TWI881698B - Lesion identification learning device and method - Google Patents

Abstract

A lesion identification learning device is disclosed, which includes a memory and a processor. The processor performs following steps: (a) selecting one of multiple candidate lesion images corresponding to a first lesion type as a question image; (b) determining whether an answer type is the same as the first lesion type; (c) when the answer type is different from the first lesion type, selecting multiple reference images similar to the question image from the multiple candidate lesion images corresponding to the first lesion type; (d) selecting the candidate lesion type that is the same to the answer type as a second lesion type, and selecting multiple similar images that are similar to the question image from the multiple candidate lesion images corresponding to the second lesion type; and (e) displaying the multiple reference images and the multiple similar images on an output interface for a user to learn lesion identification.

Description

Lesion identification learning device and method

This disclosure relates to a technology for medical applications, and in particular to a lesion identification learning device and method.

In clinical diagnosis, ultrasound, computed tomography (CT) and magnetic resonance imaging (MRI) are usually used to detect changes in lesions. Generally speaking, the evaluation of physical lesions is done by observing or comparing lesion images manually. However, in the face of such delicate lesion identification in the human body, those who aspire to become medical workers usually need to spend considerable time and energy to gain a certain degree of understanding. The entry threshold is actually very high. Of course, some people give up because the threshold is too high, or their enthusiasm gradually fades during the learning process, leading to them giving up halfway. Therefore, how to enable medical workers to learn lesion identification more quickly and efficiently to avoid the loss of potential medical talents is a very important topic in this field.

The main purpose of this disclosure is to provide a lesion identification learning device and method, which can enable medical workers to learn lesion identification more quickly and efficiently, so as to avoid the loss of potential medical talents.

In order to achieve the above-mentioned purpose, the lesion recognition learning device disclosed in the present invention comprises: a memory configured to store a plurality of instructions and a plurality of candidate lesion images respectively corresponding to each of a plurality of candidate lesion types; and a processor connected to the memory and configured to access the plurality of instructions to execute the following steps: (a) selecting one of the plurality of candidate lesion types as a first lesion type, and selecting one of the plurality of candidate lesion images corresponding to the first lesion type as a question image; (b) receiving an answer type, and determining whether the answer type is the same as the first lesion type; (c) receiving a question type, and determining whether the answer type is the same as the first lesion type; c) when the answer type is different from the first lesion type, selecting multiple reference images similar to the problem image from the multiple candidate lesion images corresponding to the first lesion type; (d) selecting the candidate lesion type that is the same as the answer type as a second lesion type, and selecting multiple similar images similar to the problem image from the multiple candidate lesion images corresponding to the second lesion type; and (e) generating the multiple reference images corresponding to the first lesion type and the multiple similar images corresponding to the second lesion type on an output interface for a user to learn lesion recognition.

In order to achieve the above-mentioned purpose, the lesion recognition learning method disclosed in the present invention is applicable to the recognition learning of multiple candidate lesion images corresponding to each of multiple candidate lesion types, including: (a) selecting one of the multiple candidate lesion types as a first lesion type, and selecting one of the multiple candidate lesion images corresponding to the first lesion type as a question image; (b) receiving an answer type, and determining whether the answer type is the same as the first lesion type; (c) when the answer type is the same as the first lesion type, At the same time, multiple reference images similar to the question image are selected from the multiple candidate lesion images corresponding to the first lesion type; (d) the candidate lesion type that is the same as the answer type is selected as a second lesion type, and multiple similar images that are similar to the multiple reference images are selected from the multiple candidate lesion images corresponding to the second lesion type; and (e) the multiple reference images corresponding to the first lesion type and the multiple similar images corresponding to the second lesion type are displayed on an output interface.

