'.join(\n '| {} | '.format(''.join(str(_) for _ in row)) for row in data)\n )\n))\n","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"fig, ax = plt.subplots(nrows=1,figsize=(12,6)) \ntmp = train.groupby('data_provider')['gleason_score'].value_counts() \ndf = pd.DataFrame(data={'Exams': tmp.values}, index=tmp.index).reset_index() \nsns.barplot(ax=ax,x = 'data_provider', y='Exams',hue='gleason_score',data=df, palette='Set1') \nplt.title(\"Number of examinations grouped on Data provider and Gleason score\") \nplt.show()","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"markdown","source":"**One Mislabeled Image?**"},{"metadata":{},"cell_type":"markdown","source":"In the above dataframe it looks like one image might have been converted to wrong ISUP grade."},{"metadata":{"trusted":true},"cell_type":"code","source":"df_train[(df_train.isup_grade==2)&(df_train.gleason_score !='3+4')]","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"All the Karolinska images in the training data is graded by the same pathologist.However,for the test set we used several pathologists who each labeled the images using ISUP(not Gleason) and derived a consensus label.The mislabeled image was one of the those images but was later moved to the training set. "},{"metadata":{},"cell_type":"markdown","source":"# Differences Between Data Providers"},{"metadata":{},"cell_type":"markdown","source":"They used different scanners with slightly different maximu microscope resolutions and worked with different pathologists for labeling their images."},{"metadata":{"trusted":true},"cell_type":"code","source":"data_providers=df_train.data_provider.unique()\nfig=plt.figure(figsize=(6,4))\nax=sns.countplot(x=\"isup_grade\",hue=\"data_provider\",data=df_train)\nplt.title(\"ISUP Grade Count by Data Provider\",fontsize=14)\nplt.xlabel(\"ISUP Grade\",fontsize=14)\nplt.ylabel(\"Count\",Fontsize=14)\nplt.show()","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"data_providers=df_train.gleason_score.unique()\nfig=plt.figure(figsize=(6,4))\nax=sns.countplot(x=\"isup_grade\",hue=\"gleason_score\",data=df_train)\nplt.title(\"ISUP Grade Count by gleason_score\",fontsize=14)\nplt.xlabel(\"ISUP Grade\",fontsize=14)\nplt.ylabel(\"Count\",Fontsize=14)\nplt.show()","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"def plot_relative_distribution(df, feature, hue, title='', size=2):\n f, ax = plt.subplots(1,1, figsize=(4*size,3*size))\n total = float(len(df))\n sns.countplot(x=feature, hue=hue, data=df, palette='Set2')\n plt.title(title)\n for p in ax.patches:\n height = p.get_height()\n ax.text(p.get_x()+p.get_width()/2.,\n height + 3,\n '{:1.2f}%'.format(100*height/total),\n ha=\"center\") \n plt.show()","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"plot_relative_distribution(df=train, feature='isup_grade', hue='data_provider', title = 'relative count plot of isup_grade with data_provider', size=2)\n","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"plot_relative_distribution(df=train, feature='gleason_score', hue='data_provider', title = 'relative count plot of gleason_score with data_provider', size=3)","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"plot_relative_distribution(df=train, feature='isup_grade', hue='gleason_score', title = 'relative count plot of isup_grade with gleason_score', size=3)","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"# Visualizing Image and Mask Samples"},{"metadata":{},"cell_type":"markdown","source":"Let's have a quick first look at the differences between the data providers in regards to the original images and the masks.