import os os.system ("pip install keras==2.4.3") import keras print('The keras version is {}.'.format(keras.__version__))

# IPython display functions import IPython from IPython.display import display, HTML, SVG, Image # General Plotting import matplotlib.pyplot as plt plt.style.use('seaborn-paper') plt.rcParams['figure.figsize'] = [10, 6] ## plot size plt.rcParams['axes.linewidth'] = 2.0 #set the value globally ## notebook style and settings display(HTML("<style>.container { width:90% !important; }</style>")) display(HTML("<style>.output_png { display: table-cell; text-align: center; vertical-align: middle; } </style>")) display(HTML("<style>.MathJax {font-size: 100%;}</style>")) # For changing background color def set_background(color): script = ( "var cell = this.closest('.code_cell');" "var editor = cell.querySelector('.input_area');" "editor.style.background='{}';" "this.parentNode.removeChild(this)" ).format(color) display(HTML('<img src onerror="{}">'.format(script)))

import os import sys import random import numpy as np import pandas as pd from os import walk # Metrics from sklearn.metrics import * # Keras library for deep learning # https://keras.io/ import tensorflow as tf import keras from keras.datasets import mnist # MNIST Data set from keras.models import Sequential # Model building from keras.layers import * # Model layers from keras.preprocessing.image import * from sklearn.model_selection import train_test_split np.float32 import warnings warnings.simplefilter(action='ignore', category=FutureWarning) from tensorflow.keras.layers import Activation from tensorflow.keras.utils import get_custom_objects class Mish(Activation): ''' Mish Activation Function. .. math:: mish(x) = x * tanh(softplus(x)) = x * tanh(ln(1 + e^{x})) Shape: - Input: Arbitrary. Use the keyword argument `input_shape` (tuple of integers, does not include the samples axis) when using this layer as the first layer in a model. - Output: Same shape as the input. Examples: >>> X = Activation('Mish', name="conv1_act")(X_input) ''' def __init__(self, activation, **kwargs): super(Mish, self).__init__(activation, **kwargs) self.__name__ = 'Mish' def mish(inputs): return inputs * tf.math.tanh(tf.math.softplus(inputs)) get_custom_objects().update({'Mish': Mish(mish)})

def displayConfusionMatrix(confusionMatrix, precisionNegative, precisionPositive, recallNegative, recallPositive, title): # Set font size for the plots. You can ignore this line. PLOT_FONT_SIZE = 14 # Set plot size. Please ignore this line plt.rcParams['figure.figsize'] = [5, 5] # Transpose of confusion matrix to align the plot with the actual precision recall values. Please ignore this as well. confusionMatrix = np.transpose(confusionMatrix) # Plotting the confusion matrix plt.imshow(confusionMatrix, interpolation='nearest',cmap=plt.cm.Blues, vmin=0, vmax=100) # Setting plot properties. You should ignore everything from here on. xticks = np.array([-0.5, 0, 1,1.5]) plt.gca().set_xticks(xticks) plt.gca().set_yticks(xticks) plt.gca().set_xticklabels(["", "Healthy\nRecall=" + str(recallNegative) , "Pneumonia\nRecall=" + str(recallPositive), ""], fontsize=PLOT_FONT_SIZE) plt.gca().set_yticklabels(["", "Healthy\nPrecision=" + str(precisionNegative) , "Pneumonia\nPrecision=" + str(precisionPositive), ""], fontsize=PLOT_FONT_SIZE) plt.ylabel("Predicted Class", fontsize=PLOT_FONT_SIZE) plt.xlabel("Actual Class", fontsize=PLOT_FONT_SIZE) plt.title(title, fontsize=PLOT_FONT_SIZE) # Add text in heatmap boxes for i in range(2): for j in range(2): text = plt.text(j, i, confusionMatrix[i][j], ha="center", va="center", color="white", size=15) ### size here is the size of text inside a single box in the heatmap plt.show()

