{"cells":[{"metadata":{},"cell_type":"markdown","source":"
\n

PANDA: Image Data Augmentation Techniques w/ MONAI

\n

Hello there!\n\nThis is a notebook of mine whose aim is to find out suitable preprocessing methods for this competition on prostate cancer detection. Prostate cancer is a deadly form of cancer which affects the *prostate gland* in men. Here I try to describe the best possible methods for preprocessing the images in the competition and I'll try to cover as many images as well as preprocessing/data augmentation techniques that are possible in this notebook.

\n
\n

I shall try to cover a lot of methods of preprocessing/data augmentation and i'll also try to use a variety of libraries including MONAI which is a very useful library for medical image processing and segmentation.

","execution_count":null},{"metadata":{},"cell_type":"markdown","source":"
\nCredit goes to Marco Vasquez for his HTML notebook.\n
","execution_count":null},{"metadata":{},"cell_type":"markdown","source":"
\n

Table of Contents

\n 1. The Affine Transform1\n 2. Elastic Deformation 2\n 3. Flipped images3\n 4. Spatial Padding4 \n 5. Experimental Method from APTOS5 \n Appendix A: Libraries and Resources.A ","execution_count":null},{"metadata":{"_uuid":"8f2839f25d086af736a60e9eeb907d3b93b6e0e5","_cell_guid":"b1076dfc-b9ad-4769-8c92-a6c4dae69d19","trusted":true,"_kg_hide-input":true,"_kg_hide-output":true},"cell_type":"code","source":"%reload_ext autoreload\n%autoreload 2\n%matplotlib inline\nimport pandas as pd\nimport numpy as np\nimport matplotlib.pyplot as plt\nimport openslide\nimport os\n!pip install monai\nimport torch","execution_count":null,"outputs":[]},{"metadata":{"_kg_hide-input":true,"trusted":true},"cell_type":"code","source":"train = pd.read_csv('../input/prostate-cancer-grade-assessment/train.csv')\ngpu = torch.device(\"cuda:0\" if torch.cuda.is_available() else \"cpu\")\ngpu","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"## Our Image Reading Function","execution_count":null},{"metadata":{},"cell_type":"markdown","source":"This is our simple function to show/read images (from Gabriel's kernel). It is going to be very useful for our image processing and we will use many variations of these functions later on for our different transforms.","execution_count":null},{"metadata":{"_uuid":"d629ff2d2480ee46fbb7e2d37f6b5fab8052498a","_cell_guid":"79c7e3d0-c299-4dcb-8224-4455121ee9b0","trusted":true,"_kg_hide-input":true},"cell_type":"code","source":"def show_images(df, read_region=(1780,1950)):\n data = df\n f, ax = plt.subplots(3,3, figsize=(16,18))\n for i,data_row in enumerate(data.iterrows()):\n image = str(data_row[1][0])+'.tiff'\n image_path = os.path.join('../input/prostate-cancer-grade-assessment',\"train_images\",image)\n image = openslide.OpenSlide(image_path)\n spacing = 1 / (float(image.properties['tiff.XResolution']) / 10000)\n patch = image.read_region(read_region, 0, (256, 256))\n ax[i//3, i%3].imshow(patch) \n image.close() \n ax[i//3, i%3].axis('off')\n ax[i//3, i%3].set_title(f'ID: {data_row[1][0]}\\nSource: {data_row[1][1]} ISUP: {data_row[1][2]} Gleason: {data_row[1][3]}')\n\n plt.show()\nimages = [\n '059cbf902c5e42972587c8d17d49efed', '06a0cbd8fd6320ef1aa6f19342af2e68', '06eda4a6faca84e84a781fee2d5f47e1',\n '037504061b9fba71ef6e24c48c6df44d', '035b1edd3d1aeeffc77ce5d248a01a53', '046b35ae95374bfb48cdca8d7c83233f',\n '074c3e01525681a275a42282cd21cbde', '05abe25c883d508ecc15b6e857e59f32', '05f4e9415af9fdabc19109c980daf5ad'] \ndata_sample = train.loc[train.image_id.isin(images)]\nshow_images(data_sample)","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"

