# 雙能肺炎檢測

## 方法一、減少資料直到類別平衡
1. 修改成以下程式碼產生資料集的DataFrame:

data = []
k = 0
for i in range(label_df.shape[0]):
    if(label_df.iloc[i]["Pneumonia"] == 1 or label_df.iloc[i]["Pneumonia"] == 0): #原本是 ["Pneumonia"] == 1 改成 ["Nodule"] == 0
        target = os.path.join(image_folder, label_df.iloc[i]["foldername"])
        target = os.path.join(target, image_type)
        if(not os.path.exists(target)):
            continue
        filename = label_df.iloc[i]["foldername"] + ".png"
        if(label_df.iloc[i]["Pneumonia"] == 1):
            k += 1
            if(k <= 203):
                data.append((label_df.iloc[i]["foldername"] + ".png", label_df.iloc[i]["Nodule"], label_df.iloc[i]["Pneumonia"]))
        else:
            data.append((label_df.iloc[i]["foldername"] + ".png", label_df.iloc[i]["Nodule"], label_df.iloc[i]["Pneumonia"]))
        #copyfile(target, os.path.join(export_folder, filename))

2. 修改拆分訓練及與驗證集的seed (42->40)：

train_list, val_list = train_test_split(df, #測試的dataframe
                                        test_size=0.2, # 這個測試集數量可以自己設定，不一定要8 : 2，這邊只是一個範例
                                        random_state=40,
                                        shuffle = True)


## 方法二、對少量類別的資料進行資料增強 (如水平翻轉)
1. 請先上傳libs.py至環境中，libs.py已上傳至Trello中

import libs as L

2. 修改成以下程式碼產生資料集的DataFrame:

data = []
k = 0
for i in range(label_df.shape[0]):
    if(label_df.iloc[i]["Pneumonia"] == 1 or label_df.iloc[i]["Pneumonia"] == 0): #原本是 ["Pneumonia"] == 1 改成 ["Nodule"] == 0
        target = os.path.join(image_folder, label_df.iloc[i]["foldername"])
        target = os.path.join(target, image_type)
        if(not os.path.exists(target)):
            continue
        filename = label_df.iloc[i]["foldername"] + ".png"
        target = os.path.join(image_folder, label_df.iloc[i]["foldername"])
        target = os.path.join(target, image_type)
        
        jsonPath = os.path.join(mask_folder, label_df.iloc[i]["foldername"] + '.json')
        
        image = L.mask_lung_generator(target, jsonPath)
        cv.imwrite(os.path.join(export_folder, filename), image)
        data.append((filename, label_df.iloc[i]["Nodule"], label_df.iloc[i]["Pneumonia"]))
        
        if(label_df.iloc[i]["Pneumonia"] == 0):
            if(k < 191):
                image = np.flip(image,axis=1)
                filename = label_df.iloc[i]["foldername"] + "_2.png"
                cv.imwrite(os.path.join(export_folder, filename), image)
                data.append((filename, label_df.iloc[i]["Nodule"], label_df.iloc[i]["Pneumonia"]))
                k = k + 1

3. 修改拆分訓練及與驗證集的seed (43)：

train_list, val_list = train_test_split(df, #測試的dataframe
                                        test_size=0.2, # 這個測試集數量可以自己設定，不一定要8 : 2，這邊只是一個範例
                                        random_state=43,
                                        shuffle = True)

## 方法三、原圖+肺部區域圖+資料增強
1. 修改成以下程式碼產生資料集的DataFrame

data = []
k = 0
for i in range(label_df.shape[0]):
    if(label_df.iloc[i]["Pneumonia"] == 1 or label_df.iloc[i]["Pneumonia"] == 0): #原本是 ["Pneumonia"] == 1 改成 ["Nodule"] == 0
        target = os.path.join(image_folder, label_df.iloc[i]["foldername"])
        target = os.path.join(target, image_type)
        if(not os.path.exists(target)):
            continue
        filename = label_df.iloc[i]["foldername"] + ".png"
 
        jsonPath = os.path.join(mask_folder, label_df.iloc[i]["foldername"] + '.json')
        
        image = L.mask_lung_generator(target, jsonPath)
        cv.imwrite(os.path.join(export_folder, filename), image)
        data.append((filename, label_df.iloc[i]["Nodule"], label_df.iloc[i]["Pneumonia"]))
        
        filename = label_df.iloc[i]["foldername"] + "_ori.png"
        copyfile(target, os.path.join(export_folder, filename))
        data.append((filename, label_df.iloc[i]["Nodule"], label_df.iloc[i]["Pneumonia"]))
        
        if(label_df.iloc[i]["Pneumonia"] == 0):
            if(k < 191):
                image = np.flip(image,axis=1)
                filename = label_df.iloc[i]["foldername"] + "_2.png"
                cv.imwrite(os.path.join(export_folder, filename), image)
                data.append((filename, label_df.iloc[i]["Nodule"], label_df.iloc[i]["Pneumonia"]))
                
                filename = label_df.iloc[i]["foldername"] + "_ori2.png"
                image = cv.imread(target)
                image = image.astype(float)
                image = -image + np.max(image)
                image = image.astype('uint8')
                cv.imwrite(os.path.join(export_folder, filename), image)
                data.append((filename, label_df.iloc[i]["Nodule"], label_df.iloc[i]["Pneumonia"]))
                k = k + 1

2. 修改拆分訓練及與驗證集的seed (60)：

train_list, val_list = train_test_split(df, #測試的dataframe
                                        test_size=0.1, # 這個測試集數量可以自己設定，不一定要8 : 2，這邊只是一個範例
                                        random_state=60, 
                                        shuffle = True)