def calculateMetricsAndPrint(predictions, predictionsProbabilities, actualLabels): # Convert label format from [0,1](label 1) and [1,0](label 0) into single integers: 1 and 0. actualLabels = [item[1] for item in actualLabels] # Get probabilities for the class with label 1. That is all we need to compute AUCs. We don't need probabilities for class 0. predictionsProbabilities = [item[1] for item in predictionsProbabilities] # Calculate metrics using scikit-learn functions. The round function is used to round the numbers up to 2 decimal points. accuracy = round(accuracy_score(actualLabels, predictions) * 100, 2) precisionNegative = round(precision_score(actualLabels, predictions, average = None)[0] * 100, 2) precisionPositive = round(precision_score(actualLabels, predictions, average = None)[1] * 100, 2) recallNegative = round(recall_score(actualLabels, predictions, average = None)[0] * 100, 2) recallPositive = round(recall_score(actualLabels, predictions, average = None)[1] * 100, 2) auc = round(roc_auc_score(actualLabels, predictionsProbabilities) * 100, 2) confusionMatrix = confusion_matrix(actualLabels, predictions) # Print metrics. .%2f prints a number upto 2 decimal points only. print("------------------------------------------------------------------------") print("Accuracy: %.2f\nPrecisionNegative: %.2f\nPrecisionPositive: %.2f\nRecallNegative: %.2f\nRecallPositive: %.2f\nAUC Score: %.2f" % (accuracy, precisionNegative, precisionPositive, recallNegative, recallPositive, auc)) print("------------------------------------------------------------------------") print("+ Printing confusion matrix...\n") # Display confusion matrix displayConfusionMatrix(confusionMatrix, precisionNegative, precisionPositive, recallNegative, recallPositive, "Confusion Matrix") print("+ Printing ROC curve...\n") # ROC Curve plt.rcParams['figure.figsize'] = [16, 8] FONT_SIZE = 16 falsePositiveRateDt, truePositiveRateDt, _ = roc_curve(actualLabels, predictionsProbabilities) plt.plot(falsePositiveRateDt, truePositiveRateDt, linewidth = 5, color='black') plt.xticks(fontsize=FONT_SIZE) plt.yticks(fontsize=FONT_SIZE) plt.xlabel("False Positive Rate", fontsize=FONT_SIZE) plt.ylabel("True Positive Rate", fontsize=FONT_SIZE) plt.show() return auc

def calculateMetrics(predictions, predictionsProbabilities, actualLabels): # Convert label format from [0,1](label 1) and [1,0](label 0) into single integers: 1 and 0. actualLabels = [item[1] for item in actualLabels] # Get probabilities for the class with label 1. That is all we need to compute AUCs. We don't need probabilities for class 0. predictionsProbabilities = [item[1] for item in predictionsProbabilities] # Calculate metrics using scikit-learn functions. The round function is used to round the numbers up to 2 decimal points. try: accuracy = round(accuracy_score(actualLabels, predictions) * 100, 2) precisionNegative = round(precision_score(actualLabels, predictions, average = None)[0] * 100, 2) precisionPositive = round(precision_score(actualLabels, predictions, average = None)[1] * 100, 2) recallNegative = round(recall_score(actualLabels, predictions, average = None)[0] * 100, 2) recallPositive = round(recall_score(actualLabels, predictions, average = None)[1] * 100, 2) except: print("An exception occurred but was caught.") auc = round(roc_auc_score(actualLabels, predictionsProbabilities) * 100, 2) return auc

def getKagglePredictions(model, kaggleData, filename): print("+ Writing kaggle test results in : results/%s..." % filename) predictions = model.predict(kaggleData) predictionProbs = [item[1] for item in predictions] # Store predictions for kaggle outputFile = open("results/" + str(filename), "w") outputFile.write("Id,Prediction\n") for i in range(0, len(predictionProbs)): outputFile.write(str(i + 1) + "," + str(predictionProbs[i]) + "\n") outputFile.close()

