Classification method and classification device for classifying the degree of age-related macular degeneration

Abstract

The present application provides a classification method and a classification device for classifying the degree of age-related macular degeneration. The classification method includes: pre-storing an object detection model and a first classification model; obtaining a fundus image; detecting a macular region in the fundus image according to the object detection model to generate a bounding box in the fundus image; calculating an intersection-union ratio of a default range in the fundus image and the bounding box; and generating a classification of the fundus image according to the first classification model in response to the intersection-union ratio being greater than a threshold value.

Description

Technical Field

[0001] The present invention relates to a classification method and a classification device for classifying the degree of age-related macular degeneration. Background Art

[0002] When using artificial intelligence (AI) models to determine the extent of age-related macular degeneration (AMD), the input data for the AI ​​models is mostly raw fundus images or pre-processed fundus images. However, when doctors judge the severity of AMD, they do not base their diagnosis on the entire eyeball in the fundus image, but rather on the macula in the fundus image. In other words, the AI ​​model and the doctor's judgment criteria are different. Therefore, the AI ​​model's judgment results may be inaccurate. Summary of the Invention

[0003] The present invention provides a classification method and a classification device for classifying the degree of age-related macular degeneration, which can classify the degree of AMD according to a reasonable range in a fundus image.

[0004] A classification device for classifying the degree of age-related macular degeneration (AMD) according to the present invention includes a processor, a storage medium, and a transceiver. The storage medium stores an object detection model and a first classification model. The processor is coupled to the storage medium and the transceiver and is configured to: obtain a fundus image via the transceiver; detect the macula in the fundus image according to the object detection model to generate a bounding box in the fundus image; calculate the intersection-union ratio between a default range in the fundus image and the bounding box; and, in response to the intersection-union ratio being greater than a threshold, generate a classification of the fundus image according to the first classification model.

[0005] In an embodiment of the present invention, the storage medium further stores a second classification model, wherein the processor is further configured to execute: generating a classification of the fundus image according to the second classification model in response to the intersection-union ratio being less than or equal to a threshold.

[0006] In one embodiment of the present invention, the processor inputs the image in the bounding box into a first classification model to generate a classification.

[0007] In one embodiment of the present invention, the processor inputs the fundus image into a second classification model to generate a classification.

[0008] In one embodiment of the present invention, the center point of the aforementioned default range is located at the geometric center of the fundus image.

[0009] In one embodiment of the present invention, the fundus image and the default range are rectangular, a first side of the default range is at a first distance from a first boundary of the fundus image, and a second side of the default range is at a first distance from a second boundary of the fundus image, wherein the second side is opposite to the first side, and wherein the second boundary is opposite to the first boundary.

[0010] In one embodiment of the present invention, the default range is a rectangle, wherein the processor is further configured to execute: obtaining an eyeball range in the fundus image; and generating a long side length and a short side length of the rectangle according to a diameter of the eyeball range.

[0011] In one embodiment of the present invention, the classification indicates that the fundus image corresponds to one of the following: the first stage, the second stage, the third stage, and the fourth stage of age-related macular degeneration.

[0012] In one embodiment of the present invention, the first classification model and the second classification model have the same convolutional neural network architecture.

[0013] A classification method for classifying the degree of age-related macular degeneration according to the present invention includes: pre-storing an object detection model and a first classification model; obtaining a fundus image; detecting the macula in the fundus image according to the object detection model to generate a bounding box in the fundus image; calculating the intersection-union ratio of a default range in the fundus image and the bounding box; and generating a classification of the fundus image according to the first classification model in response to the intersection-union ratio being greater than a threshold.

[0014] Based on the above, the classification device of the present invention can classify the degree of AMD based on the same criteria as a doctor when the macular position in the fundus image is reasonable. If the macular position in the fundus image is unreasonable, the classification device can classify the degree of AMD based on the entire fundus image. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 A schematic diagram illustrating a classification device for classifying the degree of AMD according to an embodiment of the present invention;

[0016] Figure 2 A schematic diagram showing a fundus image according to an embodiment of the present invention;

[0017] Figure 3 A schematic diagram showing a default range and a bounding box according to an embodiment of the present invention;

[0018] Figure 4 A flowchart of a classification method for classifying the degree of AMD is shown according to an embodiment of the present invention.