\n\n**Radboud:**Prostate glands are individually labelled.Valid values are:\n* 0:background (non tissue) or unknown\n* 1:stroma(connective tissue,non-epithelium tissue)\n* 2.healthy (benign) epithelium\n* 3.cancerous epithelium (Gleason 3)\n* 4.Cancerous epithelium (Gleason 4)\n* 5.Cancerous epithelium (Gleason 5)\n\n**Karolinska**:Regions are labelled.Valid values are:\n* [0]:backgound (non tissue) or unknown\n* [1]:benign tissue(stroma and epithelium combined)\n* [2]:cancerous tissue (stroma and epithelium combined)"},{"metadata":{"trusted":true},"cell_type":"code","source":"def display_images(slides): \n f, ax = plt.subplots(5,3, figsize=(18,22))\n for i, slide in enumerate(slides):\n image = openslide.OpenSlide(os.path.join(data_dir, f'{slide}.tiff'))\n spacing = 1 / (float(image.properties['tiff.XResolution']) / 10000)\n patch = image.read_region((1780,1950), 0, (256, 256))\n ax[i//3, i%3].imshow(patch) \n image.close() \n ax[i//3, i%3].axis('off')\n \n image_id = slide\n data_provider = train.loc[slide, 'data_provider']\n isup_grade = train.loc[slide, 'isup_grade']\n gleason_score = train.loc[slide, 'gleason_score']\n ax[i//3, i%3].set_title(f\"ID: {image_id}\\nSource: {data_provider} ISUP: {isup_grade} Gleason: {gleason_score}\")\n\n plt.show() \n","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"images = ['07a7ef0ba3bb0d6564a73f4f3e1c2293','037504061b9fba71ef6e24c48c6df44d','035b1edd3d1aeeffc77ce5d248a01a53','059cbf902c5e42972587c8d17d49efed','06a0cbd8fd6320ef1aa6f19342af2e68','06eda4a6faca84e84a781fee2d5f47e1','0a4b7a7499ed55c71033cefb0765e93d','0838c82917cd9af681df249264d2769c','046b35ae95374bfb48cdca8d7c83233f','074c3e01525681a275a42282cd21cbde',\n'05abe25c883d508ecc15b6e857e59f32','05f4e9415af9fdabc19109c980daf5ad','060121a06476ef401d8a21d6567dee6d','068b0e3be4c35ea983f77accf8351cc8','08f055372c7b8a7e1df97c6586542ac8'\n]\n\n","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"display_images(images)","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"def display_masks(slides): \n f, ax = plt.subplots(5,3, figsize=(18,22))\n for i, slide in enumerate(slides):\n \n mask = openslide.OpenSlide(os.path.join(mask_dir))\n mask_data = mask.read_region((0,0), mask.level_count - 1, mask.level_dimensions[-1])\n cmap = matplotlib.colors.ListedColormap(['black', 'gray', 'green', 'yellow', 'orange', 'red'])\n\n ax[i//3, i%3].imshow(np.asarray(mask_data)[:,:,0], cmap=cmap, interpolation='nearest', vmin=0, vmax=5) \n mask.close() \n ax[i//3, i%3].axis('off')\n \n image_id = slide\n data_provider = train.loc[slide, 'data_provider']\n isup_grade = train.loc[slide, 'isup_grade']\n gleason_score = train.loc[slide, 'gleason_score']\n ax[i//3, i%3].set_title(f\"ID: {image_id}\\nSource: {data_provider} ISUP: {isup_grade} Gleason: {gleason_score}\")\n f.tight_layout()\n \n plt.show()\n","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"display_masks(data_sample)","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"data_providers = ['karolinska', 'radboud']\ntrain_df = pd.read_csv(f'{BASE_PATH}/train.csv')\nmasks = os.listdir(mask_dir)\nmasks_df = pd.Series(masks).to_frame()\nmasks_df.columns = ['mask_file_name']\nmasks_df['image_id'] = masks_df.mask_file_name.apply(lambda x: x.split('_')[0])\ntrain_df = pd.merge(train_df, masks_df, on='image_id', how='outer')\ndel masks_df\nprint(f\"There are {len(train_df[train_df.mask_file_name.isna()])} images without a mask.\")\n\n## removing items where image mask is null\ntrain_df = train_df[~train_df.mask_file_name.isna()]","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"display_masks(data_sample)\n","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"markdown","source":"few images and associated masks for samples with Gleason Score(5+5)"},{"metadata":{"trusted":true},"cell_type":"code","source":"sample_images = list(train.loc[train.gleason_score==\"5+5\", \"image_id\"])\nsample_images = [ '08459aaedfda0679aab403ababbd6ece','0a848ccbbb065ef5ee59dd01710f8531', '0bbbb6734f721f4df4d2ba60ade0ed15', \n '0bd231c85b2695e2cf021299e67a6afc', '0efdb66c93d6b474d93dfe41e40be6ca', '1364c10e1e7f1ad0457f649a44d74888', \n '1e644a98460e4f7ea50717720a001efd', '1fb65315d7ded63d688194863a1b123e', '244d9617bd58fa1db73ab4c1f40d298e']\ndata_sample = train.loc[train.image_id.isin(sample_images)]\nshow_images(data_sample)\n","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"def load_and_resize_image(img_id):\n \"\"\"\n Edited from https://www.kaggle.com/xhlulu/panda-resize-and-save-train-data\n \"\"\"\n biopsy = skimage.io.MultiImage(os.path.join(data_dir, f'{img_id}.tiff'))\n return cv2.resize(biopsy[-1], (512, 512))\n\ndef load_and_resize_mask(img_id):\n \"\"\"\n Edited from https://www.kaggle.com/xhlulu/panda-resize-and-save-train-data\n \"\"\"\n biopsy = skimage.io.MultiImage(os.path.join(mask_dir, f'{img_id}_mask.tiff'))\n return cv2.resize(biopsy[-1], (512, 512))[:,:,0]","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"\nlabels = []\nfor grade in range(train.isup_grade.nunique()):\n fig, ax = plt.subplots(nrows=4, ncols=4, figsize=(22, 22))\n\n for i, row in enumerate(ax):\n idx = i//2\n temp = train_df[(train_df.isup_grade == grade) & (train_df.data_provider == data_providers[idx])].image_id.head(4).reset_index(drop=True)\n if i%2 < 1:\n labels.append(f'{data_providers[idx]} (image)')\n for j, col in enumerate(row):\n col.imshow(load_and_resize_image(temp[j]))\n col.set_title(f\"ID: {temp[j]}\")\n \n else:\n labels.append(f'{data_providers[idx]} (mask)')\n for j, col in enumerate(row):\n if data_providers[idx] == 'radboud':\n col.imshow(load_and_resize_mask(temp[j]), \n cmap = matplotlib.colors.ListedColormap(['white', 'lightgrey', 'green', 'orange', 'red', 'darkred']), \n norm = matplotlib.colors.Normalize(vmin=0, vmax=5, clip=True))\n else:\n col.imshow(load_and_resize_mask(temp[j]), \n cmap = matplotlib.colors.ListedColormap(['white', 'green', 'red']), \n norm = matplotlib.colors.Normalize(vmin=0, vmax=2, clip=True))\n \n gleason_score = train.loc[temp[j], 'gleason_score']\n col.set_title(f\"ID: {temp[j]}\")\n \n for row, r in zip(ax[:,0], labels):\n row.set_ylabel(r, rotation=90, size='large', fontsize=14)\n\n plt.suptitle(f'ISUP Grade {grade}', fontsize=20)\n plt.show()","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"Parse all images for train data to extract image characteristics"},{"metadata":{},"cell_type":"markdown","source":"# **Overlaying masks on the slides**"},{"metadata":{},"cell_type":"markdown","source":"As the masks have the same dimension as the slides ,we can overlay the masks on the tissue to directly see which areas are cancerous .This overlay can help you identifying the different growth patterns.To do this,we load both the mask and the biopsy and merge them using PIL. "},{"metadata":{},"cell_type":"raw","source":"Note : In the example below you can also observe a few pen marking slide (dark green smudges).These markings are not part of the tissue but were made by the pathologists who originally checked this case.These pen markings are available on some slides in the training set."},{"metadata":{"trusted":true},"cell_type":"code","source":"def overlay_mask_on_slide(images, center='radboud', alpha=0.8, max_size=(800, 800)):\n \"\"\"Show a mask overlayed on a slide.\"\"\"\n f, ax = plt.subplots(5,3, figsize=(18,22))\n \n \n for i, image_id in enumerate(images):\n slide = openslide.OpenSlide(os.path.join(data_dir, f'{image_id}.tiff'))\n mask = openslide.OpenSlide(os.path.join(mask_dir, f'{image_id}_mask.tiff'))\n slide_data = slide.read_region((0,0), slide.level_count - 1, slide.level_dimensions[-1])\n mask_data = mask.read_region((0,0), mask.level_count - 1, mask.level_dimensions[-1])\n mask_data = mask_data.split()[0]\n \n \n # Create alpha mask\n alpha_int = int(round(255*alpha))\n if center == 'radboud':\n alpha_content = np.less(mask_data.split()[0], 2).astype('uint8') * alpha_int + (255 - alpha_int)\n elif center == 'karolinska':\n alpha_content = np.less(mask_data.split()[0], 