System and method for detecting diabetic retinopathy

Abstract

Approaches for detecting presence of diabetic retinopathy (DR) in a patient's eye retinal image, are descried. In an example, the presence of referable DR is detected based on certain characteristics of eye indicating presence DR. To detect the presence of referable DR, a first detection model and a second detection model are trained based on a training information to predict presence of DR in an input eye image. Once trained, the first detection model and the second detection model may be used for ascertaining presence of referable DR in the input eye image.

Description

SYSTEM AND METHOD FOR DETECTING DIABETIC RETINOPATHYBACKGROUND

[0001] Eye is an organ whose function is to collect light from its surrounding and transform it into signals interpretable by the brain. Generally, an eye includes various components which enables the proper functioning, namely, cornea, iris, pupil, ciliary muscle, vitreous, retina, etc. Amongst these, the function of the retina is to transform incoming light from surrounding into electrical signals interpretable by the brain. There are some eye vision disorders or diseases which affects proper functioning of the retina of the eye. One of them is diabetic retinopathy which damages blood vessels in the retina resulting in swelling and leaking the blood vessels causing blurry vision, fluctuating vision, dark or empty areas in vision, or even vision loss. Generally, diabetic retinopathy is diagnosed by examining the retina of the eyes to detect the presence of numerous lesions, such as microneurysms, hemorrhages, soft and hard exudates and many more. However, none of the existing tests provide automatic detection of diabetic retinopathy with accurate results for large scale screening of population at an affordable cost.BRIEF DESCRIPTION OF FIGURES

[0002] Systems and / or methods, in accordance with examples of the present subject matter are now described and with reference to the accompanying figures, in which:

[0003] FIG. 1 illustrates a training system for training a first detection model and a second detection model enabling them to detect diabetic retinopathy, as per an example;

[0004] FIG. 2 illustrates a detection system for determining presence of diabetic retinopathy in an input eye image, as per another example;

[0005] FIG. 3 illustrates a method for training a first detection model and a second detection model, as per an example; and

[0006] FIG. 4A-4B illustrates a method for determining presence of referable DR in an input eye image, based on a trained first and second detection model, as per an example.DETAILED DESCRIPTION

[0007] Eyes are a sensory organ capable of reacting to visible light and allowing this light to be converted to brain interpretable signals. Brain utilizes such signals for various purposes including seeing things, keeping balance, etc. Amongst other features or components of the eyes, retina at the back of the eyes is one of the important components which causes the photoreceptors to turn the light into electrical signals on illumination with light. These electrical signals then travel from the retina through the optic nerve to the brain for further processing. Such electric signals are then processed by the brain to create a visual feed or perception of surrounding objects which we see as images or videos.

[0008] An individual may suffer from different eye vision disorders. Examples of such vision disorder may include, but are not limited to, blurred vision (refractive errors), age-related macular degeneration, glaucoma, cataract, diabetic retinopathy, etc. One such visual disorder is diabetic retinopathy, an eye condition that can cause vision loss and blindness in people who have diabetes. Diabetic retinopathy (DR) affects blood vessels in the retina and may involve the growth of abnormal blood vessels in the retina. Generally, diabetic retinopathy causes blurry vision, fluctuating vision, dark or empty areas in vision, or even vision loss.

[0009] Eye affected with diabetic retinopathy generally encompasses different types of lesions formed on the retina. Examples of such lesions include, but may not be limited to, microaneurysms, retinal hemorrhages, soft exudates, hard exudates, intraretinal microvascular anomalies (IRMA), neovascularization, fibrous proliferation, and preretinal and vitreous hemorrhage, tractional retinal detachment. DR is generally caused by high blood sugar persisted for long time due to diabetes. The damage caused byDR is irreversible, however, with proper and timely diagnosis may help in preventing progression of DR.

[0010] Conventionally, DR is best diagnosed with a comprehensive dilated eye exam in which drops placed in the patient’s eye widen its pupils to allow a medical practitioner to have a better view inside the eyes of the patient. Other ancillary diagnosis processes include fluorescein angiography, optical coherence tomography (OCT), etc. However, the above disclosed techniques or other diagnostic methods require either a specialist medical practitioner or expensive equipment. As may be understood, presence of such specialized medical practitioner and equipment is limited to tertiary level health care centre which are far away from the reach of rural population, which is highest in India. To perform screening of large population at an affordable cost, there is a need for a system which performs automatic detection of DR having an on-the-edge operable configuration to reduce cost and time of operation of such system.

[0011] Approaches for detecting presence of diabetic retinopathy (DR) in a patient’s eye retinal image, are descried. In an example, the presence of referable DR is detected based on certain characteristics of eye indicating presence of DR. To detect the presence of referable DR, a first detection model and a second detection model are trained based on a training information to predict presence of referable DR in an input eye image. In an example, the first detection model and the second detection are machine learning model which are trained based on the training information. In an example, the training information may include processed training images which may be used to train the first detection model and unprocessed training images which may be used to train the second detection model. The processed training images are the images which may have gone through a set of pre-processing steps before using for training the first detection model. On the other hand, the unprocessed training images are the images which are not pre-processed before using for training the second detection model. The training eye images included in the training information includeplurality of retinal images centred across different parts of retina, such as disc centred, macula centred, etc.

[0012] In one example, before training the models based on the training eye images, the processed training images are processed to extract processed information related to training eye images and the unprocessed training images are further processed to extract unprocessed information related to each of the training eye images. In an example, the processed information and the unprocessed information includes a plurality of eye characteristics with their corresponding attribute value. In some cases, such information may be annotated beforehand with each of the training eye images.

