The Gold Answer is not the output of the model but the expected answer in the benchmark. Probably the benachmark contains multiple consecutive images of the same patient.