def calculateClasswiseTopNAccuracy(actualLabels, predictionsProbs, TOP_N): """ TOP_N is the top n% predictions you want to use for each class """ discreteActualLabels = [1 if item[1] > item[0] else 0 for item in actualLabels] discretePredictions = [1 if item[1] > item[0] else 0 for item in predictionsProbs] predictionProbsTopNHealthy, predictionProbsTopNPneumonia = [item[0] for item in predictionsProbs], [item[1] for item in predictionsProbs] predictionProbsTopNHealthy = list(reversed(sorted(predictionProbsTopNHealthy)))[:int(len(predictionProbsTopNHealthy) * TOP_N / 100)][-1] predictionProbsTopNPneumonia = list(reversed(sorted(predictionProbsTopNPneumonia)))[:int(len(predictionProbsTopNPneumonia) * TOP_N / 100)][-1] # Calculate accuracy for both classes accuracyHealthy = [] accuracyPneumonia = [] for i in range(0, len(discretePredictions)): if discretePredictions[i] == 1: # Pneumonia if predictionsProbs[i][1] > predictionProbsTopNPneumonia: accuracyPneumonia.append(int(discreteActualLabels[i]) == 1) else: # Healthy if predictionsProbs[i][0] > predictionProbsTopNHealthy: accuracyHealthy.append(int(discreteActualLabels[i]) == 0) accuracyHealthy = round((accuracyHealthy.count(True) * 100) / len(accuracyHealthy), 2) accuracyPneumonia = round((accuracyPneumonia.count(True) * 100) / len(accuracyPneumonia), 2) return accuracyHealthy, accuracyPneumonia

# Load normal images normalImagesPath = "data/train/normal" normalImageFiles = [] for(_,_,files) in walk(normalImagesPath): normalImageFiles.extend(files) normalImagesPath2 = "data/train/normal2" for(_,_,files) in walk(normalImagesPath2): normalImageFiles.extend(files) print(len(normalImageFiles)) # Load pneumonia images pneumoniaImagesPath = "data/train/pneumonia" pneumoniaImageFiles = [] for(_,_,files) in walk(pneumoniaImagesPath): pneumoniaImageFiles.extend(files) random.shuffle(pneumoniaImageFiles) pneumoniaImageFiles = pneumoniaImageFiles[:len(normalImageFiles)] print("Normal X-ray images: %d\nPneumonia X-ray images: %d" % (len(normalImageFiles), len(pneumoniaImageFiles)))

imagesData = [] imagesLabels = [] for file in normalImageFiles: fullPath = normalImagesPath + "/" + file if os.path.exists(fullPath) == False: continue imageData = load_img(normalImagesPath + "/" + file, color_mode = "grayscale") # load_img function comes from keras library when we do "from keras.preprocessing.image import *" imageArray = img_to_array(imageData) / 255.0 imagesData.append(imageArray) imagesLabels.append(0) for file in pneumoniaImageFiles: fullPath = pneumoniaImagesPath + "/" + file if os.path.exists(fullPath) == False: continue imageData = load_img(pneumoniaImagesPath + "/" + file, color_mode = "grayscale") # load_img function comes from keras library when we do "from keras.preprocessing.image import *" imageArray = img_to_array(imageData) / 255.0 imagesData.append(imageArray) imagesLabels.append(1) imagesData = np.array(imagesData) imagesLabels = keras.utils.to_categorical(imagesLabels) print("Input data shape: %s" % (imagesData.shape,))

testImagesPath = "data/test/" testImageFiles = [] for(_,_,files) in walk(testImagesPath): testImageFiles.extend(files) testImageFiles = list(sorted(testImageFiles)) kaggleTestImages = [] for file in testImageFiles: fullPath = testImagesPath + "/" + file if os.path.exists(fullPath) == False: continue imageData = load_img(testImagesPath + "/" + file, color_mode = "grayscale") # load_img function comes from keras library when we do "from keras.preprocessing.image import *" imageArray = img_to_array(imageData) / 255.0 kaggleTestImages.append(imageArray) kaggleTestImages = np.array(kaggleTestImages) print("Number of test images: %d" % len(kaggleTestImages))

def trainTestSplit(data, labels): """ 80-20 train-test data split """ trainData, trainLabels, testData, testLabels = [], [], [], [] for i in range(0, len(data)): if i % 5 == 0: testData.append(data[i]) testLabels.append(labels[i]) else: trainData.append(data[i]) trainLabels.append(labels[i]) return np.array(trainData), np.array(testData), np.array(trainLabels), np.array(testLabels)