[0013] Examples of eye characteristics include, but may not be limited to, size, location, and area of different types of lesions, such as microaneurysms, retinal hemorrhages, soft exudates, hard exudates, intraretinal microvascular anomalies, neovascularization, fibrous profileration, and preretinal and vitreous hemorrhage, tractional retinal detachment present in the images. Based on the analysis of the attribute values of the eye characteristics of the processed training images, a confidence score is associated with the attribute values of the eye characteristics in the first detection model. Further, based on the analysis of the attribute values of the eye characteristic of the unprocessed training images, a reliance score is associated with the attribute values of the eye characteristics in the second detection model. The associated confidence score and the reliance score are used for ascertaining presence of referable DR in the input eye image.

[0014] Once trained, the first detection model and the second detection model may be used for ascertaining presence of referable DR in the input eye image. The input eye image may be a retinal image of the eye of a patient which is under screening. Such input eye image may be either stored in a database repository or may be captured by a camera device on realtime basis. In one example, input eye image which is to be screened fordetecting referable DR, is obtained. In another example, a set of retinal images may also be obtained for screening. In such a case, collective analysis of these images is used for determination of presence of referable DR.

[0015] Once obtained, initially, numerous acceptability tests are performed on the input eye image. For example, the input eye image is processed to check its quality and its field of view. Based on the determined quality and field of view, the input eye image is either accepted for further analysis or a warning is displayed to the user. In case of warning, based on the user’s input, the input eye image may be used for further analysis even it has failed acceptability tests. Thereafter, a pre-processing step is performed on the input eye image to obtain a processed image. Now, the processed image is used with the first detection model and the input eye image as the unprocessed image is used with the second detection model. In an example, the processed image and the unprocessed image are further processed to derive corresponding information, i.e., processed information from processed image and unprocessed information from unprocessed image.

[0016] In one example, such information indicates information regarding structural configuration of different lesions forming in the retina. Examples of such eye characteristics include, but may not be limited to, size, location, and color of different types of lesions present in the retinal image, wherein the types of lesions comprise one of microaneurysms, retinal hemorrhages, soft exudates, hard exudates, intraretinal microvascular anomalies, neovascularization, fibrous proliferation, and preretinal and vitreous hemorrhage, tractional retinal detachment or combination thereof. Once derived, the processed information is used to operate the first detection model to generate a confidence score and the unprocessed information is used to operate the second detection model to generate a reliance score. Once the confidence score and the reliance score are obtained, acomparison of the confidence score and the reliance score individually with a predefined value is performed.

[0017] Based on the comparison, on ascertaining one of the scores, i.e., confidence score and the reliance score, being greater than the predefined value, the patient’s eye is designated as referable DR positive eye. In case none of the scores is greater than predefined value, further calculations are performed to identify the state of the patient’s eye.

[0018] As will be explained further, the present approaches perform binary classification of the input eye image as either referable DR or DR negative / non-referable based on the eye characteristics of the input eye image. Further, the usage of two separate detection model, i.e., first detection model for processed image and the second detection model for unprocessed image, results in more efficient, reliable and accurate outcome.

[0019] The explanation provided above and the examples that are discussed further in the current description are exemplary only. For instance, some of the examples may have been described in which only one image is considered, either in training or in inference stage. However, the current approaches may be adopted for other instances or situations as well, such as a set of input eye images, a set of training eye images may be used, or such without deviating from the scope of the present subject matter.

[0020] The manner in which the first detection model and the second detection model are trained and used for ascertaining presence of referable DR in input eye image is explained with respect to FIGS. 1 -4. While aspects of described systems may be implemented in any number of different electronic devices, environments, and / or implementation, the examples are described in the context of the following example device (s). In another example, the aspects of the present subject matter may also be implemented by a standalone device having executable instructions. It may be noted that drawings of the present subject matter shown here are ofillustrative purpose and are not to be construed as limiting the scope of the subject matter claimed.

[0021] FIG. 1 illustrates a training system 102 comprising a processor or memory (not shown), for training a first detection model and a second detection model. In an example, the training system 102 (referred to as system 102) may be communicatively coupled to a repository 104 through a network 106. The repository 104 may include a training information 108. In an example, the training information 108 may include a processed training image and an unprocessed training image of a patient’s eye. The processed training image is a pre-processed image of the patient’s eye and the unprocessed image is a normal image without being pre-processed of the patient’s eye. Example of pre-processing step is subtracting background color from the eye image.

[0022] In another example, the training information 108 may include only a training image which is further pre-processed to obtain the processed training image and is treated in its original form as unprocessed training image. It may be noted that, the training information 108 is depicted to have only one pair of training image, i.e., one processed training image and one unprocessed training image, however, more than one pair of training images may also be present in the training information 108 without deviating from the scope of the present subject matter.

[0023] In another example, along with images, training information 108 may further include corresponding DR diagnosis along with the training images. For example, the processed training image is associated with a DR diagnosis. In an example, the DR diagnosis is associated based on the analysis of the images by a trained medical practitioner. T raining information 108, although depicted as being obtained from a single repository, such as repository 104, may also be obtained from multiple other sources without deviating from the scope of the present subject matter. In such cases, each of such multiple repositories may be interconnected through a network, such as the network 106.

[0024] The network 106 may be a private network or a public network and may be implemented as a wired network, a wireless network, or a combination of a wired and wireless network. The network 106 may also include a collection of individual networks, interconnected with each other and functioning as a single large network, such as the Internet. Examples of such individual networks include, but are not limited to, Global System for Mobile Communication (GSM) network, Universal Mobile Telecommunications System (UMTS) network, Personal Communications Service (PCS) network, Time Division Multiple Access (TDMA) network, Code Division Multiple Access (CDMA) network, Next Generation Network (NGN), Public Switched Telephone Network (PSTN), Long Term Evolution (LTE), and Integrated Services Digital Network (ISDN).

