Human Protein Image Classification

Table of Contents

Introduction

1. Data Pre-processing

1.1 Data Loading and Preperation

1.2 Mean and Variance Normalization

1.3 Data Augmentation

2. Model Selection and construction

2.1 MobileNet

2.2 Model Construction

3. Model training and prediction

3.1 Loss function

3.2 Optimizer[5]

3.3 Performance evaluation function

3.4 Model compile

3.5 Training

3.6 Prediction

References

---

Introduction

The project is to classify mixed patterns of proteins in microscope images. More specifically, the project is to predict protein organelle localization labels for each sample. The dataset is provided by The Human Protein Atlas. There are 28 different labels present in the dataset as shown in Table 1. All image samples are represented by four filters, the protein (green) and three cellular landmarks: nucleus (blue), microtubules (red), endoplasmic reticulum (yellow). Therefore, green filters should be used to predict labels and other filters used as references[1]. 

                                           Table 1 Categories and corresponding labels

label	Category
0	Nucleoplasm
1	Nuclear membrane
2	Nucleoli
3	Nucleoli fibrillar center
4	Nuclear speckles
5	Nuclear bodies
6	Endoplasmic reticulum
7	Golgi apparatus
8	Peroxisomes
9	Endosomes
10	Lysosomes
11	Intermediate filaments
12	Actin filaments
13	Focal adhesion sites
14	Microtubles
15	Microtubule ends
16	Cytokinetic bridge
17	Mitotic spindle
18	Microtubule organizing center
19	Centrosome
20	Lipid droplets
21	Plasma membrane
22	Cell junctions
23	Mitochondria
24	Aggresome
25	Cytosol
26	Cytoplasmic bodies
27	Rods&rings

The code to get the image and four images of one sample are shown below (size of the image in each channel: (512, 512, 3)).

test_img_sample=cv2.imread("../input/human-protein-atlas-image-classification/train/7030d472-bb9a-11e8-b2b9-ac1f6b6435d0_blue.png")
plt.imshow(test_img_sample)
test_img_sample.shape

                                  Figure 1 (a) Yellow filter, (b) Blue filter , (c) Green filter, (d) Red filter

1. Data Pre-processing

In machine learning, data and features determine the upper limit of the results, model selection and optimization is to approach this upper limit. Feature engineering is used to remove impurities and redundancy from the original data, allowing the model to fit quickly. Unlike traditional machine learning, there's no need to artificially synthesize advanced complex features in deep learning. It only needs human's prior knowledge to process first-order features, and then deep learning will learn relevant complex features by itself.  

1.1 Data Loading and Preperation

The images and their corresponding targets are stored in the following format:

path_to_train = '../input/human-protein-atlas-image-classification/train'
data = pd.read_csv('../input/human-protein-atlas-image-classification/train.csv')
print(data)

Convert the data into array using the following code:

train_dataset_info = []
for name, la