Example: Image Segmentation (Camvid Dataset)

The dataset to perform imgage segmentation can be downloaded from here

Import libraries

import os
import numpy as np
import medicalai as ai
import tensorflow as tf

Define the hyperparameters

Specify the dataset folder which further contains test & train folders each with n class object folders

datasetFolderPath = "../data/camvid/"

Specify the dimensions of image to be fed to network

(IMG_HEIGHT,IMG_WIDTH) = (256,256)

Specify the name of the model to be trained on

EXPT_NAME = '1'
AI_NAME = 'unet' 
MODEL_SAVE_NAME = '../model/'+AI_NAME+'/Medical_RSNA_'+str(IMG_HEIGHT)+'x'+str(IMG_WIDTH)+'_'+AI_NAME+'_EXPT_'+str(EXPT_NAME)

Specify remaining hyperparamters

batch_size = 32
epochs = 10
learning_rate = 0.0001

Define the augmentation for the generator

(The following augmentation is for image only)

augment = ai.AUGMENTATION(rescale= 1./255)

Load your data from folder using datasetGenFromFolder if your data is in folder structured form. Make sure flag_multi_class is set to True

dsHandler = ai.segmentaionGenerator(folder=datasetFolderPath,targetDim=(IMG_HEIGHT,IMG_WIDTH), 
                                        augmentation=augment, class_mode=None,
                                        batch_size = batch_size,
                                        image_folder_name = "image", mask_folder_name = "masks",
                                        flag_multi_class=True)

trainGen = dsHandler.load_train_generator()
testGen = dsHandler.load_test_generator()

Train model

Now our image generator is ready to be trained on our model. But first we need to define a tensorflow callback for the model

model_checkpoint = tf.keras.callbacks.ModelCheckpoint(
                                        MODEL_SAVE_NAME+'best.h5', 
                                        verbose=0,
                                        mode='auto', 
                                        save_freq=5,
                                        save_best_only=True,
                                        )
callbacks = [model_checkpoint]

TRAIN_STEPS = int(np.ceil(dsHandler.imageGen.generator.n/dsHandler.batch_size))
trainer = ai.TRAIN_ENGINE()
trainer.train_and_save_segmentation(AI_NAME=AI_NAME,
                                    MODEL_SAVE_NAME = MODEL_SAVE_NAME, 
                                    trainSet=trainGen,inputSize = (IMG_HEIGHT,IMG_WIDTH,3),
                                    TRAIN_STEPS=TRAIN_STEPS,
                                    BATCH_SIZE= BATCH_SIZE, EPOCHS= EPOCHS, 
                                    LEARNING_RATE= LEARNING_RATE, SAVE_BEST_MODEL = SAVE_BEST_MODEL, 
                                    callbacks = callbacks,
                                    showModel = False)

Use the above model to predict segmentation masks

infEngine = ai.INFERENCE_ENGINE(MODEL_SAVE_NAME)
predsG = infEngine.predict_segmentation(testGen)

Save the predicted segmentation masks in a folder

infEngine.saveResult(save_path='results',npyfile=predsG,num_class=32,flag_multi_class=True)