In order to achieve the above-mentioned purpose, the lesion recognition learning method disclosed in the present invention is applicable to the recognition learning of multiple candidate lesion images corresponding to each of multiple candidate lesion types, including: (a) selecting one of the multiple candidate lesion types as a first lesion type by a processor, and selecting one of the multiple candidate lesion images corresponding to the first lesion type as a question image; (b) receiving an answer type by the processor, and determining whether the answer type is the same as the first lesion type; (c) when the answer type is the same as the first lesion type, the processor receives an answer type and determines whether the answer type is the same as the first lesion type; and (d) when the answer type is the same as the first lesion type, the processor receives an answer type and determines whether the answer type is the same as the first lesion type. When the lesion types are different, multiple reference images similar to the problem image are selected from the multiple candidate lesion images corresponding to the first lesion type; (d) by the processor, the candidate lesion type of the same type as the answer is selected as a second lesion type, and multiple similar images similar to the problem image are selected from the multiple candidate lesion images corresponding to the second lesion type; and (e) by the processor, the multiple reference images corresponding to the first lesion type and the multiple similar images corresponding to the second lesion type are displayed on an output interface.

Compared with the related art, the present invention displays images of the correct lesion category similar to the problem image and images of the wrong lesion category similar to the problem image when the user answers the wrong lesion category, thereby achieving the technical effect of allowing the user to learn the characteristics of the correct lesion category and the characteristics of the wrong lesion category at the same time to improve the lesion identification ability. In this way, medical workers will be able to learn lesion identification more quickly and efficiently, and solve the problem of potential medical talent loss.

100: Lesion identification learning device

110: Memory

120: Processor

130: Display

140: Input circuit

tp1~tpn: Candidate lesion type

img11~imgnz: candidate lesion images

S210~S250: Steps

300: Graphical interface

310, 610: Question field

320~340: Input field

350: Drop-down menu

360:Answer field

370~380: Option pattern

600: Output interface

620~640: Reference field

650~670: Similar fields

FR: Feature area

FIG1 shows a block diagram of a lesion recognition learning device in some embodiments.

FIG2 is a flowchart of a lesion identification learning method in some embodiments.

FIG3 is a schematic diagram of a graphical interface in some embodiments.

FIG4 is a schematic diagram of a drop-down menu of a graphical interface in some embodiments.

FIG5 is a schematic diagram of an answer field of a graphical interface in some embodiments.

FIG6 is a schematic diagram of an output interface in some embodiments.

FIG7 is a schematic diagram showing multiple key feature points in some embodiments.

Referring to FIG. 1 , FIG. 1 shows a block diagram of a lesion identification learning device 100 in some embodiments, wherein the lesion identification learning device 100 can be implemented by any electronic device or server, etc. (for example, it can be a terminal processing device (i.e., a mobile phone, a desktop computer or a tablet computer, etc.), a cloud device, a server or a cloud server, etc.). As shown in FIG. 1 , the lesion identification learning device 100 includes a memory 110 and a processor 120. The processor 120 is connected to the memory 110.

In the present embodiment, the memory 110 stores a plurality of instructions and a plurality of candidate lesion images corresponding to each of the plurality of candidate lesion types tp1 to tpn. Specifically, the candidate lesion type tp1 corresponds to a plurality of candidate lesion images img11 to img1m. The candidate lesion type tp2 corresponds to a plurality of candidate lesion images img21 to img2o. Similarly, each of the other candidate lesion types tp3 to tpn also corresponds to a plurality of candidate lesion images. It is worth noting that m to z are any positive integers and are not particularly limited. In some embodiments, the memory 110 further stores a classification model (not shown). In some embodiments, the processor 120 uses a classification model to classify all candidate lesion images to classify all candidate lesion images into candidate lesion types tp1~tpn. In some embodiments, the classification model can be a model of any deep learning algorithm (for example, a convolutional neural network model, a transformer model, or a YOLO (you only look once) model, etc.).