[0025] The system 102 may further include instructions 110 and a training engine 112. In an example, the instructions 110 are fetched from a memory and executed by a processor included within the system 102. The training engine 112 may be implemented as a combination of hardware and programming, for example, programmable instructions to implement a variety of functionalities. In examples described herein, such combinations of hardware and programming may be implemented in several different ways. For example, the programming for the training engine 112 may be executable instructions, such as instructions 110. Such instructions may be stored on a non-transitory machine-readable storage medium which may be coupled either directly with the system 102 or indirectly (for example, through networked means). In an example, the training engine 112 may include a processing resource, for example, either a single processor or a combination of multiple processors, to execute such instructions. In the present examples, the non-transitory machine-readable storage medium may store instructions, such as instructions 110, that when executed by the processing resource, implement training engine 112. In other examples, the training engine 112 may be implemented as electronic circuitry.

[0026] The instructions 110, when executed by the processing resource, causes the training engine 112 to train a first detection model 114 and a second detection model 116. The instructions 110 may be executed by the processing resource for training the first detection model 114 and the second detection model 116 based on the training information 108. The system 102 may further include a processed training images 118, and unprocessed training images 120. In an example, the system 102 may obtain training information 108 from the repository 104 comprising processed training image and unprocessed training image, and the information pertaining to these images is stored as processed training images 118, and unprocessed training images 120 in the system 102.

[0027] Returning to the present example, for training, firstly, the first detection model 114 is trained based on a processed training information derived from the processed training images 118 and the second detection model 116 is trained based on the unprocessed training information derived from the unprocessed training images 120. Once trained, the first detection model 114 is to generate a confidence score and the second detection model 116 is to generate a reliance score. In an example, the processed and unprocessed training information include plurality training eye characteristics having corresponding training attribute values. Examples of training eye characteristics include, but may not be limited to, size, location, and color of different types of lesions present in the retinal image, wherein the types of lesions comprise one of microaneurysms, retinal hemorrhages, soft exudates, hard exudates, interretinal microvascular anomalies, neovascularization, fibrous profileration, and preretinal and vitreous hemorrhage, or combination thereof. It may be noted that, above disclosed training eye characteristics are exemplary, distinct characteristics based on the type of images present in the training information 108 may be used.

[0028] In operation, the system 102 obtain the training information 108 from the repository 104 which may be stored as processed training images 118, unprocessed training images 118, referable DR diagnosis in thesystem 102. In another example, the training information 108 includes only one training image which when pre-processed to subtract the background of the image is to generate the processed training image 118 and the training image in its original form is designated as unprocessed training image 120. In an example, the processed training image 118 and the unprocessed training image 120 are either already cropped to eliminate borders of the images or cropped later by the training engine 112.

[0029] Thereafter, the training engine 112 derives a processed training information from the processed training image 118 and an unprocessed training information from the unprocessed training image 120. In an example, the training engine 112 uses any feature extraction technique to derive this information. The processed training information and the unprocessed training information includes a plurality of training eye characteristics with corresponding training attribute value. In an example, the plurality of training eye characteristics includes size, location, and color of different types of lesions present in the eye image. Examples of such eye characteristics include, but may not be limited to, size, location, and color of different types of lesions present in the retinal image, wherein the types of lesions comprise one of microaneurysms, retinal hemorrhages, soft exudates, hard exudates, intraretinal microvascular anomalies, neovascularization, fibrous proliferation, and preretinal and vitreous hemorrhage, tractional retinal detachment or combination thereof.

[0030] Once derived, the training engine 112 trains the first detection model 114 based on the processed training information derived from the processed training image 118. The first detection model 114, once trained, is to generate a confidence score. For example, while training the first detection model 114, training engine 112 classifies the training attribute values corresponding to the processed training information as a first processed training eye characteristics. In an example, the first training eye characteristic belongs to a first range of attribute values of the training eye characteristics. Once classified, the training engine 112 associates the firstprocessed training eye characteristic with a first confidence score, based on the training attribute values. For example, based on the attribute values of size, location, and color of different types of lesions, the training engine 112 associates the first confidence score, with the first processed training eye characteristic.

[0031] Thereafter, the training engine 112 trains the second detection model 116 based on the unprocessed training information derived from the unprocessed training image 120. The second detection model 116, once trained, is to generate a reliance score. For example, while training the second detection model 116, training engine 112 classifies the training attribute values corresponding to the unprocessed training information as a first unprocessed training eye characteristics. In an example, the first unprocessed training eye characteristic belongs to a first range of attribute values of the training eye characteristics. Once classified, the training engine 112 associates the first unprocessed training eye characteristic with a first reliance score based on the training attribute values. For example, based on the attribute values of size, location, and color of different types of lesions, the training engine 112 associates the first reliance score with the first unprocessed training eye characteristic. In an example, the confidence score and the reliance score generated by the first detection model 114 and the second detection model 116, respectively, represents probability of detecting referable DR in the eye image.

[0032] In an example, while training, whenever a subsequent training information is received, the training engine 112 obtains a subsequent processed training image for training first detection model 114 and a subsequent unprocessed training image for training second detection model 116 included in the subsequent training information.

[0033] For example, in case of training first detection model 114 based on subsequent processed training image, the training engine 112 derives processed training information corresponding to the subsequent processed training image which further includes training attribute values. Once derived,the training engine 112 ascertains whether training attribute values corresponding to the subsequent processed training image lies in the range of the training attribute values of first processed training eye characteristic.

[0034] On ascertaining training attribute values does not correspond to the first processed training eye characteristics, the training engine 112 classifies the training attribute values corresponding to the subsequent processed training image as a second processed training eye characteristic in the first detection model 114. In an example, the second processed training eye characteristics belongs to a second range of attribute values. Once classified, the training engine 112 associates second processed training eye characteristic with a second confidence score, based on the training attribute values corresponding to the subsequent processed training image.

