Fig. 3
From: A weakly-supervised follicle segmentation method in ultrasound images
MIL loss function. It enforces the tightness constraints of the bounding boxes on the prediction map. It encourages the model to predict a high score for at least one pixel in a positive bag (a region within a bounding box that should contain the object) and a low score for all pixels in a negative bag (a region outside any bounding box that should not contain the object). \(b_i\) represents the set of mask probabilities of the pixel instances belonging to bag i, \(y_i\) = 1 if bag i is positive, and \(y_i\) = 0 otherwise.
