A disease diagnosis system includes a processor and a storage device storing a neural network. The processor trains the neural network in the storage device to output a determination value corresponding to a probability having at least one of a plurality of states using a given loss function and learning data labeled so that a given unitary unit included in a biometric image is to have at least one of the plurality of states. The neural network includes a specific layer to output a plurality of feature values corresponding to a probability that the unitary unit is to be determined as each of the plurality of states. The loss function incorporates both first and second feature values corresponding to first and second states into a dual labeling unitary unit with the first state having a higher probability and a second state having lower probability.