This is a competition in which we have to identify and segment functional tissue units (FTUs) across five human organs. We've to build our model using a dataset of tissue section images, with the best submissions segmenting FTUs as accurately as possible.
\nThis competition is evaluated on the mean Dice coefficient. This notebook tries to explain the Dice Coefficient and how it differs from another Image Segmentation score Intersection-Over-Union.
\n Dice Coefficient or Dice similarity coefficient (DSC) is also known by the name Sørensen–Dice index. It is a statistical tool which measures the similarity between two sets of data.\n
\n\nIt was independently developed by the **botanists**:\n
\n Hence, came the name Dice Coefficient or Sørensen–Dice index.\n
\n\n","metadata":{}},{"cell_type":"markdown","source":"\n The equation or formula for the Dice Coefficient is given as:\n
\n\n
\n where:\n
\n
\n
\n Also, when the same thing is applied to Boolean data, the Dice Coefficient can be written as:\n
\n\n\n\n
\n where:\n
\n
\n We can create our Segmentation model and train it, but to know about its accuracy, we need some kind of metric.
\n There are many but the essential metrics that we see often are:\n
\n
\n Now, look at the below diagram.\n
\n\n\n
\n The Dice Coefficient is simply the 2 * the area of the overlap divided by the no. of pixels in both the images.
\n Now, we may think wasn't that the same as Intersection-Over-Union as it follows the same structure:\n
\n Also, if we look at their boolean implementation formula:\n
\n\n- **Dice Coefficient**\n\n\n\n- **IoU**\n\n\n\n\n The ratio between them can be related explicitly to the IoU like:\n
\n\n\n","metadata":{}},{"cell_type":"markdown","source":"\n They may be same in the normal usage but the problem comes when taking the average or mean score over a set of inferences.\n
\n\n\n\n Generally:\n
\n\n#### IoU\n\n\n
\nHence, Dice Coefficient tends to measure something closer to average performance while the IoU score measures something closer to the worst case performance.\n
\n\n In conclusion, the most commonly used metrics for semantic segmentation are the IoU and the Dice Coefficient.\n In keras, to calculate the Dice Coefficient, we can do the following:\n
\n\n```python\nfrom keras import backend as K\n\ndef dice_coeff(y_true, y_pred, smooth=1):\n y_true_flatten = K.flatten(y_true)\n y_pred_flatten = K.flatten(y_pred)\n intersection = K.sum(y_true_flatten * y_pred_flatten)\n dice = (2. * intersection + smooth) / (K.sum(y_true_flatten) + K.sum(y_pred_flatten) + smooth)\n return dice\n```\n","metadata":{}},{"cell_type":"markdown","source":"