In some implementations, the candidate lesion types tp1-tpn may be any lesion type of a body tissue (e.g., multiple drusen or diabetic macular edema of the eye). In some implementations, the candidate lesion images may be any images of body tissues with lesions (e.g., ultrasound images of eyes with optic edema, computed tomography (CT) images of eyes with multiple drusen, or magnetic resonance imaging (MRI) images of eyes with diabetic macular edema). In some implementations, the memory 110 may be implemented by a memory unit, a flash memory, a read-only memory, a hard disk, or any equivalent storage component, but is not limited thereto. In some implementations, the above-mentioned multiple instructions may be corresponding software or firmware instruction programs.

In this embodiment, the processor 120 executes the detailed steps of the lesion recognition learning method in the subsequent paragraphs based on the above instructions. In some embodiments, the processor 120 can be implemented by a central processing unit (CPU), a micro control unit (MCU), a programmable logic controller (PLC), a system on chip (SoC) or a field programmable gate array (FPGA), etc., but is not limited thereto.

In some embodiments, the lesion identification learning device 100 further includes a display 130, which displays the output interface of the subsequent paragraphs, so that the user can watch and learn lesion identification. In some embodiments, the display 130 can be implemented by any type of display (for example, a liquid crystal display, a touch display, or a head-mounted display, etc.). In some embodiments, the output interface can be any visual graphical interface. In some embodiments, the lesion identification learning device 100 further includes an input circuit 140, which allows the user to input the answer type of the subsequent paragraph on the graphical interface. In some embodiments, in some embodiments, the graphical interface can also be displayed on the display 130. In some embodiments, the graphical interface may also be any visual graphical interface. In some embodiments, the input circuit 140 may be implemented by any input circuit (e.g., a keyboard, a mouse, or a touch panel, etc.). In some embodiments, the input circuit 140 may be implemented by a display 130 (e.g., a touch display).

The following further describes the lesion identification learning method disclosed herein. Referring to FIG. 2 , FIG. 2 shows a flow chart of the lesion identification learning method in some embodiments. The lesion identification learning method is applicable to the lesion identification learning device 100 shown in FIG. 1 .

As shown in FIG2 , the lesion recognition learning method includes steps S210 to S250. First, in step S210, the processor 120 selects one of a plurality of candidate lesion types tp1 to tpn as the first lesion type, and selects one of a plurality of candidate lesion images corresponding to the first lesion type as the problem image.

In some embodiments, the input circuit 140 receives the image type and the organ type to transmit to the processor 120. Then, the processor 120 randomly selects one of the candidate lesion types associated with the organ type as the first lesion type, and randomly selects one of the candidate lesion images corresponding to the image type from a plurality of candidate lesion images corresponding to the first lesion type as the problem image. In some embodiments, the image type can be the type of image generated by any imaging method (e.g., ultrasound image, computer tomography image, or magnetic resonance imaging, etc.). In some embodiments, the organ type can be any type of human organ (e.g., eye, ear, stomach, or kidney, etc.). In some embodiments, the graphical interface can receive the image type and organ type input by the user using the input circuit 140, so as to display the image type and organ type respectively on the input imaging type field and the input organ field in the graphical interface. In some embodiments, the graphical interface further has an input disease field for the user to input the answer type of the subsequent paragraph. In some embodiments, the graphical interface further has a question field to display the question image for the user to view.

The graphical interface is described below with an actual example. Referring to FIG. 3 , FIG. 3 is a schematic diagram of a graphical interface 300 in some embodiments. As shown in FIG. 3 , the graphical interface 300 has input fields 320 to 340. The input field 320 is used to receive an image type (i.e., optical coherence tomography) input by a user. The input field 330 is used to receive an organ type (i.e., eye) input by a user. The processor 120 randomly selects one of the candidate lesion types corresponding to the eye (i.e., diabetic macular edema) as the first lesion type, and randomly selects one of the candidate lesion images corresponding to the optical coherence tomography as the problem image from a plurality of candidate lesion images corresponding to diabetic macular edema. It is worth mentioning that the processor 120 does not display the randomly selected first lesion type, so the user cannot know the first lesion type. The graphical interface 300 further has a question field 310, which is used to display the question image selected by the processor 120 based on the input fields 320 and 330 (i.e., any candidate lesion image of the optical coherence tomography scan of diabetic macular edema). After the user views the question image displayed on the question field 310, he can enter the answer type of the subsequent paragraph in the input field 340 (i.e., view the image of the lesion in the question and identify which lesion exists in the image). In this way, the effect of training the ability to identify lesions is achieved.