[0035] On the other hand, on ascertaining training attribute values corresponding to the subsequent processed training image to correspond to the first processed training eye characteristics, the training engine 112 associates the first confidence score to the second processed training eye characteristic as well.

[0036] Similarly, in case of training second detection model 116 based on subsequent unprocessed training image, the training engine 112 derives unprocessed training information corresponding to the subsequent unprocessed training image which further includes training attribute values. Once derived, the training engine 112 ascertains whether training attribute values corresponding to the subsequent unprocessed training image lies in the range of the training attribute values of first unprocessed training eye characteristic.

[0037] On ascertaining training attribute values does not correspond to the first unprocessed training eye characteristics, the training engine 112 classifies the training attribute values corresponding to the subsequent unprocessed training image as a second unprocessed training eye characteristic in the second detection model 116. In an example, the secondunprocessed training eye characteristics belongs to a second range of attribute values. Once classified, the training engine 112 associates second unprocessed training eye characteristic with a second reliance score, based on the training attribute values corresponding to the subsequent unprocessed training image.

[0038] On the other hand, on ascertaining training attribute values corresponding to the subsequent unprocessed training image to correspond to the first unprocessed training eye characteristics, the training engine 112 associates the first confidence score to the second unprocessed training eye characteristic as well.

[0039] It may be noted that the examples as described above have been described in the context of the first confidence score, second confidence score, first reliance score, and the second reliance score. The same has been done for the ease of explanation and should not be construed as a limitation. The approaches as described above may be further implemented to further train both the models.

[0040] Once trained, the first detection model 114 and the second detection model 116 may be used to determine or predict the presence of referable DR in an input eye image. For example, processed information and unprocessed information pertaining to the input eye image may be processed based on the first detection model 114 and the second detection model 116, respectively. In an example, based on the first detection model 114 and the second detection model 116, confidence score and reliance score are generated for the input eye image. Thereafter, based on the analysis of the confidence score and the reliance score, the presence of referable DR is predicted or determined.

[0041] The manner in which the confidence score and the reliance score generated based on first detection model 114 and the second detection model 116, respectively, is further described in conjunction with FIG. 2.

[0042] FIG. 2 illustrates a clinical environment 200 with a detection system 202 (referred to as system 202), for determining the presence ofdiabetic retinopathy (DR) in an input eye image 204 of a patient 206. In an example, the system 202 includes a retinal camera for capturing retinal images, e.g., input eye image 204 is also a retinal image. The input eye image 204 may be an image of an eye of the patient 206 who is under screening for the diagnosis of DR.

[0043] Similar to the system 102, the system 202 may further include instructions 208 and an analysis engine 210. In an example, the instructions 208 are fetched from a memory and executed by a processor included within the system 202. The analysis engine 210 may be implemented as a combination of hardware and programming, for example, programmable instructions to implement a variety of functionalities. In examples described herein, such combinations of hardware and programming may be implemented in several different ways. For example, the programming for the analysis engine 210 may be executable instructions, such as instructions 208. Such instructions 208 may be stored on a non-transitory machine-readable storage medium which may be coupled either directly with the system 202 or indirectly (for example, through networked means). In an example, the analysis engine 210 may include a processing resource, for example, either a single processor or a combination of multiple processors, to execute such instructions. In the present examples, the non- transitory machine-readable storage medium may store instructions, such as instructions 208, that when executed by the processing resource, implement analysis engine 210. In other examples, the analysis engine 210 may be implemented as electronic circuitry.

[0044] The system 202 may include a first detection model, such as first detection model 1 14 and a second detection model, such as second detection model 1 16. The system 202 may further include an input eye image 212, processed information 214, unprocessed information 216, confidence score 218, and reliance score 220. In an example, the processed information 214 and the unprocessed information 216 may be derived by the analysis engine 210 as a result of the execution or analysisof the input eye image 212 based on the first detection model 114 and second detection model 1 16, respectively. The processed information 214 and the unprocessed information 216 includes an attribute value corresponding to each of a plurality of eye characteristics. In an example, eye characteristics may include, but may not be limited to, size, location, and color of different types of lesions present on retina of the eye. Examples of such eye characteristics include, but may not be limited to, size, location, and color of different types of lesions present in the retinal image, wherein the types of lesions comprise one of microaneurysms, retinal hemorrhages, soft exudates, hard exudates, intraretinal microvascular anomalies, neovascularization, fibrous proliferation, and preretinal and vitreous hemorrhage, tractional retinal detachment or combination thereof.

[0045] In operation, initially, the system 202 may obtain an image of an eye of a suspected patient who is under screening for the detection of DR. The image of the eye, i.e., input eye image 204, is stored as input eye image 212 in system 202. In an example, the input eye image 212 is captured by the system 202 using a camera device or module which is either installed on the system 202 itself or connected externally to the system 202. In another example, there are other external hardware equipment that needs to be installed along with the system 202 to capture or get the retinal view of an eye of a person. In another example, the input eye image 212 may be obtained from a database repository (not shown in FIG. 2) storing a plurality of input eye image to be tested or diagnosed for detecting presence of referable DR.

[0046] Thereafter, the analysis engine 210 processes the input eye image 212 based on the first detection model 1 14 and the second detection model 116 to determine presence of referable DR in the patient’s eye. However, before proceeding with this, the analysis engine 210 processes the input eye image 212 based on a quality assessment test to determine the quality information of the input eye image 212. This step of identifying quality of the input eye image 212 is important because it is difficult toprocess low-quality image to derive processed information 214 and unprocessed information 216. Therefore, if the quality of the input eye image 212 comes out to be low compared to standard quality requirement, a warning prompt is generated and displayed on the display device of the system 202 to indicate to quality issue to the user. On the other hand, if the quality of the input eye image 212 is as per the allowed quality standard, the analysis engine 210 proceeds with the input eye image 212 for further processing steps.