1. The Affine Transform

\n\nThe affine transform is any function than can preserve collinearity. There are multiple examples of affine transforms which you would have heard in high school mathematics:\n+ Shear\n+ Translate\n+ Reflect\n+ Rotate\n+ Geometric contraction \n\net. al.\n\nWe can apply a simple affine transform to our images over here (not necessarily one of the geometric transforms given above).","execution_count":null},{"metadata":{"trusted":true,"_kg_hide-input":true},"cell_type":"code","source":"from monai.transforms import *\n\ndef show_images_affine(df, read_region=(1780,1950)):\n data = df\n f, ax = plt.subplots(3,3, figsize=(16,18))\n for i,data_row in enumerate(data.iterrows()):\n image = str(data_row[1][0])+'.tiff'\n image_path = os.path.join('../input/prostate-cancer-grade-assessment',\"train_images\",image)\n image = openslide.OpenSlide(image_path)\n spacing = 1 / (float(image.properties['tiff.XResolution']) / 10000)\n patch = image.read_region(read_region, 0, (256, 256))\n patch = np.array(patch)\n # MONAI transforms always take channel-first data: [channel x H x W]\n im_data = np.moveaxis(patch, -1, 0) # make them channel first\n # create an Affine transform\n affine = Affine(rotate_params=np.pi/4, scale_params=(1.2, 1.2), translate_params=(200, 40), \n padding_mode='zeros', device=torch.device('cuda:0'))\n # convert both image and segmentation using different interpolation mode\n new_img = affine(im_data, (256, 256), mode='bilinear')\n \n ax[i//3, i%3].imshow(np.moveaxis(new_img.astype(int), 0, -1)) \n image.close() \n ax[i//3, i%3].axis('off')\n ax[i//3, i%3].set_title(f'Affine Transformed\\n Gleason: {data_row[1][3]}')\n plt.show()\n \nimages = [\n '059cbf902c5e42972587c8d17d49efed', '06a0cbd8fd6320ef1aa6f19342af2e68', '06eda4a6faca84e84a781fee2d5f47e1',\n '037504061b9fba71ef6e24c48c6df44d', '035b1edd3d1aeeffc77ce5d248a01a53', '046b35ae95374bfb48cdca8d7c83233f',\n '074c3e01525681a275a42282cd21cbde', '05abe25c883d508ecc15b6e857e59f32', '05f4e9415af9fdabc19109c980daf5ad'] \ndata_sample = train.loc[train.image_id.isin(images)]\nshow_images_affine(data_sample)","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"**YOU ARE HERE:**\n
\n
20%
\n
","execution_count":null},{"metadata":{},"cell_type":"markdown","source":"

2. Elastic Deformation

\n\nElastic deformation basically applies an invisible \"stress\" to our image in such a manner that the image gets deformed due to the virtual stress applied. We can simulate this sort of \"stress\" with MONAI's `Rand2DElastic` transform. This transform actually is a concept in real life: apply stress to an object in such a way that the original position is recoverable even after having applied stress (elasticity basically).\n\nThis is an example of strain in real life:\n![](https://upload.wikimedia.org/wikipedia/commons/thumb/3/3c/DeformationOfRod_plain.svg/400px-DeformationOfRod_plain.svg.png)\n\nWe can similarly apply such virtual stress to our image to deform it.\n\nWe can use 2-dimensional elastic deformation where the strain is the measure of deformation. The strain can be simply represented by:\n![](https://wikimedia.org/api/rest_v1/media/math/render/svg/77b3bbb9d1485775abe0dc7069df39c807ba6988)\n\nFor more on the mathematical side, you can see the Wikipedia page [here](https://en.wikipedia.org/wiki/Deformation_(mechanics)).","execution_count":null},{"metadata":{"_kg_hide-input":true,"trusted":true},"cell_type":"code","source":"def show_images_elastic2d(df, read_region=(1780,1950)):\n data = df\n f, ax = plt.subplots(3,3, figsize=(16,18))\n for i,data_row in enumerate(data.iterrows()):\n image = str(data_row[1][0])+'.tiff'\n image_path = os.path.join('../input/prostate-cancer-grade-assessment',\"train_images\",image)\n image = openslide.OpenSlide(image_path)\n spacing = 1 / (float(image.properties['tiff.XResolution']) / 10000)\n patch = image.read_region(read_region, 0, (256, 256))\n patch = np.array(patch)\n im_data = np.moveaxis(patch, -1, 0) # make them channel first\n # create an elastic transform\n deform = Rand2DElastic(prob=1.0, spacing=(30, 30), magnitude_range=(5, 6),\n rotate_range=(np.pi/4,), scale_range=(0.2, 0.2), translate_range=(100, 100), \n padding_mode='zeros', device=torch.device('cuda:0'))\n new_img = deform(im_data, (256, 256), mode='nearest')\n \n ax[i//3, i%3].imshow(np.moveaxis(new_img.astype(int), 0, -1)) \n image.close() \n ax[i//3, i%3].axis('off')\n ax[i//3, i%3].set_title(f'Deformed Transformed\\n Gleason: {data_row[1][3]}')\n plt.show()\n \nimages = [\n '059cbf902c5e42972587c8d17d49efed', '06a0cbd8fd6320ef1aa6f19342af2e68', '06eda4a6faca84e84a781fee2d5f47e1',\n '037504061b9fba71ef6e24c48c6df44d', '035b1edd3d1aeeffc77ce5d248a01a53', '046b35ae95374bfb48cdca8d7c83233f',\n '074c3e01525681a275a42282cd21cbde', '05abe25c883d508ecc15b6e857e59f32', '05f4e9415af9fdabc19109c980daf5ad'] \ndata_sample = train.loc[train.image_id.isin(images)]\nshow_images_elastic2d(data_sample)","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"**YOU ARE HERE:**\n
\n
40%
\n
","execution_count":null},{"metadata":{},"cell_type":"markdown","source":"