Return to Fig. 2, in step S220, processor 120 receives answer type, and judges whether answer type is identical with the first lesion type.In some embodiments, answer type is one of candidate lesion types relevant to the organ type of input.In some embodiments, processor 120 receives one of candidate lesion types relevant to the organ type of input from a graphical interface.In some embodiments, processor 120 generates a drop-down menu when receiving a click instruction to an input field relevant to the disease of the graphical interface from input circuit 140.Then, processor 120 uses the candidate lesion type clicked as answer type when receiving a click instruction to one of candidate lesion types relevant to the organ type in the drop-down menu from input circuit 140. In other embodiments, the processor 120 may also display the input answer type on the graphical interface when receiving a direct input of the answer type from the input circuit 140.

The following is an actual example to illustrate the graphical interface. Referring to FIG. 4 , FIG. 4 is a schematic diagram of a drop-down menu 350 of the graphical interface 300 in some embodiments. As shown in FIG. 3 and FIG. 4 , continuing the example of FIG. 3 , when the user clicks the input field 340 on the graphical interface 300 using the input circuit 140, the processor 120 can generate a drop-down menu 350 on the graphical interface 300, and the drop-down menu 350 includes candidate lesion types related to the eye (for example, multiple drusen related to the eye, diabetic macular edema or papillary edema, etc.). Next, the user can view the question image displayed in the question field 310 and use the input circuit 140 to select one of the eye-related multiple drusen, diabetic macular edema, or optic disc edema as the answer type (i.e., the user can select which candidate lesion type the question image should be).

Returning to FIG. 2 , in step S230, when the answer type is different from the first lesion type, the processor 120 selects a plurality of reference images similar to the question image from a plurality of candidate lesion images corresponding to the first lesion type. In some embodiments, the processor 120 uses a similarity algorithm to select a plurality of candidate lesion images similar to the question image from a plurality of candidate lesion images corresponding to the first lesion type as a plurality of reference images. In some embodiments, the processor 120 uses a similarity algorithm to identify a specific number of candidate lesion images with the highest similarity to the question image from a plurality of candidate lesion images corresponding to the first lesion type as a plurality of reference images. In some embodiments, the similarity algorithm may be any algorithm for calculating similarity (e.g., Euclidean distance algorithm, Chebyshev distance algorithm, or principal component analysis algorithm, etc.). In some embodiments, the specific number may be any number pre-set by the user. In some embodiments, when the answer type is different from the first lesion type, the processor 120 generates an answer field on the graphical interface and displays the first lesion type in the answer field.

The answer field is explained below using an actual example. Referring to FIG. 5 , FIG. 5 is a schematic diagram of an answer field 360 of a graphical interface 300 in some embodiments. As shown in FIG. 5 , when a user inputs an answer type (i.e., multiple drusen) on the graphical interface 300 using the input circuit 140, the processor 120 receives the answer type from the graphical interface 300 and compares the answer type with the first lesion type (i.e., diabetic macular edema) to see if they are the same. At this time, the processor 120 determines that the input answer type is not diabetic macular edema. Then, the processor 120 generates an answer field 360 on the graphical interface 300 and displays diabetic macular edema in the answer field 360 (i.e., displays the correct answer). In addition, the graphical interface 300 further has option icons 370~380. Option icon 370 is used for the user to select to generate the output interface of the subsequent paragraph. The subsequent paragraph will further explain the output interface, so it is not repeated here. Option icon 380 is used for the user to select to generate the graphical interface of the next question. The graphical interface of the next question is similar to the above graphical interface, so it is not repeated here.

