Learning device and medical image processing device

A learning device comprising:

• circuitry configured to
• acquire training images in which a first training image acquired by capturing light from a subject irradiated with light in a first wavelength band and a second training image acquired by capturing light from the subject irradiated with light in a second wavelength band different from the first wavelength band are paired;
• specify a singular area in the second training image;
• extract feature data of a singular-corresponding area at a pixel position in the first training image, the pixel position corresponding to the singular area in specified in the second training image;
• and
• generate a learning model by performing machine learning on the singular-corresponding area based on the feature data,
• wherein
• the feature data includes feature data related to at least one of color and luminance, and
• the circuitry is configured to make a weight of the feature data related to a blue color component lower than a weight of the feature data related to another color component when performing the machine learning on the singular-correspondence area and generating a learning model.

A medical image processing device comprising:

• circuitry configured to
• acquire a captured image acquired by capturing light from an observation target irradiated with light in a first wavelength band;
• extract feature data of each area of the captured image;
• and
• specify a singular-corresponding area in the captured image by using a learning model constructed by machine learning based on the feature data, wherein
• the learning model is generated by the machine learning on the singular-corresponding area by utilization of training images, in which a first training image acquired by capturing light from a subject irradiated with the light in the first wavelength band and a second training image acquired by capturing light from the subject irradiated with light in a second wavelength band different from the first wavelength band are paired, based on the feature data of the singular-corresponding area at a pixel position, the pixel position corresponding to a singular area in the second training image, in the first training image, wherein
• the singular area is an area in which a pixel level is equal to or larger than a specific threshold in the second training image, the singular area includes a first singular area in which the pixel level is within a first range, and
• a second singular area in which the pixel level is within a second range higher than the first range, and
• the learning model is generated by machine learning on each of a first singular-corresponding area, which is the singular-corresponding area in the first training image and is at a pixel position corresponding to the first singular area, and a second singular-corresponding area, which is the singular-corresponding area in the first training image and is at a pixel position corresponding to the second singular area, based on the feature data of the first singular-corresponding area and the second singular-corresponding area.

A medical image processing device comprising:

• circuitry configured to
• acquire a captured image acquired by capturing light from an observation target irradiated with light in a first wavelength band;
• extract feature data of each area of the captured image;
• and
• specify a singular-corresponding area in the captured image by using a learning model constructed by machine learning based on the feature data, wherein
• the learning model is generated by the machine learning on the singular-corresponding area by utilization of training images, in which a first training image acquired by capturing light from a subject irradiated with the light in the first wavelength band and a second training image acquired by capturing light from the subject irradiated with light in a second wavelength band different from the first wavelength band are paired, based on the feature data of the singular-corresponding area at a pixel position, the pixel position corresponding to a singular area in the second training image, in the first training image,
• wherein
• the feature data includes feature data related to at least one of color and luminance, and
• the learning model is generated by machine learning on the singular-corresponding area in a state in which a weight of the feature data related to a blue color component is made lower than a weight of the feature data related to another color component.

A learning method comprising:

• acquiring training images in which a first training image acquired by capturing light from a subject irradiated with light in a first wavelength band and a second training image acquired by capturing light from the subject irradiated with light in a second wavelength band different from the first wavelength band are paired;
• specifying a singular area in the second training image, wherein the singular area is an area in which a pixel level is equal to or larger than a specific threshold in the second training image, the singular area includes a first singular area in which the pixel level is within a first range each of a first singular area in which the pixel level is within a first range and a second singular area in which the pixel level is within a second range higher than the first range;
• extracting feature data of each of a first singular-corresponding area, which is the singular-corresponding area in the first training image and is at a pixel position corresponding to the first singular area, and a second singular-corresponding area, which is the singular-corresponding area in the first training image and is at a pixel position corresponding to the second singular area, specified in the second training image;
• and
• generating a learning model by performing machine learning on the first singular-corresponding area and the second singular-corresponding area based on the feature data.

A learning method comprising:

• acquiring training images in which a first training image acquired by capturing light from a subject irradiated with light in a first wavelength band and a second training image acquired by capturing light from the subject irradiated with light in a second wavelength band different from the first wavelength band are paired;
• specifying a singular area in the second training image;
• extracting feature data of a singular-corresponding area at a pixel position in the first training image, the pixel position corresponding to the singular area in specified in the second training image;
• and
• generating a learning model by performing machine learning on the singular-corresponding area based on the feature data,
• wherein
• the feature data includes feature data related to at least one of color and luminance, and
• the method further comprises setting a weight of the feature data related to a blue color component lower than a weight of the feature data related to another color component when performing the machine learning on the singular-correspondence area and generating a learning model.