CODEWITHSUNDEEP
pythontensorflowdeep-learningconv-neural-networktf.keras
Bruno Miguel Gonçalves
Bruno Miguel Gonçalves

Reputation: 165

Good performance with Accuracy but not with Dice loss in Image Segmentation

I'm doing image segmentation with U-Net like architecture on Tensorflow w/Keras but I'm new in Deep Learning.
I've got this dataset with the following set shapes:

Got some examples of these images and each channel, further down.

--> with 20% positive examples and 80% negative examples equally in each set

I run some series but for the best filter combination it made the plot for BCE with good accuracy: enter image description here

The plot for custom functions, Dice_Loss by Dice_Coeff:

enter image description here

And some images generated from the best model trained with test images:

enter image description here

The problem is when I change to dice loss and coefficient, there aren´t good predictions as we seen in the image plot and now it isn´t in the image prediction as we may see. enter image description here

Why it performs so badly in the dice loss? What other function do you recommend?

My dice loss and coefficient function:

def dice_coeff(y_true, y_pred, smooth=1):
    intersection = K.sum(K.abs(y_true * y_pred), axis=-1)
    return (2. * intersection + smooth) / (K.sum(K.square(y_true),-1) + K.sum(K.square(y_pred),-1) + smooth)


def dice_loss(y_true, y_pred):
    return 1-dice_coeff(y_true, y_pred)

Upvotes: 4

Views: 9706

Answers (1)

statmlben
statmlben

Reputation: 1

This is because most segmentation frameworks: minimizing loss functions (such as pixel-wise cross entropy) and subsequent truncation at 0.5 or argmax, are designed for pixel-wise classification accuracy.

However, the commonly-used objective of segmentation is the Dice (or IoU) metric, and these classification-based frameworks naturally cannot ensure satisfactory results based on the Dice metric. My recent paper in JMLR (https://www.jmlr.org/papers/v24/22-0712.html) has proven that for Dice-segmentation, the correct approach should be: (1) rank the estimated pixel probabilities; (2) determine the number of segmented pixels for each image through computations, see Page 6 in the paper.

Intuitively, Dice is a metric with a certain global property. It necessitates the handling of segmentation prediction on a global scale (the decision to choose a pixel for segmentation depends on its probability, while also considering the overall distribution of other pixels.), instead of simply truncating each pixel individually and independently at 0.5 like pixel-wise classification.

See more details:

Upvotes: 0

Related Questions