In some embodiments, when the answer type is the same as the first lesion type, the processor 120 uses a similarity algorithm to select a candidate lesion image that is most similar to the question image from multiple candidate lesion images corresponding to each of the candidate lesion types different from the first lesion type as a new question image. Then, the processor 120 uses the candidate lesion type corresponding to the new question image as the new first lesion type. Then, the processor 120 executes step S220 again.

Returning to FIG. 2 , in step S240, the processor 120 selects a candidate lesion type that is the same as the answer type as the second lesion type, and selects a plurality of similar images that are similar to the problem image from a plurality of candidate lesion images corresponding to the second lesion type. In some embodiments, the processor 120 uses a similarity algorithm to identify a specific number of candidate lesion images that have the highest similarity to the problem image from a plurality of candidate lesion images corresponding to the second lesion type as a plurality of similar images. In other words, the processor 120 uses a similarity algorithm to find out a few candidate lesion images that are most similar to the problem image from a plurality of candidate lesion images corresponding to the second lesion type as these similar images (i.e., find the candidate lesion images with the highest similarity to the problem image). In some embodiments, this specific number may also be an arbitrary number pre-set by the user. In some embodiments, the number of similar images is equal to the number of reference images.

In step S250, the processor 120 generates a plurality of reference images corresponding to the first lesion type and a plurality of similar images corresponding to the second lesion type on the output interface for the user to learn lesion recognition. In some embodiments, the processor 120 uses a classification model to identify a plurality of key feature points related to the lesion in the problem image (for example, the coordinates of a plurality of key feature points corresponding to the lesion are extracted from a plurality of convolutional layers of a convolutional neural network model). In some embodiments, the processor 120 further generates a problem image on the output interface, and generates corresponding multiple colors at the positions of multiple key feature points related to the lesion in the problem image (for example, the key feature points with higher lesion tissue density have colors closer to red, and the key feature points with lower lesion tissue density have colors closer to green). In some embodiments, the processor 120 displays the output interface on the display 130 for the user to learn lesion identification.

The output interface is explained below with an actual example. Referring to FIG. 6 , FIG. 6 is a schematic diagram of an output interface 600 in some embodiments. As shown in FIG. 6 , the output interface 600 has a question field 610. The question field 610 is used to display the above-mentioned question image (i.e., one of the candidate lesion images of diabetic macular edema), and generate a feature region FR at the positions of multiple key feature points related to the lesion on the question image, wherein the feature region FR will display colors corresponding to the multiple key feature points, wherein the key feature points with higher lesion tissue density in the feature region FR have colors closer to red at their positions, and the key feature points with lower lesion tissue density have colors closer to green at their positions. With reference to the diagram of FIG. 7 , FIG. 7 illustrates a schematic diagram of a plurality of key feature points in some embodiments. In the diagram of FIG. 7 , the key feature points with higher lesion tissue density have a color closer to the first color (e.g., red), while the key feature points with lower lesion tissue density have a color closer to the second color (green). In other words, the problem field 610 shows a colored problem image after feature enhancement of the problem image of the problem field 310 on the graphical interface 300 of FIG. 5 . As shown in FIG. 6 , the output interface 600 further has reference fields 620 to 640. Reference fields 620-640 are used to display the three reference images most similar to the problem image (i.e., the top three diabetic macular edema images with the highest similarity to the problem image), wherein the reference image displayed in reference field 620 has the highest similarity to the problem image (i.e., the one diabetic macular edema image most similar to the problem image). In other words, the diabetic macular edema image displayed in reference field 620 has the highest similarity to the problem image.