[0019] Description of Reference Numerals

[0020] 100: classification device;

[0021] 110: processor;

[0022] 120: storage medium;

[0023] 121: First classification model;

[0024] 122: Second classification model;

[0025] 123: Object detection model;

[0026] 130: transceiver;

[0027] 200: fundus image;

[0028] 21: First boundary;

[0029] 210: Eyeball range;

[0030] 22: Second boundary;

[0031] 220: macula;

[0032] 23: Third Boundary;

[0033] 24: Fourth Boundary;

[0034] 300: default range;

[0035] 31: first side;

[0036] 310: center point;

[0037] 32: Second side;

[0038] 33: third side;

[0039] 34: fourth side;

[0040] 400: bounding box;

[0041] D1: first distance;

[0042] D2: second distance;

[0043] S401, S402, S403, S404, S405: steps. DETAILED DESCRIPTION

[0044] Reference will now be made in detail to exemplary embodiments of the present invention, examples of which are illustrated in the accompanying drawings. Whenever possible, the same reference numerals are used in the drawings and the description to refer to the same or like parts.

[0045] Figure 1According to one embodiment of the present invention, a schematic diagram of a classification device 100 for classifying the degree of AMD is shown. The classification device 100 may include a processor 110, a storage medium 120, and a transceiver 130. The classification device 100 may be used to determine the degree of AMD corresponding to a fundus image. The classification device 100 may classify the fundus image input into the classification device 100 as stage 1, stage 2, stage 3, or stage 4 AMD.

[0046] The processor 110 may be, for example, a central processing unit (CPU), or other programmable general-purpose or special-purpose microcontrol unit (MCU), microprocessor, digital signal processor (DSP), programmable controller, application-specific integrated circuit (ASIC), graphics processing unit (GPU), image signal processor (ISP), image processing unit (IPU), arithmetic logic unit (ALU), complex programmable logic device (CPLD), field programmable gate array (FPGA), or other similar components or combinations thereof. The processor 110 may be coupled to the storage medium 120 and the transceiver 130 to access and execute multiple modules and various applications stored in the storage medium 120.

[0047] The storage medium 120 is, for example, any type of fixed or removable random access memory (RAM), read-only memory (ROM), flash memory, hard disk drive (HDD), solid state drive (SSD), or similar device, or a combination thereof, and is used to store multiple modules or various applications executable by the processor 110. In this embodiment, the storage medium 120 can store multiple models including a first classification model 121, a second classification model 122, and an object detection model 123.

[0048] The object detection model 123 can be used to detect the macula in the fundus image and generate a bounding box corresponding to the macula on the fundus image. The first classification model 121 can be used to classify the fundus image based on the image within the bounding box. In other words, the first classification model 121 classifies the fundus image based on a portion of the fundus image. The second classification model 122 can be used to classify the fundus image based on the entire fundus image.

[0049] The transceiver 130 transmits and receives signals wirelessly or by wire. The transceiver 130 may also perform operations such as low noise amplification, impedance matching, frequency mixing, up or down frequency conversion, filtering, amplification, and the like.

[0050] The processor 110 may obtain the fundus image through the transceiver 130 . Figure 2 According to one embodiment of the present invention, a schematic diagram of a fundus image 200 is shown. The fundus image 200 may include an eyeball range 210 and a macula 220. In one embodiment, the processor 110 may obtain the eyeball range 210 from the fundus image 200 according to the Hough transform. After obtaining the fundus image 200, the processor 110 may input the fundus image 200 into the object detection model 123. The object detection model 123 is, for example, a machine learning model. The object detection model 123 may detect the macula in the fundus image 200 to generate a bounding box on the fundus image 200, such as Figure 3 shown. Figure 3 A schematic diagram illustrating a default range 300 and a bounding box 400 is shown according to an embodiment of the present invention.