[0047] Returning to the present example, on ascertaining the quality of the input eye image 212 to be acceptable, the analysis engine 210 determines the field of view of the input eye image 212. On determining the field of view of the input eye image 212 as one of disc centered and macula centered, the analysis engine 210 accepts the input eye image 212 and proceeds to further processing steps. On the other hand, on determining the field of view of the input eye image 212 as non-macula centered and nondisc centered, the analysis engine 210 displays the warning prompt on the display device of the system 202 to direct the user to re-capture the input eye image 212 or obtain another input eye image from the repository.

[0048] It may be noted that, while processing the input eye image 212 to detect presence of referable DR, it is important to have either a macula centered or a disc centered view of image because most of the changes in retinal structure in response to DR are identified or captured in disc or macula region only. Therefore, in order to efficiently and correctly process the input eye image 212, the image needs to be one of macula centered and disc centered view. In another example, before proceeding further, the input eye image 212 may be processed by the analysis engine 210 to crop the borders.

[0049] As described in conjunction with FIG. 1 , the first detection model 114 is trained based on a processed image while the second detection model 116 is trained based on an unprocessed image. Therefore, processed image is used for operating the first detection model 114 andunprocessed image is used for operating the second detection model 116. In an example, once the quality and field of view is determined to be acceptable, the input eye image 212 is pre-processed by the analysis engine 210 to subtract the background from the input eye image 212. Therefore, the processed image is the one in which the background of the image is subtracted. On the other hand, the unprocessed image is the one in which the background of the image is not subtracted. In another example, in case of obtaining input eye image 212 from the data repository, the analysis engine 210 obtains a pair of images including the processed image and unprocessed and there is no need to process the input eye image 212 to obtain the processed image.

[0050] Once the processed image and the unprocessed image are obtained, in an example, the analysis engine 210 derives a processed information, such as processed information 214 from the processed image and an unprocessed information, such as unprocessed information 216 from the unprocessed image using any feature extraction technique. In an example, the processed information 214 and the unprocessed information 216 includes a plurality of eye characteristics having corresponding attribute values. Examples of such eye characteristics include, but may not be limited to, size, location, and color of different types of lesions present in the retinal image, wherein the types of lesions comprise one of microaneurysms, retinal hemorrhages, soft exudates, hard exudates, intraretinal microvascular anomalies, neovascularization, fibrous proliferation, and preretinal and vitreous hemorrhage, tractional retinal detachment or combination thereof.

[0051] Thereafter, the analysis engine 210 generate a confidence score, such as confidence score 218 by operating the first detection model 114 based on the processed information 214. In an example, the processed information is derived from the processed image of the patient’s eye. For example, to generate the confidence score 218, the analysis engine 210 processes the attribute values included in the processed information 214based on the first detection model 114. As a result, a weight for each of the plurality of eye characteristics is assigned to generate a first weighted information. Thereafter, based on the first weighted information, the analysis engine 210 generate the confidence score 218 for the processed image using trained first detection model 114.

[0052] Similarly, the analysis engine 210 generate a reliance score, such as reliance score 220 by operating the second detection model 116 based on the unprocessed information 216. In an example, the unprocessed information 216 is derived from the unprocessed image of the patient’s eye. For example, to generate the reliance score 220, the analysis engine 210 processes the attribute values included in the unprocessed information 216 based on the second detection model 116. As a result, a weight for each of the plurality of eye characteristics is assigned to generate a second weighted information. Thereafter, based on the second weighted information, the analysis engine 210 generate the reliance score 220 for the unprocessed image using trained second detection model 116.

[0053] Once generated, the analysis engine 210 compares the confidence score 218 and the reliance score 220 individually with a predefined value. The comparison is performed to ascertain whether one of confidence score 218 and the reliance score 220 is greater than the predefined value. In an example, the value of confidence score 218 and the reliance score 220 varies between 0 and 1 . It may be noted that, values mentioned for confidence score and the reliance score 220 is exemplary and any other value may be used without deviating from the scope of the present subject matter.

[0054] Continuing further, on ascertaining one of the scores being greater than the predefined value, the analysis engine 210 designated the input eye image as referable DR. On the other hand, on ascertaining none of the scores being greater than the predefined value, the analysis engine 210 further analyze the confidence score 218 and the reliance score 220 based on a linear combination. For example, the analysis engine 210designate the input eye image 212 as referable DR if following condition satisfies: x*(reliance score) + y*(confidence score) -z > 0

[0055] Further, on ascertaining none of the scores being greater than the predefined value, the analysis engine 210 designates the input eye image 212 as DR negative / non-referable if following condition satisfies: x*(reliance score) + y*(confidence score) -z <= 0 wherein x, y, and z represent any real number value except ‘0’.

[0056] Therefore, in such a way the DR positiveness is identified by the system 202 to indicate its presence in the input eye image 212 to the patient or the person operating the system 202. Based on the result, the patient or the medical person appropriately take further steps relating to treatment of DR so that it may not progress further.

[0057] In another example, the system 202 may be communicatively coupled to a central computing server through a network (not shown in FIG. 2). The network may be a private network or a public network and may be implemented as a wired network, a wireless network, or a combination of a wired and wireless network, and may be similar to the network 106 (as depicted in FIG. 1 ). All the above disclosed steps which may be performed by the analysis engine 210 of the system 202, may be implemented or performed by the central computing server on behalf of the system 202 to reduce computing load on the edge of the network.

[0058] FIG. 3 illustrates example method 300 for training a first detection model and a second detection model for enabling them to predict presence of diabetic retinopathy, in accordance with examples of the present subject matter. The order in which the above-mentioned method is described is not intended to be construed as a limitation, and some of the described methodblocks may be combined in a different order to implement the method, or alternative method.

