kaggleのデータセットを使って、レントゲンの肺炎画像を診断する。
使用したデータセット： Chest X-Ray Images (Pneumonia)
今回挑戦すること
kaggleで公開されているデータセットを使用し、kaggleに慣れる。
kerasで自作モデルを作り、VGG-16転移学習モデルと比較を行う。
評価は混同行列の適合率及び再現率で行い、より再現率を重視する。
ディレクトリ構造
各ディレクトリ構造
X-ray
├─ train
├─ NORMAL(1341)
└─ PNEMONIA(3785)
├─ test
├─ NORMAL(234)
└─ PNEMONIA(390)
└─ val
├─ NORMAL(8)
└─ PNEMONIA(8)
下準備
ライブラリをインポート
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
from keras.utils.np_utils import to_categorical
from keras.applications.vgg16 import VGG16
from keras.layers import Dense, Dropout, Flatten, Conv2D, MaxPooling2D, Input, SeparableConv2D
from keras.layers.normalization import BatchNormalization
from keras.models import Sequential, Model
from keras.optimizers import Adam, SGD, RMSprop
from mlxtend.plotting import plot_confusion_matrix
from sklearn.metrics import confusion_matrix
import cv2
%matplotlib inline
import os
test_size = 0.8
ディレクトリを変数に代入
test_dir = "../input/chest-xray-pneumonia/chest_xray/test"
train_dir = "../input/chest-xray-pneumonia/chest_xray/train"
val_dir = "../input/chest-xray-pneumonia/chest_xray/val"
画像を取得する関数
def data_path(path, n_p):
'''
path: ディレクトリパス
n_p: path以下のNORMAL/PHEMONIAフォルダ
'''
# 引数からディレクトリパスを作成し、変数に格納
path = path + "/" + n_p
files = os.listdir(path)
# 画像を格納するリスト
t_list = []
# ファイルから画像を1つずつ取得
for imgfile in files:
img = path + "/" + imgfile
# 画像読み込み
img = cv2.imread(img)
# 画像サイズ変換
img = cv2.resize(img, (256, 256))
# 画像をリストに追加
t_list.append((img))
return t_list
自作モデルで学習・評価
Most Votesの方のコードを参考に作成してみました。
モデルクラスでモデルを作成
def my_model(shape):
# 入力層
input_img = Input(shape=shape, name="img_input")
# conv1 2回畳み込み+プーリング
x = Conv2D(128, (3,3), activation = "relu", padding= "same")(input_img)
x = Conv2D(128,(2,2), activation = "relu", padding= "same")(x)
x = MaxPooling2D((2,2))(x)
# conv2 2回畳み込み+プーリング
x = SeparableConv2D(128, (3,3), activation="relu", padding="same")(x)
x = SeparableConv2D(128, (3,3), activation="relu", padding = "same")(x)
x = MaxPooling2D((2,2))(x)
# conv3 2回畳み込み+正規化+畳み込み+プーリング
x = SeparableConv2D(256, (2,2), activation="relu", padding="same")(x)
x = SeparableConv2D(256, (2,2), activation="relu", padding="same")(x)
x = BatchNormalization()(x)
x = SeparableConv2D(256, (3, 3), activation="relu", padding = "same")(x)
x = MaxPooling2D((4,4))(x)
# conv4 畳み込み+正規化+畳み込み+正規化+畳み込み
x = SeparableConv2D(512, (3, 3), activation="relu", padding = "same")(x)
x = BatchNormalization()(x)
x = SeparableConv2D(512, (3, 3), activation="relu", padding = "same")(x)
x = BatchNormalization()(x)
x = SeparableConv2D(512, (3, 3), activation="relu", padding = "same")(x)
# 平坦化
x = Flatten()(x)
# 全結合層
x = Dense(512, activation="relu")(x)
x = Dropout(0.7)(x)
x = Dense(256, activation="relu")(x)
x = Dropout(0.6)(x)
x = Dense(64, activation="relu")(x)
x = Dropout(0.4)(x)
x = Dense(2, activation="softmax")(x)
model = Model(inputs = input_img, outputs=x)
return model
モデルを確認
model = my_model(shape=(256, 256,3))
model.summary()
>>
Model: "model_1"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
img_input (InputLayer) (None, 256, 256, 3) 0
_________________________________________________________________
conv2d_1 (Conv2D) (None, 256, 256, 128) 3584
_________________________________________________________________
conv2d_2 (Conv2D) (None, 256, 256, 128) 65664
_________________________________________________________________
max_pooling2d_1 (MaxPooling2 (None, 128, 128, 128) 0
_________________________________________________________________
separable_conv2d_1 (Separabl (None, 128, 128, 128) 17664
_________________________________________________________________
separable_conv2d_2 (Separabl (None, 128, 128, 128) 17664
_________________________________________________________________
max_pooling2d_2 (MaxPooling2 (None, 64, 64, 128) 0
_________________________________________________________________
separable_conv2d_3 (Separabl (None, 64, 64, 256) 33536