[0051] After generating the bounding box 400, the processor 110 may calculate the intersection over union (IOU) ratio between the default range 300 and the bounding box 400 in the fundus image 200. If the IOU ratio between the default range 300 and the bounding box 400 is greater than a threshold, the processor 110 may generate a classification for the fundus image 200 based on the first classification model 121. Specifically, the processor 110 may input the image within the bounding box 400 (i.e., the image of the macula 220) into the first classification model 121. The first classification model 121 may generate a classification for the fundus image 200 based on the image within the bounding box 400. The processor 110 may output the classification of the fundus image 200 via the transceiver 130 for the user's reference. The user may determine whether the fundus image 200 corresponds to stage 1, stage 2, stage 3, or stage 4 AMD based on the classification output by the transceiver 130.

[0052] If the intersection-union ratio of the default range 300 and the bounding box 400 is less than or equal to the threshold, the processor 110 may generate a classification for the fundus image 200 based on the second classification model 122. Specifically, the processor 110 may input the fundus image 200 (i.e., an image of the entire eyeball) into the second classification model 122. The second classification model 122 may generate a classification for the fundus image 200 based on the fundus image 200. The processor 110 may output the classification of the fundus image 200 via the transceiver 130 for user reference. The user may determine whether the fundus image 200 corresponds to stage 1, stage 2, stage 3, or stage 4 AMD based on the classification output by the transceiver 130.

[0053] The first classification model 121 or the second classification model 122 is, for example, a machine learning model. In one embodiment, the first classification model 121 and the second classification model 122 may have the same convolutional neural network architecture. However, because the first classification model 121 and the second classification model 122 are trained based on different training data sets or hyperparameters, the convolutional neural network in the first classification model 121 and the convolutional neural network in the second classification model 122 may have different weights.

[0054] The fundus image 200 and the default range 300 may be rectangular. The processor 110 may determine the position of the default range 300 based on the geometric center of the fundus image 200. The center point 310 of the default range 300 may be located at the geometric center of the fundus image 200. The fundus image 200 may have a first boundary 21, a second boundary 22, a third boundary 23, and a fourth boundary 24. The first boundary 21 and the second boundary 22 may be the short sides of the rectangle, and the third boundary 23 and the fourth boundary 24 may be the long sides of the rectangle. The second boundary 22 may be the opposite side of the first boundary 21, and the fourth boundary 24 may be the opposite side of the third boundary 23. On the other hand, the default range 300 may have a first side 31, a second side 32, a third side 33, and a fourth side 34. The first side 31 and the second side 32 may be the short sides of the rectangle, and the third side 33 and the fourth side 34 may be the long sides of the rectangle. The second side 32 may be an opposite side to the first side 31 , and the fourth side 34 may be an opposite side to the third side 33 .

[0055] In one embodiment, the processor 110 may determine the default range 300 based on the boundaries of the fundus image 200. Specifically, assuming that the default range 300 is a rectangle, the storage medium 120 may pre-store a first distance D1 and a second distance D2. The processor 110 may determine that the first side 31 of the default range 300 is spaced a first distance D1 from the first boundary 21 of the fundus image 200, and that the second side 32 of the default range 300 is spaced a first distance D1 from the second boundary 22 of the fundus image 200. Furthermore, the processor 110 may determine that the third side 33 of the default range 300 is spaced a second distance D2 from the third boundary 23 of the fundus image 200, and that the fourth side 34 of the default range 300 is spaced a second distance D2 from the fourth boundary 24 of the fundus image 200.

[0056] In one embodiment, the processor 110 may determine the default range 300 based on the diameter of the eye area 210. Assuming the default range 300 is a rectangle, the processor 110 may calculate the length of the long side and the length of the short side of the rectangle (i.e., the default range 300) based on the diameter of the eye area 210. For example, the processor 110 may multiply the diameter of the eye area 210 by 0.9 to calculate the length of the long side of the default range 300 (i.e., the length of the third side 33 or the fourth side 34). The processor 110 may multiply the diameter of the eye area 210 by 0.8 to calculate the length of the short side of the default range 300 (i.e., the length of the first side 31 or the second side 32).

