Applying Deep Learning to Histology Images

The kidney

Filter the blood
Process starts with the glomeruli
Also the site of many kidney diseases

Glomeruli are undercounted

Pathologists consistently miss 50% of glomeruli

Simple classification model: performance baseline

Positive
Negative

Simple classification model: performance baseline

Segmentation

Is this pixel part of a glomerulus?

Fully convolutional neural networks appear useful

Segmentation

Maxpooling to expand the perceptive field

Concatenate higher resolution features to refine boundary

U-net model


def unet_generator(input_shape=(1024,1024,3), num_classes=1, levels=7):
    inputs = Input(shape=input_shape)
    
    upsampling = []
    connection = inputs
    for i in range(levels):
        down = Conv2D(8*2**i, (3, 3), padding='same')(connection)
        down = BatchNormalization()(down)
        down = Activation('relu')(down)
        down = Conv2D(8*2**i, (3, 3), padding='same')(down)
        down = BatchNormalization()(down)
        down = Activation('relu')(down)
        upsampling.append(down)
        connection = MaxPooling2D((2, 2), strides=(2, 2))(down)
        
    center = Conv2D(8*2**levels, (3, 3), padding='same')(connection)
    center = BatchNormalization()(center)
    center = Activation('relu')(center)
    center = Conv2D(8*2**levels, (3, 3), padding='same')(center)
    center = BatchNormalization()(center)
    connection = Activation('relu')(center)
        
    for i in range(levels-1, -1, -1):
        up = UpSampling2D((2, 2))(connection)
        up = concatenate([upsampling.pop(), up], axis=1)
        up = Conv2D(8*2**i, (3, 3), padding='same')(up)
        up = BatchNormalization()(up)
        up = Activation('relu')(up)
        up = Conv2D(8*2**i, (3, 3), padding='same')(up)
        up = BatchNormalization()(up)
        up = Activation('relu')(up)
        up = Conv2D(8*2**i, (3, 3), padding='same')(up)
        up = BatchNormalization()(up)
        connection = Activation('relu')(up)
        
    classify = Conv2D(num_classes, (1, 1), activation='sigmoid')(connection)

    model = Model(inputs=inputs, outputs=classify)

    model.compile(optimizer=RMSprop(lr=1e-3), 
                    loss=bce_dice_loss, 
                    metrics=[dice_coeff])

    return model