Therefore, the first CNN 31 is trained using not only the data indicating a correct infarction region but also the task of estimating the MR estimated image. That is, not only the knowledge of infarction but also the knowledge of estimating the MR estimated image is reflected in the discrimination result of the infarction region output from the second CNN 32. As a result, the accuracy of the discrimination result is improved. Therefore, in this embodiment, it is possible to discriminate a disease region with high accuracy using a limited amount of data even in an image in which it is difficult to prepare a large amount of data indicating a correct disease region.

In the above-described embodiment, two CNNs 32 and 33 are provided as a plurality of learning units according to the present disclosure. However, the technology according to the present disclosure is not limited thereto and two or more CNNs may be provided. For example, a learning unit that can be used to extract an infarction region from a CT image may be further provided in order to extract an infarction region from, for example, a PET image, an ultrasound image, a T1 image, and a T2 image.

In the above-described embodiment, the disease is infarction. However, the technology according to the present disclosure is not limited thereto. For example, the disease may be thrombus or bleeding.