The output interface 600 also has similar fields 650-670. The similar fields 650-670 are respectively used to display similar images that are most similar to the problem image of the problem field 310 on the graphical interface 300 of FIG. 5 (ie, the image of multiple drusen that is most similar to the problem image of diabetic macular edema). In other words, the image of multiple drusen displayed in the similar field 650 has the highest similarity with the image of diabetic macular edema displayed in the problem field 310 (i.e., the top three images of multiple drusen with the highest similarity to the problem image displayed in the problem field 310 of FIG. 5 ), wherein the similar image displayed in the similar field 650 has the highest similarity with the problem image displayed in the problem field 310 of FIG. 5 (i.e., the image of multiple drusen that is most similar to the problem image). In other words, the image of multiple drusen displayed in the similar field 650 has the highest similarity with the problem image displayed in the problem field 310 of FIG. 5 . The images of multiple drusen displayed in the similarity fields 660-670 respectively have the second and third highest similarities with the problem image displayed in the problem field 310 of FIG. 5 .

Through the above steps, the user can view the location of the lesion characteristics in the problem image on the output interface displayed by the display 130, and can also view the reference image similar to the problem image (same lesion type) and the similar image that is difficult to distinguish from the problem image (different lesion type) on the output interface displayed by the display 130. The user can directly observe the difference between the reference image and the similar image of different lesion types that is most similar to the problem image. In this way, the user can not only learn the characteristics of the image of the lesion of the first lesion type, but also learn the characteristics of the image of the lesion of the second lesion type, so that the user can more accurately and quickly identify the lesion and lesion type of the image. In this way, such a method will overcome the problem of difficulty in identifying lesions due to too many image types in the past.

In summary, the lesion identification learning device and method proposed in the present disclosure first set the lesion type of the question and the image of the question corresponding to the lesion type, and display the image on the display for the user to answer. When receiving an incorrect answer, the lesion identification learning device and method proposed in the present disclosure will find an image similar to the image of the question, wherein this image also corresponds to the lesion type of the question. The lesion identification learning device and method proposed in the present disclosure will further find images similar to the image of the question in other lesion types. Thereby, the user can observe the difference and characteristics of two lesion types in paired similar images on the display. In addition, the lesion identification learning device and method proposed in the present disclosure will further display the key feature points in the image of the question for the user to observe various characteristics of the lesion type of the question. In this way, the lesion identification learning device and method proposed in this disclosure can enable medical workers to learn lesion identification more quickly and efficiently, and solve the problem of potential medical talent loss.

Although the present disclosure has been disclosed as above by way of embodiments, it is not intended to limit the present disclosure. Any person with ordinary knowledge in the relevant technical field may make some changes and modifications within the spirit and scope of the present disclosure. Therefore, the protection scope of the present disclosure shall be subject to the scope defined by the attached patent application.

100: Lesion identification learning device

110: Memory

120: Processor

130: Display

140: Input circuit

tp1~tpn: Candidate lesion type

img11~imgnz: candidate lesion images

Claims (12)