# In our context, since we have a private test data on kaggle, our test data here would actually mean validation data. We will use results on this validation(test) data to see how our model would perform on the actual test data. # Split data into 80% training and 20% testing trainData, testData, trainLabels, testLabels = trainTestSplit(imagesData, imagesLabels)

def createParameterizedConvolutionalNeuralNetwork(trainImages, numLayers, numFilters, kernelSize, maxPooling, dropoutValue, learningRate, numClasses): # Create model object model = Sequential() # Add the first layer with dropout model.add(Conv2D(numFilters, kernel_size=(kernelSize, kernelSize), activation='tanh', padding = 'same', input_shape=trainImages.shape[1:])) model.add(MaxPooling2D(pool_size=(maxPooling, maxPooling))) model.add(Dropout(dropoutValue)) while numLayers > 1: model.add(Conv2D(numFilters, kernel_size=(kernelSize, kernelSize), activation='tanh', padding = 'same')) model.add(MaxPooling2D(pool_size=(maxPooling, maxPooling))) model.add(Dropout(dropoutValue)) numLayers = numLayers - 1 # Convolutional layers are done, adding the remaining stuff. Please note that after conv layers, you should always use a Flatten() layer. model.add(Flatten()) model.add(Dense(64, activation='tanh')) model.add(BatchNormalization(momentum=0.99)) model.add(Dropout(dropoutValue)) model.add(Dense(numClasses, activation='softmax')) # Compile model. You can skip this line. model.compile(loss=keras.losses.categorical_crossentropy, optimizer=keras.optimizers.Adam(lr=learningRate), metrics=['accuracy']) # Return model return model

def createNuancedConvolutionalNeuralNetwork(trainImages, numClasses): """ You should try to edit this model as much as you can. Try adding/removing layers, setting different parameters for different layers etc. You have complete control of the model and you should try different things to see what works and what does not. """ # Create model object model = Sequential() # Add the first layer with dropout model.add(Conv2D(filters = 64, kernel_size=(5, 5), activation='elu', padding = 'same', input_shape=trainImages.shape[1:])) model.add(MaxPooling2D(pool_size=(3, 3))) model.add(Dropout(0.5)) # Second layer with diffefiltersrent parameters model.add(Conv2D(filters = 32, kernel_size=(3, 3), activation='elu', padding = 'same')) model.add(MaxPooling2D(pool_size=(2, 2))) model.add(Dropout(0.2)) # Convolutional layers are done, adding the remaining stuff. Please note that after conv layers, you should always use a Flatten() layer. model.add(Flatten()) model.add(Dense(64, activation='elu')) model.add(Dense(32, activation='elu')) model.add(Dropout(0.5)) model.add(Dense(numClasses, activation='softmax')) # Compile model. You can skip this line. model.compile(loss=keras.losses.categorical_crossentropy, optimizer=keras.optimizers.Adam(lr=0.0001), metrics=['accuracy']) # Return model return model

set_background('#fce53a') ##################################################################################################################################################### # Things you can change ##################################################################################################################################################### # You can change all these parameters for different results. Please go to the following links to read more about each parameter: # https://keras.io/preprocessing/image/ # https://www.pyimagesearch.com/2019/07/08/keras-imagedatagenerator-and-data-augmentation/ dataAugmentation = ImageDataGenerator( rotation_range=12, width_shift_range=0.08, height_shift_range=0.08, horizontal_flip=True, vertical_flip=False, shear_range=0.0, zoom_range=0.38)

set_background('#fce53a') ##################################################################################################################################################### # Things you can change ##################################################################################################################################################### numLayers = 3 # Number of layers in the neural network numFilters = 55 # Number of units in each layer kernelSize = 22 # filter size of a single filter dropoutValue = 0.29 # Dropout probability maxPooling = 3 # Max pooling numClasses = 2 # Don't change this value for pneumonia since we have 2 classes i.e we are trying to recognize a digit among 10 digits. But for any other data set, this should be changed batchSize = 64 # How many images should a single batch contain learningRate = 0.00004807256976 # How fast should the model learn epochs = 200 # Number of epochs to train your model for USE_DATA_AUGMENTATION = True # You can set it to false if you do not want to use data augmentation. We recommend trying both true and false. ##################################################################################################################################################### # Please do not change this line. dataAugmentation.fit(trainData) # Training the augmentor in case we set USE_DATA_AUGMENTATION to True.

# Create model parameterizedModel = createParameterizedConvolutionalNeuralNetwork(trainData, numLayers, numFilters, kernelSize, maxPooling, dropoutValue, learningRate, numClasses = 2) print("+ Your parameterized model has been created...")

# You can create the other model with the following line nonParameterizedModel = createNuancedConvolutionalNeuralNetwork(imagesData, numClasses = 2) print("+ Your non parameterized model has been created...")

