Page 73 - Big Data Analytics for Intelligent Healthcare Management

66 CHAPTER 4 TRANSFER LEARNING AND SUPERVISED CLASSIFIER

4.5 IMPLEMENTATION
System Description: Proposed model is implemented on a system with 8GB RAM core i7 processor.
Tools: Python programming language with python image processing package scikit-mage [39], ma-
chine learning package scikit-learn [40], and for ConvNets, a deep learning framework, Keras, is used.

4.5.1 FEATURE EXTRACTION
I. Every image is resized by the required maximum input size for the four ConvNet models. After that,
image pixels are rescaled to [ 1, +1] (Table 4.2).
II. Then the resized and rescaled images are passed through the ConvNets Models and 1D feature
vector is collected (Table 4.3).
All the ConvNets models are available in keras deep learning framework.

4.5.2 DIMENSIONALITY REDUCTION
PCA is applied on the feature vectors to reduce feature dimension. Before applying PCA, the features
are standardized. PCA was trained only on a training set and projection to lower dimensional space was
applied to both the training and test set. PCA is applied with an explained variance ratio of 0.95. PCA is
available in the scikit-learn package (Table 4.4).

Table 4.2 Input Image Size Required for Four ConvNet
Models
Feature Extractor Input Size
ResNet50 224 224
InceptionV3 299 299
Inception-ResNet-v2 299 299
Xception 299 299

Table 4.3 Feature Dimension by Four Feature Extractors
Feature Extractor Feature Dimension
ResNet50 2048
InceptionV3 2048
Inception ResnetV2 1536
Xception 2048