A lesion recognition learning device comprises: a memory configured to store a plurality of instructions and a plurality of candidate lesion images respectively corresponding to each of a plurality of candidate lesion types; and a processor connected to the memory and configured to access the plurality of instructions to execute the following steps: (a) selecting one of the plurality of candidate lesion types as a first lesion type, and selecting one of the plurality of candidate lesion images corresponding to the first lesion type as a question image; (b) receiving an answer type, and determining whether the answer type is the same as the first lesion type; (c) when the answer type is When the first lesion type is different, multiple reference images similar to the question image are selected from the multiple candidate lesion images corresponding to the first lesion type using a similarity algorithm; (d) the candidate lesion type of the same type as the answer type is selected as a second lesion type using the similarity algorithm, and multiple similar images similar to the question image are selected from the multiple candidate lesion images corresponding to the second lesion type; and (e) the multiple reference images corresponding to the first lesion type and the multiple similar images corresponding to the second lesion type are displayed on an output interface. The lesion recognition learning device as described in claim 1, wherein the processor is further configured to perform the following steps: when the answer type is the same as the first lesion type, using a similarity algorithm, from the multiple candidate lesion images corresponding to each of the multiple candidate lesion types different from the first lesion type, select the candidate lesion image that is most similar to the problem image as a new problem image; and Use the candidate lesion type corresponding to the new problem image as a new first lesion type, and re-execute step (b). The lesion recognition learning device as described in claim 1 further includes an input circuit connected to the processor and configured to receive an image type and an organ type, wherein in step (a), the processor is further configured to perform the following steps: randomly selecting one of the multiple candidate lesion types associated with the organ type as the first lesion type, and randomly selecting one of the multiple candidate lesion images corresponding to the first lesion type that corresponds to the image type as the problem image. The lesion recognition learning device as described in claim 1, wherein in step (c), the processor is further configured to perform the following steps: using the similarity algorithm, identifying a specific number of candidate lesion images with the highest similarity to the problem image from the multiple candidate lesion images corresponding to the first lesion type as the multiple reference images. The lesion recognition learning device as described in claim 1, wherein in step (d), the processor is further configured to perform the following steps: using the similarity algorithm, identifying a specific number of the candidate lesion images with the highest similarity to the problem image from the multiple candidate lesion images corresponding to the second lesion type as the multiple similar images. The lesion recognition learning device as described in claim 1 further includes a display connected to the processor and configured to display the output interface, wherein the processor is further configured to perform the following steps: displaying the multiple reference images and the multiple similar images on the output interface; and displaying the problem image on the output interface, and generating multiple colors at a position of multiple key feature points related to a lesion in the problem image. A lesion recognition learning method is applicable to recognition learning of multiple candidate lesion images corresponding to each of multiple candidate lesion types, comprising: (a) selecting one of the multiple candidate lesion types as a first lesion type by a processor, and selecting one of the multiple candidate lesion images corresponding to the first lesion type as a question image; (b) receiving an answer type by the processor, and determining whether the answer type is the same as the first lesion type; (c) using the processor to identify the answer type by using a similarity model when the answer type is different from the first lesion type. The algorithm selects multiple reference images similar to the problem image from the multiple candidate lesion images corresponding to the first lesion type; (d) the processor selects the candidate lesion type of the same type as the answer type as a second lesion type using the similarity algorithm, and selects multiple similar images similar to the problem image from the multiple candidate lesion images corresponding to the second lesion type; and (e) the processor displays the multiple reference images corresponding to the first lesion type and the multiple similar images corresponding to the second lesion type on an output interface. The lesion recognition learning method as described in claim 7 further includes: When the answer type is the same as the first lesion type, the processor uses a similarity algorithm to select the candidate lesion image that is most similar to the problem image from the multiple candidate lesion images corresponding to each of the multiple candidate lesion types different from the first lesion type as a new problem image; and the processor uses the candidate lesion type corresponding to the new problem image as a new first lesion type, and returns to step (a). The lesion recognition learning method as described in claim 7, wherein step (a) includes: randomly selecting, by the processor, one of the multiple candidate lesion types associated with an organ type as the first lesion type, and randomly selecting one of the multiple candidate lesion images corresponding to an image type as the problem image from the multiple candidate lesion images corresponding to the first lesion type. The lesion recognition learning method as described in claim 7, wherein step (c) includes: using the processor, using the similarity algorithm, identifying a specific number of candidate lesion images with the highest similarity to the problem image from the multiple candidate lesion images corresponding to the first lesion type as the multiple reference images. The lesion recognition learning method as described in claim 7 further includes: using the processor to use the similarity algorithm to identify a specific number of candidate lesion images with the highest similarity to the problem image from the multiple candidate lesion images corresponding to the second lesion type as the multiple similar images. The lesion recognition learning method as described in claim 7 further includes: The processor displays the multiple reference images and the multiple similar images on the output interface; and the processor displays the problem image on the output interface and generates multiple colors at a position of multiple key feature points related to a lesion in the problem image.