3. Flipped images

","execution_count":null},{"metadata":{},"cell_type":"markdown","source":"Flipping as a transform is rather self explanatory. We can either flip our set of images across the x-axis or we can flip our set of images across the y-axis.\n\n![An x-axis flip](data:image/jpeg;base64,/9j/4AAQSkZJRgABAQAAAQABAAD/2wCEAAkGBxAOERUTERETDQ4NDg8ODhgWEg8VDhYQFhEaGRUXFxcYHSgsJBolHRUVITEhJSktLi4uFx8zODMsNygtLisBCgoKDg0OGxAQGy0iHyUtLS0tLS0tLS0rLS0tLS0rLS0tLS0rLS0tLS0tLS0tLS0tLS0tLS0tLS0tLS0rLS0tLf/AABEIAMwAzAMBEQACEQEDEQH/xAAbAAEAAQUBAAAAAAAAAAAAAAAABgECAwUHBP/EAEcQAAEDAgEGCgYIBAUFAQAAAAEAAgMEEQUGEiExUZITFiJBUlNxgZHSFCNhocHRJDJCVGKCk7EHc3SjFTNDcvElY7LC4TT/xAAaAQEAAgMBAAAAAAAAAAAAAAAAAQQCAwUG/8QAMhEBAAIBAgQDBwIHAQEAAAAAAAECAwQREiExURRBYQUTMlJxodEiwRUzQoGRsfDhI//aAAwDAQACEQMRAD8A7igICDlGUuKztrJS2V7OCkzGWeQ0Ac1lI2mEZdyMs2dnDN6TdDx2jUfcmwmeGYvBUi8Ugcedup47WnSoGwQEBAQEBAQEBAQEBAQEBAQEBAQEBAQEHH8ox9Nm/n/JSJXieS8Ety0cC/a0cnvb/wAKBFq/AqmmOcAXhukPjztHt2hTuNjhGWs8VhLapZ7TaUdjufvTYTXCcoqarsGPzHn7DuS/u29ygbhAQEBAQEBAQEBAQEBAQEBAQEBAQEHIcox9Nm/n/JSOikKBRBq8SwCCouS3Mf0maHd41FBFsRyXni0s+kM/COXu/JTuK4XlXVU3JJ4Zg+zJfOHsDtYQTXCcraaosHO9Hkd9l50dztR9yjYSFAQEBAQEBAQEBAQEBAQEBAQEBByPKQfTZv53yUjoxCgUsgWQLIPDiGEQ1A9YwE9IaHjvQRbEskZY7mF3DDonQ/5H3Kdx4sOxyronZgc4ButkgNvA6R3IJjhGWsEthMPR37Sbxb3N3psJMyQOALSHA6iDcHsIUDIgICAgICAgICAgICAgICAg5JlL/wDtm/nfJSOjki9ri+u1+VbbZNp23N/IsoCyBZAsgqGIMVXh8U7c2WNrx7Ryh2HWEEWxPIk6XU7/AMj/AIO+ancaOCqq8Pfblwnokeqd3aj2hSJVhGXMb7NqG8Eek25Z3t1j3qNhLKWqjmaHRubIw87TcKBnQEBAQEBAQEBAQEBAQEHJspB9Nm/nfALKsTM7QiZ2jeWxqaszzF4Jbp0WNnADUu/jxRjx8Dk3yTe/E2lNi0sY5dpmN1nU8D9iqWTSY7/Dyn7LFNRavXnH3b+CTPaHAENcL6RY6Vzb14bTWfJerPFG7IGrFKoaguAQVAQXAIMVVDHI3Nka14dzEXagieJ5HNfd0DuCPQdpZ3HWPep3EddFV0D78uF3Sb9R3fqPYgkWE5dHQ2oZf8bNB72/LwTYTDD8ShqG50UjX7QDyh2t1hQPYgICAgICAgICAgICDluOt+mzHZJ8Ar/s/FxZOLsq6u/DTburSrr3c6rZ0MHDyNj+z9eT/YObvOhVM1/d0m3n0j6t+OvHaK/5S8BcR1FUFUFQEFUFpfsQWIKWQUfGCLOAcHawRdp7kEexPJGGS7oiad+waYj+Xm7lO4i9XhNVRuz7Obm6nxklveRq70G3wnLeWOzZ2iZnSGiQfA+5NhMcMxqCqHq5Bn9E6HjuKgbNAQEBAQEBAQEBBzDKDRVTfik+AXd9nU2xb95cvV23ybdmKAq1aGiEqyUg5D5T/qvs3