_loss, optimizer=optimizer, metrics=[&#39;mse&#39;,ssim_loss,psnr_loss])</pre><pre>model.summary()</pre><h3>Custom loss function which is a combination of SSIM and MSE</h3><pre>def mse_ssim_loss(y_true, y_pred):<br>return tf.reduce_mean(tf.math.squared_difference(y_true, y_pred)) - tf.reduce_mean(tf.image.ssim(y_true, y_pred, 1.0))</pre><p>This loss function is generating better results for spiral galaxies as well.</p><p>val_mse: 0.0023 — val_ssim_loss: 0.0468 — val_psnr_loss: 30.6308. Loss curves.</p><p><strong>U-Net Results on Test Set Improved Model: </strong>more results below</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/743/0*44FaMmVWRrrHk4mU" /><figcaption><strong>Improved model with custom loss function generating patterns as well |</strong> <strong>Galaxy Images</strong> : Mandelbaum, Rachel, Lackner, Claire, Leauthaud, Alexie, &amp; Rowe, Barnaby. (2012). COSMOS real galaxy dataset [Data set]. Zenodo. <a href="http://doi.org/10.5281/zenodo.3242143">http://doi.org/10.5281/zenodo.3242143</a></figcaption></figure><h3>Insights and Future work:</h3><p>We tried to generate a source from our trained model for original strong gravitational lensed images from the internet as well, but we need more preprocessing and noise removal for that data. We will also try gaussian noise removal as well as augmenting more data (Data based approach). We will also experiment with RGB images and continue to refine our models for both parts. So we can get closer to predict how these source galaxies actually look like. Also, we will make simulated data publicly available.</p><h3>Contributions</h3><p>Git Repository : <a href="https://github.com/madhumitadange/Generate-Source-Galaxy-Images-from-Strong-Gravitational-Lens-Images">Generate-Source-Galaxy-Images-from-Strong-Gravitational-Lens-Images</a></p><p>Data : <a href="https://drive.google.com/drive/folders/1av08dbI44aa5R2LmY8hXxSSZbEaUuh8U?usp=sharing">lensingData</a></p><h3>Acknowledgements</h3><p>My team member<a href="https://profiles.stanford.edu/anand-bhavsar"> Anand Bhavsar</a> and I would like to thank our mentor Mr. <a href="https://profiles.stanford.edu/jelle-aalbers">Jelle Aalbers</a>, Kavli Postdoctoral Fellow at Stanford University, domain expert in astrophysics for the idea, helping us set up our data, providing guidance and advice to build our models, and giving feedback on our test results.</p><h3>More Results:</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/749/0*9Rl2TIjZK9LMySTF" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/743/0*zPTFqzqS4JUDtTEB" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/743/0*vcizXYriDxdbpkV2" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/743/0*SNE5L6giQDsSyWoV" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/743/0*d1UK4-pGtRkis4pb" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/743/0*L8IrTBLjzXtPYlVj" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/743/0*XCBqSdTiYym-78Ds" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/743/0*h39YUBGoAHBH0Bar" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/743/0*86uPqd4pHU9-h35b" /><figcaption><strong>Sample Output Images |</strong> <strong>Galaxy Images</strong> : Mandelbaum, Rachel, Lackner, Claire, Leauthaud, Alexie, &amp; Rowe, Barnaby. (2012). COSMOS real galaxy dataset [Data set]. Zenodo. <a href="http://doi.org/10.5281/zenodo.3242143">http://doi.org/10.5281/zenodo.3242143</a></figcaption></figure><h3>References:</h3><p><a href="https://en.wikipedia.org/wiki/Strong_gravitational_lensing">Strong gravitational lensing - Wikipedia</a></p><p><a href="https://zenodo.org/record/3242143">COSMOS real galaxy dataset</a></p><ul><li><a href="https://www.nature.com/articles/nature23463">Fast automated analysis of strong gravitational lenses with convolutional neural networks</a></li><li><a href="https://github.com/sibirrer/lenstronomy">sibirrer/lenstronomy</a></li></ul><p><a href="https://en.wikipedia.org/wiki/Structural_similarity">Structural Similarity</a> (SSIM)</p><p><a href="https://en.wikipedia.org/wiki/Peak_signal-to-noise_ratio">Peak Signal to Noise Ratio</a> (PSNR)</p><ul><li><a href="https://kipac.stanford.edu/">Welcome | Kavli Institute for Particle Astrophysics and Cosmology (KIPAC)</a></li><li><a href="https://arxiv.org/abs/2002.01479">What does strong gravitational lensing? The mass and redshift distribution of high-magnification lenses</a></li><li><a href="http://spiff.rit.edu/classes/phys240/lectures/grav_lens/grav_lens.html">Gravitational Lensing Theory</a></li></ul><iframe src="https://drive.google.com/viewerng/viewer?url=https%3A//arxiv.org/pdf/1511.08861.pdf&amp;embedded=true" width="600" height="780" frameborder="0" scrolling="no"><a href="https://medium.com/media/a172fbb40d69683609e3a31a001492e3/href">https://medium.com/media/a172fbb40d69683609e3a31a001492e3/href</a></iframe><ul><li><a href="https://naokishibuya.medium.com/up-sampling-with-transposed-convolution-9ae4f2df52d0">Up-sampling with Transposed Convolution</a></li><li><a href="https://up42.com/blog/tech/image-similarity-measures">Comparing Ground Truth With Predictions Using Image Similarity Measures</a></li><li><a href="https://towardsdatascience.com/understanding-semantic-segmentation-with-unet-6be4f42d4b47">Understanding Semantic Segmentation with UNET</a></li><li><a href="https://towardsdatascience.com/nucleus-segmentation-using-u-net-eceb14a9ced4">Nucleus Segmentation using U-Net</a></li><li><a href="https://www.tensorflow.org/tutorials/images/">Convolutional Neural Network (CNN) | TensorFlow Core</a></li><li><a href="https://www.nature.com/articles/nature23463">Fast automated analysis of strong gravitational lenses with convolutional neural networks</a></li><li><a href="https://hubblesite.org/resource-gallery/images">Images</a></li><li><a href="https://www.freepik.com/">Freepik | Graphic Resources for everyone</a></li><li><a href="https://www.flaticon.com/">Free Icons and Stickers - Millions of resources to download</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1d8221100601" width="1" height="1" alt=""><hr><p><a href="https://medium.com/data-science/reconstruct-source-galaxy-images-from-strong-gravitational-lens-images-using-u-net-1d8221100601">Reconstruct Source Galaxy Images from Strong Gravitational Lens Images using U-Net</a> was originally published in <a href="https://medium.com/data-science">TDS Archive</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>