_________________________________________________________________
separable_conv2d_4 (Separabl (None, 64, 64, 256) 66816
_________________________________________________________________
batch_normalization_1 (Batch (None, 64, 64, 256) 1024
_________________________________________________________________
separable_conv2d_5 (Separabl (None, 64, 64, 256) 68096
_________________________________________________________________
max_pooling2d_3 (MaxPooling2 (None, 16, 16, 256) 0
_________________________________________________________________
separable_conv2d_6 (Separabl (None, 16, 16, 512) 133888
_________________________________________________________________
batch_normalization_2 (Batch (None, 16, 16, 512) 2048
_________________________________________________________________
separable_conv2d_7 (Separabl (None, 16, 16, 512) 267264
_________________________________________________________________
batch_normalization_3 (Batch (None, 16, 16, 512) 2048
_________________________________________________________________
separable_conv2d_8 (Separabl (None, 16, 16, 512) 267264
_________________________________________________________________
flatten_1 (Flatten) (None, 131072) 0
_________________________________________________________________
dense_1 (Dense) (None, 512) 67109376
_________________________________________________________________
dropout_1 (Dropout) (None, 512) 0
_________________________________________________________________
dense_2 (Dense) (None, 256) 131328
_________________________________________________________________
dropout_2 (Dropout) (None, 256) 0
_________________________________________________________________
dense_3 (Dense) (None, 64) 16448
_________________________________________________________________
dropout_3 (Dropout) (None, 64) 0
_________________________________________________________________
dense_4 (Dense) (None, 2) 130
=================================================================
Total params: 68,203,842
Trainable params: 68,201,282
Non-trainable params: 2,560
_________________________________________________________________
モデルを予測する関数
def pred_img(img):
img = cv2.resize(img, (256, 256))
pred = np.argmax(model.predict(np.array([img])))
if pred == 0:
return 'Normal'
else:
return 'Pneumonia'
学習用データの準備
# trainデータのラベル化
# 0: NORMAL, 1: PNEUMONIA
'''
trainデータをtrainとtestに分割
valデータが少なく、またtrainデータに偏りがあるためtrainデータを分割して使用。
'''
train_nor = data_path(train_dir, "NORMAL")
train_pne = data_path(train_dir, "PNEUMONIA")
X = np.array(train_nor + train_pne)
Y = np.array([0]*len(train_nor) + [1]*len(train_pne))
rand_index = np.random.permutation(np.arange(len(X)))
X = X[rand_index]
Y = Y[rand_index]
X_train = X[:int(len(X)*test_size)]
Y_train = Y[:int(len(Y)*test_size)]
X_test = X[int(len(X)*test_size):]
Y_test = Y[int(len(Y)*test_size):]
Y_train = to_categorical(Y_train)
Y_test = to_categorical(Y_test)
モデル学習
# 損失関数は2値分類なのでbinary_crossentropyを使用
model.compile(loss="binary_crossentropy",
optimizer = Adam(lr=1e-4, decay=1e-5), metrics=['accuracy'])
model.fit(X_train, Y_train, batch_size=16, epochs = 10)
>>
Epoch 1/10
4172/4172 [==============================] - 84s 20ms/step - loss: 0.4250 - accuracy: 0.7893
Epoch 2/10
4172/4172 [==============================] - 80s 19ms/step - loss: 0.2329 - accuracy: 0.9099
Epoch 3/10
4172/4172 [==============================] - 80s 19ms/step - loss: 0.1688 - accuracy: 0.9458
Epoch 4/10
4172/4172 [==============================] - 80s 19ms/step - loss: 0.1349 - accuracy: 0.9545
Epoch 5/10
4172/4172 [==============================] - 80s 19ms/step - loss: 0.1061 - accuracy: 0.9612
Epoch 6/10
4172/4172 [==============================] - 80s 19ms/step - loss: 0.0868 - accuracy: 0.9676
Epoch 7/10
4172/4172 [==============================] - 80s 19ms/step - loss: 0.0890 - accuracy: 0.9722