[0059] Furthermore, the above-mentioned method may be implemented in suitable hardware, computer-readable instructions, or combination thereof. The steps of such method may be performed by either a system under the instruction of machine executable instructions stored on a non- transitory computer readable medium or by dedicated hardware circuits, microcontrollers, or logic circuits. For example, the method may be performed by a training system, such as system 102. In an implementation, the method may be performed under an “as a service” delivery model, where the system 102, operated by a provider, receives programmable code. Herein, some examples are also intended to cover non-transitory computer readable medium, for example, digital data storage media, which are computer readable and encode computer-executable instructions, where said instructions perform some or all the steps of the above- mentioned method.

[0060] In an example, the method 300 may be implemented by the system 102 for training the first detection model 114 and the second detection model 116 based on a training information, such as training information 108. At block 302, a training information is obtained from a repository. For example, the system 102 obtain the training information 108 from the repository 104 which may be stored as processed training image 1 18, and unprocessed training image 120. In another example, the training information 108 includes only one training image which when pre-processed to subtract the background of the image is to generate the processed training image 1 18 and the training image in its original form is designated as unprocessed training image 120. In an example, the processed training image 118 and the unprocessed training image 120 are either already cropped to eliminate borders of the images or cropped later by the training engine 112.

[0061] At block 304, a processed training information corresponding to the processed training image and an unprocessed training information corresponding to the unprocessed training image are derived. For example, the training engine 112 derives a processed training information from the processed training image 1 18 and an unprocessed training information from the unprocessed image 120. In an example, the training engine 112 uses any feature extraction technique to derive these information. The processed training information and the unprocessed training information includes a plurality of training eye characteristics with corresponding training attribute value. In an example, the plurality of training eye characteristics includes size, location, and color of different types of lesions present in the eye image. Examples of such eye characteristics include, but may not be limited to, size, location, and color of different types of lesions present in the retinal image, wherein the types of lesions comprises one of microaneurysms, retinal hemorrhages, soft exudates, hard exudates, intraretinal microvascular anomalies, neovascularization, fibrous proliferation, and preretinal and vitreous hemorrhage, tractional retinal detachment or combination thereof.

[0062] At block 306, a first detection model is trained based on the processed training information. For example, the training engine 112 trains the first detection model 114 based on the processed training information derived from the processed training image 118. The first detection model 114, once trained, is to generate a confidence score. For example, while training the first detection model 114, training engine 112 classifies the training attribute values corresponding to the processed training information as a first processed training eye characteristics. In an example, the first training eye characteristic belongs to a first range of attribute values of the training eye characteristics. Once classified, the training engine 112 associates the first processed training eye characteristic with a first confidence score, based on the training attribute values. For example, based on the attribute values of size, location, and color of different typesof lesions, the training engine 112 associates the first confidence score, with the first processed training eye characteristic.

[0063] At block 308, a second detection model is trained based on the unprocessed training information. For example, the training engine 112 trains the second detection model 116 based on the unprocessed training information derived from the unprocessed training image 120. The second detection model 116, once trained, is to generate a reliance score. For example, while training the second detection model 116, training engine 112 classifies the training attribute values corresponding to the unprocessed training information as a first unprocessed training eye characteristics. In an example, the first unprocessed training eye characteristic belongs to a first range of attribute values of the training eye characteristics. Once classified, the training engine 112 associates the first unprocessed training eye characteristic with a first reliance score, based on the training attribute values. For example, based on the attribute values of size, location, and color of different types of lesions, the training engine 112 associates the first reliance score, with the first unprocessed training eye characteristic.

[0064] At block 310, a subsequent training information is obtained. For example, while training, whenever a subsequent training information is received, the training engine 112 obtains a subsequent processed training image for training first detection model 114 and a subsequent unprocessed training image §or training second detection model 116 included in the subsequent training information.

[0065] At block 312, the first detection model is trained based on a subsequent processed training image. For example, in case of training first detection model 114 based on subsequent processed training image, the training engine 112 derives processed training information corresponding to the subsequent processed training image which further includes training attribute values. Once derived, the training engine 112 ascertains whether training attribute values corresponding to the subsequent processedtraining image lies in the range of the training attribute values of first processed training eye characteristic.

[0066] On ascertaining training attribute values does not correspond to the first processed training eye characteristics, the training engine 112 classifies the training attribute values corresponding to the subsequent processed training image as a second processed training eye characteristic in the first detection model 116. In an example, the second processed training eye characteristics belongs to a second range of attribute values. Once classified, the training engine 112 associates second processed training eye characteristic with a second confidence score, based on the training attribute values corresponding to the subsequent processed training image.

[0067] On the other hand, on ascertaining training attribute values corresponding to the subsequent processed training image to correspond to the first processed training eye characteristics, the training engine 112 associates the first confidence score to the second processed training eye characteristic as well.

[0068] At block 314, the second detection model is trained based on a subsequent unprocessed training image. For example, in case of training second detection model 116 based on subsequent unprocessed training image, the training engine 112 derives unprocessed training information corresponding to the subsequent unprocessed training image which further includes training attribute values. Once derived, the training engine 112 ascertains whether training attribute values corresponding to the subsequent unprocessed training image lies in the range of the training attribute values of first unprocessed training eye characteristic.

[0069] On ascertaining training attribute values does not correspond to the first unprocessed training eye characteristics, the training engine 112 classifies the training attribute values corresponding to the subsequent unprocessed training image as a second unprocessed training eye characteristic in the second detection model 116. In an example, the secondunprocessed training eye characteristics belongs to a second range of attribute values. Once classified, the training engine 112 associates second unprocessed training eye characteristic with a second reliance score, based on the training attribute values corresponding to the subsequent unprocessed training image.

[0070] On the other hand, on ascertaining training attribute values corresponding to the subsequent unprocessed training image to correspond to the first unprocessed training eye characteristics, the training engine 112 associates the first confidence score to the second unprocessed training eye characteristic as well.