[0057] Figure 4 According to an embodiment of the present invention, a flow chart of a classification method for classifying the degree of AMD is shown, wherein the classification method can be performed as follows: Figure 1 The classification device 100 shown in FIG. 1 is implemented as follows. In step S401, an object detection model and a first classification model are pre-stored. In step S402, a fundus image is obtained. In step S403, the macula in the fundus image is detected using the object detection model to generate a bounding box in the fundus image. In step S404, the intersection-union ratio between the default range in the fundus image and the bounding box is calculated. In step S405, a classification of the fundus image is generated based on the first classification model in response to the intersection-union ratio being greater than a threshold.

[0058] In summary, the classification device of the present invention can store a first classification model and a second classification model. The first classification model can judge the degree of AMD based on the bounding box in the fundus image, and the second classification model can judge the degree of AMD based on the entire fundus image. If the object detection model judges that the macula in the fundus image appears in the default range, it means that the position of the macula in the fundus image is reasonable. Accordingly, the classification device can classify the fundus image according to the first classification model. If the object detection model judges that the macula in the fundus image does not appear in the default range, it means that the position of the macula in the fundus image is unreasonable. Accordingly, the classification device can classify the fundus image according to the second classification model. In other words, if the position of the macula in the fundus image is reasonable, the classification device can classify the degree of AMD according to the same judgment criteria as the doctor.

[0059] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A classification device for classifying the degree of age-related macular degeneration, characterized in that: include: transceiver; a storage medium storing the object detection model, the first classification model, and the second classification model; as well as a processor coupled to the storage medium and the transceiver, wherein the processor is configured to execute: Obtaining an eye fundus image through the transceiver; detecting the macula in the fundus image according to the object detection model to generate a bounding box corresponding to the macula in the fundus image; Calculating the intersection-union ratio of a default range in the fundus image and the bounding box; In response to the intersection-union ratio being greater than a threshold, inputting the image in the bounding box into the first classification model to generate a classification of the fundus image; as well as In response to the intersection-union ratio being less than or equal to the threshold, the fundus image is input into the second classification model to generate the classification of the fundus image. 2 . The classification device according to claim 1 , wherein the center point of the default range is located at the geometric center of the fundus image.

3. The classification device according to claim 2, wherein the fundus image and the default range are rectangular, a first side of the default range is at a first distance from a first boundary of the fundus image, and a second side of the default range is at the first distance from a second boundary of the fundus image, wherein the second side is the opposite side of the first side, and wherein the second boundary is the opposite side of the first boundary.

4. The classification device according to claim 1, wherein the default range is a rectangle, and wherein the processor is further configured to perform: Obtaining the eyeball range in the fundus image; and The length of the long side and the length of the short side of the rectangle are generated according to the diameter of the eyeball range. 5 . The classification device according to claim 1 , wherein the classification indicates that the fundus image corresponds to one of the following: a first stage, a second stage, a third stage, and a fourth stage of the age-related macular degeneration.

6. The classification device according to claim 1, wherein the first classification model and the second classification model have the same convolutional neural network architecture.

7. A method for classifying the degree of age-related macular degeneration, characterized in that: include: Pre-stored object detection model, first classification model, and second classification model; Obtain fundus images; detecting the macula in the fundus image according to the object detection model to generate a bounding box corresponding to the macula in the fundus image; Calculating the intersection-union ratio of a default range in the fundus image and the bounding box; as well as In response to the intersection-union ratio being greater than a threshold, inputting the image in the bounding box into the first classification model to generate a classification of the fundus image; as well as In response to the intersection-union ratio being less than or equal to the threshold, the fundus image is input into the second classification model to generate the classification of the fundus image.

Images