Epoch 8/10
4172/4172 [==============================] - 80s 19ms/step - loss: 0.0671 - accuracy: 0.9779
Epoch 9/10
4172/4172 [==============================] - 80s 19ms/step - loss: 0.0630 - accuracy: 0.9772
Epoch 10/10
4172/4172 [==============================] - 80s 19ms/step - loss: 0.0514 - accuracy: 0.9823
trainデータから分割したデータを予測
scores = model.evaluate(X_test, Y_test, verbose=1)
print('Test loss:', scores[0])
print('Test accuracy:', scores[1])
>>
1044/1044 [==============================] - 5s 5ms/step
Test loss: 0.060373730071681155
Test accuracy: 0.9779693484306335
テストデータの準備
test_nor = data_path(test_dir, "NORMAL")
test_pne = data_path(test_dir, "PNEUMONIA")
X_T = np.array(test_nor + test_pne)
Y_T = np.array([0]*len(test_nor) + [1]*len(test_pne))
rand_index = np.random.permutation(np.arange(len(X_T)))
X_T = X_T[rand_index]
Y_T = Y_T[rand_index]
Y_T = to_categorical(Y_T)
pred_img関数に写真を渡して予測
path_train_pne = os.listdir("../input/chest-xray-pneumonia/chest_xray/train/PNEUMONIA/")
for i in range(3):
img = cv2.imread(train_dir + "/PNEUMONIA/" + path_train_pne[i])
# print(img)
b,g,r = cv2.split(img)
img = cv2.merge([r,g,b])
plt.imshow(img)
plt.show()
print(pred_img(img))
テストデータの評価
scores = model.evaluate(X_T, Y_T, verbose=1)
print('loss:', scores[0])
print('accuracy:', scores[1])
>>
624/624 [==============================] - 3s 4ms/step
loss: 1.051183255819174
accuracy: 0.7644230723381042
混同行列
pred = model.predict(X_T)
pred = np.argmax(pred, axis=-1)
orig_test_label = np.argmax(Y_T, axis=-1)
cm = confusion_matrix(orig_test_label, pred)
plt.figure()
plot_confusion_matrix(cm,figsize=(12,8), hide_ticks=True, cmap=plt.cm.RdPu)
plt.xticks(range(2), ['Normal', 'Pneumonia'], fontsize=16)
plt.yticks(range(2), ['Normal', 'Pneumonia'], fontsize=16)
plt.show()
VGG16を用いた転移学習
VGG16の転移学習
input_tensor = Input(shape=(256, 256, 3))
vgg16 = VGG16(include_top=False, weights='imagenet', input_tensor=input_tensor)
top_model = Sequential()
top_model.add(Flatten(input_shape=vgg16.output_shape[1:]))
top_model.add(Dense(128, activation='relu'))
top_model.add(Dropout(0.2))
top_model.add(Dense(64, activation = "relu"))
top_model.add(Dense(2, activation = "softmax"))
モデルを学習
model2 = Model(input=vgg16.input, output=top_model(vgg16.output))
for layer in model.layers[:15]:
layer.trainable = False
model2.compile(loss='categorical_crossentropy',
optimizer=SGD(lr=1e-4, momentum=0.9),
metrics=['accuracy'])
model2.fit(X_train, Y_train, batch_size=100, epochs=10)
>>
Epoch 1/10
4172/4172 [==============================] - 50s 12ms/step - loss: 0.3822 - accuracy: 0.8993
Epoch 2/10
4172/4172 [==============================] - 40s 10ms/step - loss: 0.0939 - accuracy: 0.9652
Epoch 3/10
4172/4172 [==============================] - 40s 10ms/step - loss: 0.0565 - accuracy: 0.9789
Epoch 4/10
4172/4172 [==============================] - 40s 10ms/step - loss: 0.0387 - accuracy: 0.9851
Epoch 5/10
4172/4172 [==============================] - 40s 10ms/step - loss: 0.0227 - accuracy: 0.9935
Epoch 6/10
4172/4172 [==============================] - 40s 10ms/step - loss: 0.0225 - accuracy: 0.9916
Epoch 7/10
4172/4172 [==============================] - 40s 10ms/step - loss: 0.0147 - accuracy: 0.9957
Epoch 8/10
4172/4172 [==============================] - 40s 10ms/step - loss: 0.0089 - accuracy: 0.9976
Epoch 9/10
4172/4172 [==============================] - 40s 10ms/step - loss: 0.0080 - accuracy: 0.9976
Epoch 10/10
4172/4172 [==============================] - 41s 10ms/step - loss: 0.0092 - accuracy: 0.9966
モデルの確認
model2.summary()
>>
Model: "model_2"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
input_1 (InputLayer) (None, 256, 256, 3) 0
_________________________________________________________________
block1_conv1 (Conv2D) (None, 256, 256, 64) 1792