/Y3R+9/Bcj2hf8AVFI8v9y6Gkr+mbd29suetqoFkC6C0oFkFLIFkCyBZAIQaTE8mIJ7kDgX9Jn1T2t1H3JuIpiOTlTTnOaDKxunOZnZw9pbrCkenCsr6mCzZPXsGiztDx2O+d02EywrKWnqbAP4OQ/Yfod3HUVA3SAgICAgICAg5TjU2fUSn/vPHgbfBeo09eHFWPRw81t8kz6sbH2BPsWUxzREuh4RBwcEbejG33i/xXm9Rfiy2n1dnDXhpEej1rS2KoKIFkCyClkCyBZAsgWQLIFkFLINZieAwVGl7LP6beS/v5j3oInieSU8VzF9IZsGiXd5+5TuMOG5S1VIc0uL2t0Fkgdo7DrCCY4XldTT2DzwD9jjyO5+rxsmwkAN9I0qBcgICAgIONTvu9x6T3H3levrH6YeemecsmdyT2LHbmy8kujyxhDQOCk5IaObmC8pbnMu9HRXjpD1UnixRslTjrD1UniE2DjrD1UnixNhTjtD1UnixA47Q9VJ4sUCnHeHqZPFiCnHiHqZPFinYOPEPUyeLE2FOPMPUyeLE2FOPUPUyeLE2Dj1D1MnixNhTj3D1Mm8xRsHHuHqZN5inYU4+Q9RJvMTYOPkPUSbzE2HjxDKajqRaWmefxZzA8djhpQRuodHneqz8zY/NzvEa1I92F47PTf5chzei7lM8ObuUCc5OZTNrHcG5vBzBt9Bux1tdkEjUAgICDis+h7gdbXuB7iV7GvOsT6PO25TLLfkn/asfNPkm1HkbTSMY/Pl5bGv1s5wD0V5PLyvMesu/Sd6xLNxHpunLvM8qx3ZHEam6ybeZ5U3DiNTdZNvM8qbhxGpusm3meVNw4i0vTm3meVQHESl6cu8zyoKcRKXpzbzPKgcQ6XpzbzPKp3DiHS9ObeZ5U3DiFSdObfZ5U3FOIVJ05t5nkTcU4hUnTm32eRNw4hUnTm32eRNw4g0nTm32eRRuKcQKTpzb7PIp3FRkDSdObfZ5E3FRkFSdObeZ5E3F3ESk6c2+zypuNhg+TsFG4ujznPItnPdcgbBYAKBuUBBY+QDWQO02U7TPREzELPSY+m3eanDbscUd3IseZmVMwGrhnkfmN/ivWaWeLDWfRwM8bZLfVjhKysiHRskcQa+kjznND47wuuWg8k2HusvO6/DNc9to5Tz/wAuvpckTijfy5N6yQHSCD2G6pzEx1WYndZPM2Npe9wYxgu4k2aANqVrNpitY3mS1orG89Gnw/KeCV+Y4OgL/wDJMgzRIOYg8x9h0q3l0GSleKOffbyV6aqlp2nl23826ZK131XA9hBVSazHWFiJiWVQlhNQwaC9viFPDbsjigbOwmwc0nYCCk1tHWDeGZQlifM1psXNB9pAUxWZ6QjeFPSWdNu8FPDbsjijuzLFkxukDdbg3tICbTPREzELfSY+m3eap4bdjijuyA30jSoSo94AuSANpNkiJnob7LPSY+m3eap4bdkcUd17Hgi4II9huomJjqnfdUm2tBj9Jj6bd5qnht2RxR3PSY+m3eanDbscUd2ZQlEP4g4RPWRRNgj4VzJi9wzo22GaRflEbV0/Zeox4L2nJO28evf0UNfhvlpEUjfn/wB1QfiRiH3X+5Teddv+Kab5/tb8OX4DP8n3j8suP4XJStgD25j30zWvFwbPYSHC4JGot8Vhpc9c034Z3ji+0s82KccVieu3+nmp3LdaGFZWOwmWplzYIxI8sz7Z0bdA16XED/lPEUxU3vO0f3/ZjbBbJbasbz/b909yJpH4dSSelNFN690mlzHcjgmC/IJ52kW1rhe0MldVnr7meLlt5957uro6TgxT7zlz3+0dnmqq11c8OeCymYbxRnW4jU9/wbzLdjxV09do52857ekflhfJOWd5+Hyj95X1sUb43CUAx2JN+YAa/Yox2tW0cPUvFZj9XRg/h3hUsUkspiMcFRG3gSSzOcM4kXaDcaNoCn2rqKXrWnFvaJ5o0GG1bTbbasxyT5cV1HJMcyRrpqiV7KfPZJM97DwkAuCdBsX3Xp9N7R09MVazfnER5T+HBz6LNbJa0V5b94/L1ZI5LVlPWRSywcHGwy57s+A2vC9o0NcTrIWrX6/BfT2pW28zt5T3j0Z6XSZaZq2tXaI+nZ1FeddtzfLjJurq6syQw8LHwLGXz4W6Re4s5wPOF3vZutw4cHDe207z5T+0OPrNNlyZeKtd42jt+7S0mRuIMkY402aGyMcTwlNoaHAnU9XcntLTTSYi/lPlP4VqaHPFonh8484/LsS8o9Chf8QsGqKwRcBFw3Bl+fyo22va313DYur7L1OLBxe8nbfbv+zna/BfLFeCN/8AH7oZxIxD7r/cpvOuz/FNN8/2t+HN8Bn+T7x+XVcCgdFTQRvGa+KnhY8XBs5rACLj2heW1F4vmvavSZmfu72Cs1x1rPWIhrsuMPlqqR0cTOEkL4zm5zG6A650uICsezs1MWeLXnaNp/7k063FbJimtY3nk53xIxD7r/cpvOvQ/wAU03z/AGt+HH8Bn+T7x+XQ8hsOlpaQRys4KThJHWzmO0E6NLSQvP8AtHNTNn4qTvG0f9zdjRYrY8XDaNp5tri8JkglY0XfJC9rRcC5LSALlVcNorkraekTDflibUmI7OTjIfEPuv8AcpvOvVfxTTfP9rfh5/wGf5PvH5OJGIfdf7lN50/imm+f7W/CfAZ/k+8flMsvcoeBb6PEfWyj1hH2WHm7T+y4/szR8c+9v0jp6z/46et1HDHBXrPVDcOykq6a3Byuczov5bPA6R3ELsZdFgy/FXn3jk52PU5adJSzDf4hxnRPGWfiZym7p0/uuXm9jWjnjtv9V7H7Rj+uP8M2VUtPiNIXwSslfT+uaAbPzbcsFp0jRp7lr0MZdLniuSNoty9PRnqZpnxb0neY5oFRv0Lv5IcqstpRVLoZI5WXJhkaSB9Z0Z0Pb3i6qZaRkpak+cffyb6WmtotHl/0tviGIyVr8592xMPqo+Yfjftd7gqmHBXBXavWes/tHp/tvyZZy23np5Q9MDljaGULaiI1UsdK36r/AFlSdkDTq7XEWUVtGHHOafLlH1/8Jj3l4xx9Z+n/AKnkbA0AAWAFgNgGpcOZmZ3l1YjZkQEBAQEBAQEBAQEBAQEBAQc9yyySe5zqinvKSS+ZhJL+1l9fZ4bF3fZ/tGsRGLJy7T+fy5Wr0czM3pz7x+EALl3nLUzlOyAPtqNk23N2Wjks63SWF68mVZ5t1A5VLQsQ2EDlotDbD2GoEbC9x0MFz3LTwTaeGPNnxRWN5azBsdmpzJI0Mz6hzXEua4kMH1WjSNC16ylLzFI6V/35yz097Vibectrxzq9kW47zKn4WnqseIurx0qtkW47zKPC09TxFzjpVbItx3mTwtPU8RdXjpVbItx3mTwtPU8Rc451WyLcd5k8LT1PEXOOdVsi3HeZPC09TxFzjnVbItx3mTwtPU8RdXjlVbItx3mTwtPU8Rc45VWyLcd5k8LT1PEXOOVVsi3HeZPC09TxF1eOVVsi3HeZPC09TxFzjlVbItx3mTwtPU8Rc441WyLcf5k8LT1Pf3V441WyLcd5k8LT1PEXOONVsi3H+ZPC09T3924yZyglqZCyRrfqZ4LQRax57k7Vpz4K0jeG3Flm07SlCqrAghuVuRrKq8sAEdTrcNTJO3Y72+K6uh9pWw/oyc6/eFDVaKMn6qcrf7cvqYXRPLJGlj2GxaRYhenpat44qzvDiWrNZ2nlLDnLLZAHqdkbt7RTZ7QVRyV2nZZpO8NnA5V7Q3QzQU/pdQyD/RZmzVPYDyWd59y1Xv7jFOTznlH5/szrX3t4p5Rzn8Jd/hVN1Ee4FxfeX7ulwV7H+FU3UR7gT3lu5wV7LhhVN1Ee41R7y/dPBXsuGE03URbjU95fucFey4YTTdRF+m35KPeX7p4K9l4wmm6iL9NnyUe8v3k4K9lwwmm6iLcZ8k95fvKeCvZcMIpuoh/TZ8lHvL95OCvZcMIpfu8P6bPko97fvKfd17Lhg9L93h/TZ8lHvb95Pd17QuGD0v3eH9NnyUe9v3lPu69oV/wel+7w/ps+Se9yd5/ye7p2hjmwyjjGc+GBgGsmOID9lMZMk8omUTSkdYhGsSxqgZdsNLFM7aYmBndouVapiyz8Vpj+7RbJjjpCPWlqX8iIX6MUYa0dzR7yrHKkc5/y087TyhI8MyMcbOnfmDotNz3nV+6rX1Uf0t1NPP8AUldBh8VO3NiYGbTzntJ0lVL3taecrNaRXo9ixZCAgj+U2TMOIM5Q4Odo9W8DlD2Ha32K5o9dk01uXOvnCtqNLTNHPr3ccxGikppHxyWz4jY2Nwbc4OxevwZqZqRenR5/Jjtjtw26vLnLds1vbhVVmOzTqfq7VpzU3jeGzHbadm/dUCNhcfsjx2BUYpNrbQtcW0bpbknhxggzpP8APqDw0nsv9VvcP3K4+uzxkybV+GvKPy6GlxTSm89Z5y3N1SWVQUFwKgXAoleCoSvBUC4FQleFAuChK8IlhrK2KBudK9sY9p/Yc6VrNp2hE2iOqLYnluNLadmd+J4s3tDdfirVNL8yvbUfKjzRV4g/7c58Im/AKz/88Udmn9eSe6SYVkUBZ1Q/PPRYSG97tZ9yrX1U/wBLfXT/ADJVS0kcTc2NjY27ALeKqWtNp3lYisR0ehQkQEBAQEHHcqohJWTNdzzeGrUuzpctsVYtVzM9IvaYlqMosn5aJ+nlwuPIeP2dsK72k1tNRHLlbzj8OZqNNbFPp3aa6uK6UZNxmskZnD1dOWySX+q54PIHuuVy9beMFJ2625R6d13TVnLaN+kdXRm1g5+T/wCK83wOzxMoeDqN1jsndcCguBUJXgqBeCiV4KhK8FYi4Il5sQxWGmF5ZAz2a3nsaNKmuO1ukIteK9UTxTLeR9xAzgh0nWL+4ah71bppYj4le2on+lqaPDKuvdnAOlv9uQuDPE/BbLZKY42a4pe6W4XkXDHYznh37NUfhzqrfVWn4eSxTTxHXmk0UTWANa0MaNQAAA7AFV3merfEbdGVEiAgICAgICDj2UZ+nTfz/kuri/lx9HOyfHKe1cDZGlj2h7HaCCLtKrUvasxas7SsWrFo2lzjKXJN1OTJCDLDzt1vZ8x7V6PRe0a5f05OVvtLj6nRzT9VOcJXk5hfolO1hHLfy5D+MjV3CwXI1mo9/lm3lHKPov6fD7um3n5tg4Ks3rNWpShljqHDnzu1YzWGUTL0sq9o8FhNU8TOydu1YzWWW8MzXjaFjsndkaVDJ4MSx6Cm+u+7+g3lP/8Anes6YrW6MLZK16olimWM0lxF9HZtGmXx5u5WaaesdebRbNM9OTBhuTdXWHOILA7SXyk6faBrKyvnpTl/pFcV7c0ywrI+mgsXj0h+145F/YzV43VK+pvbpyWaYKx15pGBbQNC0Ny5AQEBAQEBAQEBBxvKI/T5f6n4hdbF/Lj6Odk+OXRZAqcLTC4KWLG4KUMZClCwtWQAILwFiMgUJazEMooINAPCv6Lfi7UFnXFaWM5IhG6/KKec5rSYg7Q1rL5x9lxpK31xVrzapyWl7cJyNqqizpPo8Z53aZD2N+dlrvqaV6c2dMFrdeSbYTkvS01iGcJIPtP0u7QNQVK+e91mmKtW9WptEBAQEBAQEBAQEBAQcXyhP0+X+p+IXWxfy4+jnZPjl0t4VJaYXBZIY3BShYQiFtlIo4gC5IaG6yTZqIaTEMp4Y7iP1z/ZoYO/nW2uKZ6sJyRHRHqnFKmsdmAuOdqZGDp7QNa2xStI3a5ta3JvcHyDmks6dwgZ0RypT8B7+xaMmqrHw822mnmeqb4VgNPSD1UYz+kdMh/MVTvltfrK1XHWvRtVrZiAgICAgICAgICAgICAg4plAf8AqEv9V8Quti/lx9HOyfHP1dPeFRW2NwWSGMhShhqJWxtznuDA3nJspiJnoxnkjWJZXRsuIG8KekdDO4az7lvrhnza5yx5I/wtXiD81ufMei0chvaNQ7Stv6Mcbzya/wBV5SrBv4euNnVUlh0GaT+Z/N3eKq5NXHSrfTTfMm+G4VBTNzYo2x7SByj2nWVTve1+srNaRXo9yxZCAgICAgICAgICAgICAgICDiOPn/qEv9V/7BdfH/Kj6Odf45+rqrwuetvFX10VO3Ole2Ie08o9g1lZ1rNujGZiOqIYrltrbAz87/g35+Cs1wfM0WzdmkpaSsxJ92h9Qb2LibRN9l9Q7FttamOOfJhFbXnkmeC/w7jZZ1S/hXdBlwwdrtZ9yp5NXM8qrFNNEfEmlHSRwNDImNiYNQaAB/yqlrTad5WYiI5Q9KhIgICAgICAgICAgICAgICAgICDhmUUlq+Zx1NqrnsDgV18X8uPo5uT45+rcYxl5LJcQM4FnSdZ0vhqHvWumniPiZ2zTPRq8MwKuxF2e1rntdrkkJDO4nSe5bL5aY+X2Y1x3vzTrA/4f08NnTn0mTZbNhH5efv8FSyaq1vh5LNNPWOvNMIYmsaGtaGNGgAABo7AFVmd+qxEbMqAgICAgICAgICAgICAgICAgICAgIObY5kDPPVPkjkjEMz+Eu4nObfWM22lXseqrWkRPVUvp7TbeG+wTIakprOe30mUfakAzb/hZqHvWjJqb26cobaYK19UqAWhuVQEBAQEBAQEBAQEBAQEBAQEBAQEBAQEBAQEBAQEBAQEBAQEBAQEBAQEBAQEBAQf/9k=)\n\nThe above image flips across the x-axis.","execution_count":null},{"metadata":{"_kg_hide-input":true,"trusted":true},"cell_type":"code","source":"def show_images_rotate(df, read_region=(1780,1950)):\n data = df\n f, ax = plt.subplots(3,3, figsize=(16,18))\n for i,data_row in enumerate(data.iterrows()):\n image = str(data_row[1][0])+'.tiff'\n image_path = os.path.join('../input/prostate-cancer-grade-assessment',\"train_images\",image)\n image = openslide.OpenSlide(image_path)\n spacing = 1 / (float(image.properties['tiff.XResolution']) / 10000)\n patch = image.read_region(read_region, 0, (256, 256))\n patch = np.array(patch)\n im_data = np.moveaxis(patch, -1, 0) # make them channel first\n rotater = Flip(spatial_axis=1)\n new_img = rotater(im_data)\n \n ax[i//3, i%3].imshow(np.moveaxis(new_img.astype(int), 0, -1)) \n image.close() \n ax[i//3, i%3].axis('off')\n ax[i//3, i%3].set_title(f'Rotated Transformed\\n Gleason: {data_row[1][3]}')\n plt.show()\n \nimages = [\n '059cbf902c5e42972587c8d17d49efed', '06a0cbd8fd6320ef1aa6f19342af2e68', '06eda4a6faca84e84a781fee2d5f47e1',\n '037504061b9fba71ef6e24c48c6df44d', '035b1edd3d1aeeffc77ce5d248a01a53', '046b35ae95374bfb48cdca8d7c83233f',\n '074c3e01525681a275a42282cd21cbde', '05abe25c883d508ecc15b6e857e59f32', '05f4e9415af9fdabc19109c980daf5ad'] \ndata_sample = train.loc[train.image_id.isin(images)]\nshow_images_rotate(data_sample)","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"**YOU ARE HERE:**\n
\n
60%
\n
","execution_count":null},{"metadata":{},"cell_type":"markdown","source":"