[0071] It may be noted that the example method as described above have been described in the context of the first confidence score, second confidence score, first reliance score, and the second reliance score. The same has been done for the ease of explanation and should not be construed as a limitation. The approaches as described above may be further implemented to further train both the models.

[0072] Once trained, the first detection model 114 and the second detection model 116 may be used to determine or predict the presence of referable DR in an input eye image. For example, processed information and unprocessed information pertaining to the input eye image may be processed based on the first detection model 114 and the second detection model 116, respectively. In an example, based on the first detection model 114 and the second detection model 116, confidence score and reliance score are generated for the input eye image. Thereafter, based on the analysis of the confidence score and the reliance score, the presence of referable DR is predicted or determined.

[0073] FIG. 4A-4B illustrates another example method 400 for determining presence of referable DR in an input eye image, in accordance with an example of the present subject matter. Based on the present approaches as described in the context of example method 400, the eye characteristics of an input eye image is analyzed based on the trained firstdetection model 114 and the second detection model 116 to ascertain presence of referable DR in the input eye image. The order in which the above-mentioned method is described is not intended to be construed as a limitation, and some of the described method blocks may be combined in a different order to implement the method, or alternative method.

[0074] Further, the above-mentioned method 400 may be implemented in suitable hardware, computer-readable instructions, or combination thereof. The steps of such method may be performed by either a system under the instruction of machine executable instructions stored on a non- transitory computer readable medium or by dedicated hardware circuits, microcontrollers, or logic circuits. For example, the method may be performed by a detection system, such as system 202. In an implementation, the method may be performed under an “as a service” delivery model, where the system 202, operated by a provider, receives programmable code. Herein, some examples are also intended to cover non-transitory computer readable medium, for example, digital data storage media, which are computer readable and encode computer-executable instructions, where said instructions perform some or all the steps of the above-mentioned method.

[0075] At block 402, an input eye image is obtained. For example, the system 202 may obtain an image of an eye of a suspected patient who is under screening for the detection of DR. The image of the eye, i.e., input eye image 204, is stored as input eye image 212 in the system 202. In an example, the input eye image 212 is captured by the system 202 using a camera device or module which is either installed on the system 202 itself or connected externally to the system 202. In another example, there are other external hardware equipment that needs to be installed along with the system 202 to capture or get the retinal view of an eye of a person. In another example, the input eye image 212 may be obtained from a database repository (not shown in FIG. 2) storing a plurality of input eye image to be tested or diagnosed for detecting presence of referable DR.

[0076] At block 404, certain quality assessment tests are performed on the input eye image to ascertain its acceptability. For example, the analysis engine 210 processes the input eye image 212 based on a quality assessment test to determine the quality information of the input eye image 212. This step of identifying quality of the input eye image 212 is important because it is difficult to process low-quality image to derive processed information 214 and unprocessed information 216. Therefore, if the quality of the input eye image 212 comes out to be low compared to standard quality requirement, a warning prompt is generated and displayed on the display device of the system 202 to indicate to quality issue to the user. On the other hand, if the quality of the input eye image 212 is as per the allowed quality standard, the analysis engine 210 proceeds with the input eye image 212 for further processing steps.

[0077] Returning to the present example, on ascertaining the quality of the input eye image 212 to be acceptable, the analysis engine 210 determines the field of view of the input eye image 212. On determining the field of view of the input eye image 212 as one of disc centered and macula centered, the analysis engine 210 accepts the input eye image 212 and proceeds to further processing steps. On the other hand, on determining the field of view of the input eye image 212 as non-macula centered and nondisc centered, the analysis engine 210 displays the warning prompt on the display device of the system 202 to direct the user to re-capture the input eye image 212 or obtain another input eye image from the repository.

[0078] It may be noted that, while processing the input eye image 212 to detect presence of referable DR, it is important to have either a macula centered or a disc centered view of image because most of the changes in retinal structure in response to DR are identified or captured in disc or macula region only. Therefore, in order to efficiently and correctly process the input eye image 212, the image needs to be one of macula centered and disc centered view. In another example, before proceeding further, theinput eye image 212 may be processed by the analysis engine 210 to crop the borders.

[0079] At block 406, input eye image is pre-processed to obtain a processed image. For example, as described in conjunction with FIG. 1 , the first detection model 114 is trained based on a processed image while the second detection model 116 is trained based on an unprocessed image. Therefore, processed image is used for operating the first detection model 114 and unprocessed image is used for operating the second detection model 116. In an example, once the quality and field of view is determined to be acceptable, the input eye image 212 is pre-processed by the analysis engine 210 to subtract the background from the input eye image 212. Therefore, the processed image is the one in which the background of the image is subtracted. On the other hand, the unprocessed image is the one in which the background of the image is not subtracted. In another example, in case of obtaining input eye image 212 from the data repository, the analysis engine 210 obtains a pair of images including the processed image and unprocessed and there is no need to process the input eye image 212 to obtain the processed image.

[0080] At block 408, no processing is done in this block. Just the input eye image 212 is transferred as unprocessed image to block 410 along with the processed image obtained at block 406.

[0081] At block 410, a processed information from the processed image and an unprocessed information from the unprocessed image are derived. For example, the analysis engine 210 derives a processed information, such as processed information 214 from the processed image and an unprocessed information, such as unprocessed information 216 from the unprocessed image using any feature extraction technique. In an example, the processed information 214 and the unprocessed information 216 includes a plurality of eye characteristics having corresponding attribute values. Examples of eye characteristics include, but may not be limited to, size, location, and color of different types of lesions present on the retina ofthe patient’s eye. Examples of such eye characteristics include, but may not be limited to, size, location, and color of different types of lesions present in the retinal image, wherein the types of lesions comprises one of microaneurysms, retinal hemorrhages, soft exudates, hard exudates, intraretinal microvascular anomalies, neovascularization, fibrous proliferation, and preretinal and vitreous hemorrhage, tractional retinal detachment or combination thereof.

