Classify the microscopy Images into Malaria and No_Malaria using Tensorflow2.X
Model: "sequential_1"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
sequential (Sequential) (None, 40, 40, 3) 0
_________________________________________________________________
rescaling (Rescaling) (None, 40, 40, 3) 0
_________________________________________________________________
conv2d (Conv2D) (None, 38, 38, 32) 896
_________________________________________________________________
max_pooling2d (MaxPooling2D) (None, 19, 19, 32) 0
_________________________________________________________________
conv2d_1 (Conv2D) (None, 17, 17, 64) 18496
_________________________________________________________________
max_pooling2d_1 (MaxPooling2 (None, 8, 8, 64) 0
_________________________________________________________________
conv2d_2 (Conv2D) (None, 6, 6, 128) 73856
_________________________________________________________________
max_pooling2d_2 (MaxPooling2 (None, 3, 3, 128) 0
_________________________________________________________________
conv2d_3 (Conv2D) (None, 1, 1, 128) 147584
_________________________________________________________________
dropout (Dropout) (None, 1, 1, 128) 0
_________________________________________________________________
flatten (Flatten) (None, 128) 0
_________________________________________________________________
dense (Dense) (None, 128) 16512
_________________________________________________________________
dense_1 (Dense) (None, 64) 8256
_________________________________________________________________
dense_2 (Dense) (None, 32) 2080
_________________________________________________________________
dense_3 (Dense) (None, 10) 330
_________________________________________________________________
dense_4 (Dense) (None, 1) 11
=================================================================
Total params: 268,021
Trainable params: 268,021
Non-trainable params: 0
_________________________________________________________________
Dataset contains two classes (Malaria, Nomalaria) with a total of 5493(Class Malaria) +7258(Class Nomalaria) RGB Images. Each image is of 40x40 resolution.
- Hyper-parameter Tuning : Hyperparameters like batch_size, learning_rate, etc can be tested with different values to result into more optimal model.
- Data Augmentation : It can be seen in the image samples that, Images having the parasites are like dark spot in center. So data augmentation techniques like transforming brightness, hue or contrast can be used to enhance the dark spot and helps in improving model performance.
- Deeper Model : While experimenting, I found that deeper network gives more better results. So more conv layers can be added. Deeper the network, more coarse features it will extract. Ofcourse a balance should be there to not overfit the model. As complex model are more prone to overfit.
- Improve training strategy : Currently full network is trained for 1 epoch and after that only head is getting trained for remaining epoch. During experiment, I found that training the full network for more number of epoch imporves the model performance. Similarly, other training strategy can be applied with different optimization technique like "AdamW".
- Try different Activation fucntion : If working with deeper model "Swish Activation Function" designed by Google can be used for challenging dataset. Swish tends to work more better than Leaky_ReLU and ReLU activation function. It has been found that simply replacing ReLUs with Swish units improves top-1 classification accuracy on ImageNet by 0.9% for Mobile NASNetA and 0.6% for Inception-ResNet-v2.
- Features Concatenation : Extracted features from different pretrained network like ResNet, VGG16 or InceptionNetV3, can be concatenated to caputure more useful details in features, that possibly be missing in one set of features. Several successful reasearch has already been done in this area, for ref - Biomedical image classification based on a feature concatenation and ensemble of deep CNNs
- Use of Advanced Deep Learning Models like RetinaNet, UNet or MaskRCNN : Advanced models like RetinaNet are build on top Feature Pyramid Network, which makes them applicable to small scale images and "Skip-Connection/Lateral Concatenation" in there network architecture makes it more robust towards classificatio and regression. These models proved to be successfull in obejct detection and localization.