_________________________________________________________________
block1_conv2 (Conv2D) (None, 256, 256, 64) 36928
_________________________________________________________________
block1_pool (MaxPooling2D) (None, 128, 128, 64) 0
_________________________________________________________________
block2_conv1 (Conv2D) (None, 128, 128, 128) 73856
_________________________________________________________________
block2_conv2 (Conv2D) (None, 128, 128, 128) 147584
_________________________________________________________________
block2_pool (MaxPooling2D) (None, 64, 64, 128) 0
_________________________________________________________________
block3_conv1 (Conv2D) (None, 64, 64, 256) 295168
_________________________________________________________________
block3_conv2 (Conv2D) (None, 64, 64, 256) 590080
_________________________________________________________________
block3_conv3 (Conv2D) (None, 64, 64, 256) 590080
_________________________________________________________________
block3_pool (MaxPooling2D) (None, 32, 32, 256) 0
_________________________________________________________________
block4_conv1 (Conv2D) (None, 32, 32, 512) 1180160
_________________________________________________________________
block4_conv2 (Conv2D) (None, 32, 32, 512) 2359808
_________________________________________________________________
block4_conv3 (Conv2D) (None, 32, 32, 512) 2359808
_________________________________________________________________
block4_pool (MaxPooling2D) (None, 16, 16, 512) 0
_________________________________________________________________
block5_conv1 (Conv2D) (None, 16, 16, 512) 2359808
_________________________________________________________________
block5_conv2 (Conv2D) (None, 16, 16, 512) 2359808
_________________________________________________________________
block5_conv3 (Conv2D) (None, 16, 16, 512) 2359808
_________________________________________________________________
block5_pool (MaxPooling2D) (None, 8, 8, 512) 0
_________________________________________________________________
sequential_1 (Sequential) (None, 2) 4202818
=================================================================
Total params: 18,917,506
Trainable params: 18,917,506
Non-trainable params: 0
_________________________________________________________________
テストデータの評価
scores2 = model2.evaluate(X_T, Y_T, verbose=1)
print('loss2:', scores[0])
print('accuracy2:', scores[1])
>>
624/624 [==============================] - 3s 5ms/step
loss2: 1.051183255819174
accuracy2: 0.7644230723381042
混同行列
pred2 = model2.predict(X_T)
pred2 = np.argmax(pred2, axis=-1)
cm2 = confusion_matrix(orig_test_label, pred2)
plt.figure()
plot_confusion_matrix(cm2,figsize=(12,8), hide_ticks=True, cmap=plt.cm.RdPu)
plt.xticks(range(2), ['Normal', 'Pneumonia'], fontsize=16)
plt.yticks(range(2), ['Normal', 'Pneumonia'], fontsize=16)
plt.show()
適合率と再現率
'''
今回の場合
適合率(precision): 肺炎と予測した画像のうち、本当に肺炎画像である割合
再現率(recall): 肺炎画像を正しく肺炎と認識した割合
再現率が高く、適合率が低いと、画像を正しく認識していることになる。
=> 再現率と適合率はトレードオフだが、肺炎画像は見落としてはいけないので、今回の場合は再現率が重要視される。
'''
# 自作モデルの適合率と再現率
tn, fp, fn, tp = cm.ravel()
precision = tp/(tp+fp)
recall = tp/(tp+fn)
print("--my model--")
print("Recall {:.2f}".format(recall))
print("Precision {:.2f}".format(precision))
# VGG-16転移学習モデルの適合率と再現率
tn2, fp2, fn2, tp2 = cm2.ravel()
precision2 = tp2/(tp2+fp2)
recall2 = tp2/(tp2+fn2)
print("--VGG16--")
print("Recall {:.2f}".format(recall2))
print("Precision {:.2f}".format(precision2))
>>
--my model--
Recall 0.98
Precision 0.73
--VGG16--
Recall 1.00
Precision 0.71
結果
今回重要視したいRecallは、VGG-16で転移学習したモデルで100%、自作モデルは98%であった。
参考資料
Kerasで始めるModel作成方法の違い
Kerasでちょっと難しいModelやTrainingを実装するときのTips
Kerasによる2クラス分類（Pima Indians Diabetes）
具体例で覚える畳み込み計算（Conv2D、DepthwiseConv2D、SeparableConv2D、Conv2DTranspose）
混同行列(Confusion Matrix) とは 〜 2 値分類の機械学習のクラス分類について
公式
モデル
モデルについて
Functional APIのガイド
ModelクラスAPI
SequentialモデルでKerasを始めてみよう
SequentialモデルAPI