1).astype('uint8') * alpha_int + (255 - alpha_int)\n\n alpha_content = PIL.Image.fromarray(alpha_content)\n preview_palette = np.zeros(shape=768, dtype=int)\n\n if center == 'radboud':\n # Mapping: {0: background, 1: stroma, 2: benign epithelium, 3: Gleason 3, 4: Gleason 4, 5: Gleason 5}\n preview_palette[0:18] = (np.array([0, 0, 0, 0.5, 0.5, 0.5, 0, 1, 0, 1, 1, 0.7, 1, 0.5, 0, 1, 0, 0]) * 255).astype(int)\n elif center == 'karolinska':\n # Mapping: {0: background, 1: benign, 2: cancer}\n preview_palette[0:9] = (np.array([0, 0, 0, 0, 1, 0, 1, 0, 0]) * 255).astype(int)\n\n mask_data.putpalette(data=preview_palette.tolist())\n mask_rgb = mask_data.convert(mode='RGB')\n overlayed_image = PIL.Image.composite(image1=slide_data, image2=mask_rgb, mask=alpha_content)\n overlayed_image.thumbnail(size=max_size, resample=0)\n\n \n ax[i//3, i%3].imshow(overlayed_image) \n slide.close()\n mask.close() \n ax[i//3, i%3].axis('off')\n \n data_provider = train.loc[image_id, 'data_provider']\n isup_grade = train.loc[image_id, 'isup_grade']\n gleason_score = train.loc[image_id, 'gleason_score']\n ax[i//3, i%3].set_title(f\"ID: {image_id}\\nSource: {data_provider} ISUP: {isup_grade} Gleason: {gleason_score}\")\n","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"overlay_mask_on_slide(images)","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"# Expolring images with pen markers"},{"metadata":{},"cell_type":"markdown","source":"It is mentioned that in training dataset,there are few images with pen markers on them.The organizers left us with a Note as described below.\nNote:that slightly different procedures were in place for the images used in the test set than training set.Some of the training set images have stray pen marks on them,but the test slides are free of pen marks."},{"metadata":{"trusted":true},"cell_type":"code","source":"overlay_mask_on_slide(images)","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"pen_marked_images = [\n \n 'ebb6a080d72e09f6481721ef9f88c472',\n 'ebb6d5ca45942536f78beb451ee43cc4',\n 'ea9d52d65500acc9b9d89eb6b82cdcdf',\n 'e726a8eac36c3d91c3c4f9edba8ba713',\n 'e90abe191f61b6fed6d6781c8305fe4b',\n 'fd0bb45eba479a7f7d953f41d574bf9f',\n 'ff10f937c3d52eff6ad4dd733f2bc3ac',\n 'feee2e895355a921f2b75b54debad328',\n 'feac91652a1c5accff08217d19116f1c',\n 'fb01a0a69517bb47d7f4699b6217f69d',\n 'f00ec753b5618cfb30519db0947fe724',\n 'e9a4f528b33479412ee019e155e1a197',\n 'f062f6c1128e0e9d51a76747d9018849',\n 'f39bf22d9a2f313425ee201932bac91a',\n]","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"overlay_mask_on_slide(pen_marked_images)","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"import time\nstart_time = time.time()\nslide_dimensions, spacings, level_counts = [], [], []\n\nfor image_id in train.image_id:\n image = str(image_id)+'.tiff'\n image_path = os.path.join(PATH,\"train_images\",image)\n slide = openslide.OpenSlide(image_path)\n spacing = 1 / (float(slide.properties['tiff.XResolution']) / 10000)\n slide_dimensions.append(slide.dimensions)\n spacings.append(spacing)\n level_counts.append(slide.level_count)\n slide.close()\n del slide\n\ntrain['width'] = [i[0] for i in slide_dimensions]\ntrain['height'] = [i[1] for i in slide_dimensions]\ntrain['spacing'] = spacings\ntrain['level_count'] = level_counts\n\nend_time = time.time()\nprint(f\"Total processing time: {round(end_time - start_time,2)} sec.