4. Spatial Padding

\n\nSpatial padding is used in Convolutional layers. Why? Well, convolutional layers often tend to lose pixels on the side of the image sometimes. With a small amount of Convolutional layers, it's fine but with many convolutional layers the pixel loss tends to add up over and over again and it can eventually spiral out of control.\n\nOne simple solution to this is to add some extra pixels around the borders so that even if we lose information in pixels it will not affect our model much. To keep it relatively useless besides being a dummy (in a metaphorical sense) we keep most of the values 0. The following image gives a good demo of padding:\n![](https://d2l.ai/_images/conv-pad.svg)\n\nIn the above image we have a padded input, which is then processed into a kernel and then comes out as a 4 x 4.\n\nConvolutional layers often pad with odd numbers such as 1, 3, 5 or 7. Thus, we can preserve the spatial dimensionality by padding with odd numbers. But don't think padding is only used in Convolutional layers, we can also use padding as a transform with MONAI.","execution_count":null},{"metadata":{"_kg_hide-input":true,"trusted":true},"cell_type":"code","source":"def show_images_rotate(df, read_region=(1780,1950)):\n data = df\n f, ax = plt.subplots(3,3, figsize=(16,18))\n for i,data_row in enumerate(data.iterrows()):\n image = str(data_row[1][0])+'.tiff'\n image_path = os.path.join('../input/prostate-cancer-grade-assessment',\"train_images\",image)\n image = openslide.OpenSlide(image_path)\n spacing = 1 / (float(image.properties['tiff.XResolution']) / 10000)\n patch = image.read_region(read_region, 0, (256, 256))\n patch = np.array(patch)\n im_data = np.moveaxis(patch, -1, 0) # make them channel first\n rotater = SpatialPad(spatial_size=(300, 300), mode='mean')\n new_img = rotater(im_data)\n \n ax[i//3, i%3].imshow(np.moveaxis(new_img.astype(int), 0, -1)) \n image.close() \n ax[i//3, i%3].axis('off')\n ax[i//3, i%3].set_title(f'Rotated Transformed\\n Gleason: {data_row[1][3]}')\n plt.show()\n \nimages = [\n '059cbf902c5e42972587c8d17d49efed', '06a0cbd8fd6320ef1aa6f19342af2e68', '06eda4a6faca84e84a781fee2d5f47e1',\n '037504061b9fba71ef6e24c48c6df44d', '035b1edd3d1aeeffc77ce5d248a01a53', '046b35ae95374bfb48cdca8d7c83233f',\n '074c3e01525681a275a42282cd21cbde', '05abe25c883d508ecc15b6e857e59f32', '05f4e9415af9fdabc19109c980daf5ad'] \ndata_sample = train.loc[train.image_id.isin(images)]\nshow_images_rotate(data_sample)","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"We can see dummy pixels being placed at the end (borders) of our images. Padding in action, baby!","execution_count":null},{"metadata":{},"cell_type":"markdown","source":"**YOU ARE HERE:**\n
\n
80%
\n
","execution_count":null},{"metadata":{},"cell_type":"markdown","source":"

5. Experimental Method from APTOS

\n
\n\nI have read Neuron Engineer's notebook on APTOS and I have also read about a very useful preprocessing method used in that competition to great effect. I am not exactly sure of how well it will work but I am willing to try it out.","execution_count":null},{"metadata":{"_kg_hide-input":true,"trusted":true},"cell_type":"code","source":"import cv2\n \nimage = cv2.imread('../input/panda-resize-and-save-train-data/0005f7aaab2800f6170c399693a96917.png')\nimage = cv2.cvtColor(image, cv2.COLOR_BGR2HSV)\nimage = cv2.resize(image, (256, 256))\nimage=cv2.addWeighted ( image,4, cv2.GaussianBlur( image , (0,0) , 256/10) ,-4 ,128) # the trick is to add this line\nplt.imshow(image)\nplt.title('Ben Graham Method\\n Transformed image')\nplt.show()\n","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"**YOU ARE HERE:**\n
\n
100%
\n
","execution_count":null},{"metadata":{},"cell_type":"markdown","source":"

APPENDIX A: A list of useful resources and credits

\n\n## Credits to the libraries used:\n\n+ **Pandas: https://pandas.pydata.org/ **\n+ **Numpy: https://numpy.org/**\n+ **Openslide: https://openslide.org/api/python/**\n* **Matplotlib: https://matplotlib.org/**\n+ **Monai: https://monai.readthedocs.io/en/latest/index.html**\n\n## Useful resources\n\n+ **https://www.urologyhealth.org/urologic-conditions/prostate-cancer**\n+ **https://www.healthline.com/health/prostate-cancer**","execution_count":null}],"metadata":{"kernelspec":{"language":"python","display_name":"Python 3","name":"python3"},"language_info":{"pygments_lexer":"ipython3","nbconvert_exporter":"python","version":"3.6.4","file_extension":".py","codemirror_mode":{"name":"ipython","version":3},"name":"python","mimetype":"text/x-python"}},"nbformat":4,"nbformat_minor":4}