##################################################################################################################################################### # Things you can change ##################################################################################################################################################### # Please assign model the deep learning model you want to use i.e parameterizedModel or nonParameterizedModel model = parameterizedModel

bestAcc = 0.0 bestEpoch = 0 bestAccPredictions, bestAccPredictionProbabilities = [], [] print("+ Starting training. Each epoch can take about 2-5 minutes, hold tight!") print("-----------------------------------------------------------------------\n") for epoch in range(epochs): #################################################### Model Training ############################################################### if USE_DATA_AUGMENTATION == True: # Use data augmentation in alternate epochs if epoch % 2 == 0: # Alternate between training with and without augmented data. Training just on the augmented data might not be the best way to go. ############ You can change the "epoch % 2" to some other integer value to train on top of the augmented data ############ after a certain number of epochs e.g "epoch % 3" will train on augmented data after every 2 epochs ############ model.fit_generator(dataAugmentation.flow(trainData, trainLabels, batch_size=batchSize), steps_per_epoch=len(trainData) / batchSize, epochs=1, verbose = 2) else: model.fit(trainData, trainLabels, batch_size=batchSize, epochs=1, verbose = 2) else: # Do not use data augmentation model.fit(trainData, trainLabels, batch_size=batchSize, epochs=1, verbose = 2) #################################################### Model Testing ############################################################### # Calculate test accuracy accuracy = round(model.evaluate(testData, testLabels)[1] * 100, 3) predictions = model.predict(testData) #get epoch-level AUCs AccPredictionProbabilities = model.predict(testData) AccPredictions = [1 if item[1] > item[0] else 0 for item in AccPredictionProbabilities] epochAUC = calculateMetrics(AccPredictions, AccPredictionProbabilities, testLabels) print("+ Test accuracy at epoch %d is: %.2f " % (epoch, accuracy)) print("+ Test AUC at epoch %d is: %.3f " % (epoch, epochAUC)) if accuracy > bestAcc: bestEpoch = epoch bestAcc = accuracy bestAccPredictions = [1 if item[1] > item[0] else 0 for item in predictions] bestAccPredictionProbabilities = predictions ##################################### Store predictions for kaggle ########################################################### kaggleResultsFileName = "epoch-" + str(epoch) + "-results.csv" getKagglePredictions(model, kaggleTestImages, kaggleResultsFileName) ############################################################################################################################## print('\n') print("------------------------------------------------------------------------") ##################################################### Printing best metrics ########################################################## # Get more metrics for the best performing epoch print("\n*** Printing our best validation results that we obtained in epoch %d ..." % bestEpoch) calculateMetricsAndPrint(bestAccPredictions, bestAccPredictionProbabilities, testLabels)

################################## You can change values inside the following list ########################### topNValues = [5, 10, 20, 30,40, 50, 60, 70, 80, 90,100] ############################################################################################################## accuraciesHealthy, accuraciesPneumonia = [], [] for topn in topNValues: accuracyHealthy, accuracyPneumonia = calculateClasswiseTopNAccuracy(testLabels, bestAccPredictionProbabilities, topn) accuraciesHealthy.append(accuracyHealthy) accuraciesPneumonia.append(accuracyPneumonia) print("+ Accuracy for top %d percent predictions for healthy: %.2f, pneumonia: %.2f" % (topn, accuracyHealthy, accuracyPneumonia)) # Plot results x = np.arange(len(accuraciesHealthy)) plt.plot(x, accuraciesHealthy, linewidth = 3, color = '#e01111') scatterHealthy = plt.scatter(x, accuraciesHealthy, marker = 's', s = 100, color = '#e01111') plt.plot(x, accuraciesPneumonia, linewidth = 3, color = '#0072ff') scatterPneumonia = plt.scatter(x, accuraciesPneumonia, marker = 'o', s = 100, color = '#0072ff') plt.xticks(x, topNValues, fontsize = 15) plt.yticks(fontsize = 15) plt.xlabel("Top N%", fontsize = 15) plt.ylabel("Accuracy", fontsize = 15) plt.legend([scatterHealthy, scatterPneumonia], ["Accuracy for Healthy", "Accuracy for Pneumonia"], fontsize = 17) plt.ylim(0, 110) plt.show()

print(model.summary())