\")","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"markdown","source":"# Model"},{"metadata":{"trusted":true},"cell_type":"code","source":"\n\nclass MishFunction(torch.autograd.Function):\n @staticmethod\n def forward(ctx, x):\n ctx.save_for_backward(x)\n return x * torch.tanh(F.softplus(x)) # x * tanh(ln(1 + exp(x)))\n\n @staticmethod\n def backward(ctx, grad_output):\n x = ctx.saved_variables[0]\n sigmoid = torch.sigmoid(x)\n tanh_sp = torch.tanh(F.softplus(x)) \n return grad_output * (tanh_sp + x * sigmoid * (1 - tanh_sp * tanh_sp))\n\nclass Mish(nn.Module):\n def forward(self, x):\n return MishFunction.apply(x)\n\ndef to_Mish(model):\n for child_name, child in model.named_children():\n if isinstance(child, nn.ReLU):\n setattr(model, child_name, Mish())\n else:\n to_Mish(child)\n","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"def _resnext(url,block,layers,pretrained,progress,**kwargs):\n model=ResNet(block,layers,**kwargs)\n return model\nclass Model(nn.Module):\n def __init__(self,arch='resnext50_32x4d',n=6,pre=True):\n super().__init__()\n m=_resnext(semi_supervised_model_urls[arch],Bottleneck,[3,4,6,3],False,progress=False,groups=32,width_per_group=4)\n self.enc=nn.Sequential(*list(m.children())[:-2])\n nc=list(m.children())[-1].in_features\n self.head = nn.Sequential(AdaptiveConcatPool2d(),Flatten(),nn.Linear(2*nc,512),\n Mish(),nn.BatchNorm1d(512),nn.Dropout(0.5),nn.Linear(512,n))\n \n def forward(self, x):\n shape = x.shape\n n = shape[1]\n x = x.view(-1,shape[2],shape[3],shape[4])\n x = self.enc(x)\n shape = x.shape\n x = x.view(-1,n,shape[1],shape[2],shape[3]).permute(0,2,1,3,4).contiguous()\\\n .view(-1,shape[1],shape[2]*n,shape[3])\n x = self.head(x)\n return x\nmodels = []","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"models = []\nfor path in MODELS:\n state_dict = torch.load(path,map_location=torch.device('cpu'))\n model = Model()\n model.load_state_dict(state_dict)\n model.float()\n model.eval()\n model.cuda()\n models.append(model)","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"del state_dict","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"def tile(img):\n shape = img.shape\n pad0,pad1 = (sz - shape[0]%sz)%sz, (sz - shape[1]%sz)%sz\n img = np.pad(img,[[pad0//2,pad0-pad0//2],[pad1//2,pad1-pad1//2],[0,0]],\n constant_values=255)\n img = img.reshape(img.shape[0]//sz,sz,img.shape[1]//sz,sz,3)\n img = img.transpose(0,2,1,3,4).reshape(-1,sz,sz,3)\n if len(img) < N:\n img = np.pad(img,[[0,N-len(img)],[0,0],[0,0],[0,0]],constant_values=255)\n idxs = np.argsort(img.reshape(img.shape[0],-1).sum(-1))[:N]\n img = img[idxs]\n return img\n\nmean = torch.tensor([1.0-0.90949707, 1.0-0.8188697, 1.0-0.87795304])\nstd = torch.tensor([0.36357649, 0.49984502, 0.40477625])\n\nclass PandaDataset(Dataset):\n def __init__(self, path, test):\n self.path = path\n self.names = list(pd.read_csv(test).image_id)\n\n def __len__(self):\n return len(self.names)\n\n def __getitem__(self, idx):\n name = self.names[idx]\n img = skimage.io.MultiImage(os.path.join(DATA,name+'.tiff'))[-1]\n tiles = torch.Tensor(1.0 - tile(img)/255.0)\n tiles = (tiles - mean)/std\n return tiles.permute(0,3,1,2), name","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"# Prediction"},{"metadata":{"trusted":true},"cell_type":"code","source":"sub_df = pd.read_csv(SAMPLE)\nif os.path.exists(DATA):\n ds = PandaDataset(DATA,TEST)\n dl = DataLoader(ds, batch_size=bs, num_workers=nworkers, shuffle=False)\n names,preds = [],[]\n\n with torch.no_grad():\n for x,y in tqdm(dl):\n x = x.cuda()\n #dihedral TTA\n x = torch.stack([x,x.flip(-1),x.flip(-2),x.flip(-1,-2),\n x.transpose(-1,-2),x.transpose(-1,-2).flip(-1),\n x.transpose(-1,-2).flip(-2),x.transpose(-1,-2).flip(-1,-2)],1)\n x = x.view(-1,N,3,sz,sz)\n p = [model(x) for model in models]\n p = torch.stack(p,1)\n p = p.view(bs,8*len(models),-1).mean(1).argmax(-1).cpu()\n names.append(y)\n preds.append(p)\n \n names = np.concatenate(names)\n preds = torch.cat(preds).numpy()\n sub_df = pd.DataFrame({'image_id': names, 'isup_grade': preds})\n \n","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"sub_df.to_csv(\"submission.csv\", index=False)\nsub_df.head(5)","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"","execution_count":null,"outputs":[]}],"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} |