About eval_seg about medical-sam-adapter HOT 3 CLOSED

it mainly considers the case of "point/click prompt". one click is commonly assigned with one class

from medical-sam-adapter.

but if it is unbalanced class in my dataset, i.e. liver and liver tumor, the random choice of class in one slice will tend to the liver, which cause a non-objective results. And in evaluation, you take [0.1, 0.3 ,0.5,0.7,0.9] for threshold, what is the insights.

from medical-sam-adapter.

but if it is unbalanced class in my dataset, i.e. liver and liver tumor, the random choice of class in one slice will tend to the liver, which cause a non-objective results. And in evaluation, you take [0.1, 0.3 ,0.5,0.7,0.9] for threshold, what is the insights.

you can easily modify the ratio of clicked classes. multi-threshold is a standard evaluation strategy, to mitigate the influence of different thresholdings in the evaluation.

from medical-sam-adapter.