[0082] At block 412, a confidence score is generated by operating a first detection model based on the processed information. For example, the analysis engine 210 generate a confidence score, such as confidence score 218 by operating the first detection model 114 based on the processed information 214. In an example, the processed information is derived from the processed image of the patient’s eye. For example, to generate the confidence score 218, the analysis engine 210 processes the attribute values included in the processed information 214 based on the first detection model 114. As a result, a weight for each of the plurality of eye characteristics is assigned to generate a first weighted information. Thereafter, based on the first weighted information, the analysis engine 210 generate the confidence score 218 for the processed image using trained first detection model 114.

[0083] At block 414, a reliance score is generated by operating a second detection model based on the processed information. For example, the analysis engine 210 generate a reliance score, such as reliance score 220 by operating the second detection model 116 based on the unprocessed information 216. In an example, the unprocessed information 216 is derived from the unprocessed image of the patient’s eye. For example, to generate the reliance score 220, the analysis engine 210 processes the attribute values included in the unprocessed information 216 based on the second detection model 116. As a result, a weight for each of the plurality of eye characteristics is assigned to generate a second weighted information. Thereafter, based on the second weighted information, the analysis engine210 generate the reliance score 220 for the unprocessed image using trained second detection model 1 16.

[0084] At block 416, confidence score and the reliance score are compared individually with a predefined value. For example, the analysis engine 210 compares the confidence score 218 and the reliance score 220 individually with a predefined value. The comparison is performed to ascertain whether one of confidence score 218 and the reliance score 220 is greater than the predefined value. In an example, the value of confidence score 218 and the reliance score 220 varies between 0 and 1. It may be noted that, values mentioned for confidence score and the reliance score 220 is exemplary and any other value may be used without deviating from the scope of the present subject matter.

[0085] At block 418, input eye image is designated as either referable DR or non-referable based on the comparison. For example, on ascertaining one of the scores being greater than the predefined value, the analysis engine 210 designate the input eye image as referable DR. On the other hand, on ascertaining none of the scores being greater than the predefined value, the analysis engine 210 further analyze the confidence score 218 and the reliance score 220 based on a linear combination. For example, the analysis engine 210 designate the input eye image 212 as referable DR if following condition satisfies: x*(reliance score) + y*(confidence score) -z > 0

[0086] Further, on ascertaining none of the scores being greater than the predefined value, the analysis engine designate the input eye image 212 as non-referable DR if following condition satisfies: x*(reliance score) + y*(confidence score) -z <= 0 wherein x, y, and z represents any real number value except ‘0’.

[0087] Therefore, in such a way the DR positiveness is identified by the system 202 to indicate its presence in the input eye image 212 to the patient or the person operating the system 202. Based on the result, the patient or the medical person appropriately take further steps relating to treatment of DR so that it may not progress further.

[0088] Although examples for the present disclosure have been described in language specific to structural features and / or methods, it is to be understood that the appended claims are not necessarily limited to the specific features or methods described. Rather, the specific features and methods are disclosed and explained as examples of the present disclosure.

Claims

AMENDED CLAIMS received by the International Bureau on February 4, 2025 (04.02.2025) l / We Claim:

1. A system (202) comprising: a processor; and an analysis engine (210) coupled to the processor, wherein the analysis engine (210) is to: generate a confidence score (218), indicating probability of presence of diabetic retinopathy in a processed image of a patient’s eye, by operating a first detection model (114) based on a processed information (214), wherein the processed information (214) indicates information related to structural configuration of lesions forming in a retina of the patient’s eye and is extracted from the processed image (214) of the patient’s eye; generate a reliance score (220), indicating probability of presence of diabetic retinopathy in an unprocessed image of the patient’s eye, by operating a second detection model (116) based on an unprocessed information (216), wherein the unprocessed information (216) indicates information related to structural configuration of lesions forming in a retina of the patient’s eye and is extracted from the unprocessed image of the patient’s eye; compare the confidence score (218) and the reliance score (220) individually with a predefined value to ascertain whether one of confidence score (218) and the reliance score (220) is greater than the predefined value; and on ascertaining one of the scores being greater than the predefined value, designate the patient’s eye as diabetic retinopathy positive eye.

2. The system (202) as claimed in claim 1 , wherein the analysis engine (210) is to: obtain an input eye image (212) from a repository, wherein the repository comprises a plurality of input eye images for diagnosing diabetic retinopathy; or capture, via a camera module, an input eye image (212) of the patient’s eye, wherein the patient is under screening for diagnosing diabetic retinopathy; execute a quality assessment test on the input eye image (212) to ascertain its quality; on ascertaining the quality of the retinal image as acceptable, determine field of view of the retinal image; andon determining the field of view as one of disc centered and macula centered, process the input eye image (212) to crop black border to obtain a cropped input eye image (212); and execute a pre-processing step on the cropped input eye image (212) to obtain the processed image.

3. The system (202) as claimed in claim 1 and 2, wherein the processed image is pre-processed using the pre-processing step and the unprocessed image is not pre- processed, wherein the analysis engine (210) to obtain the processed image using the pre-processing step is to: process the cropped input eye image (212) to subtract background color to obtain the processed image.

4. The system (202) as claimed in claim 2, on ascertaining the quality of the retinal image as non-acceptable or on determining the field of view is not one of disc centered and macula centered, the analysis engine (210) is to: generate a warning prompt to be displayed on a display device to the patient indicating need to recapture the input eye image (212).

5. The system (202) as claimed in claim 1 , wherein on ascertaining none of the scores being greater than the predefined value, the analysis engine (210) is to: designate the patient’s eye as diabetic retinopathy positive when following condition satisfies: x*(reliance score) + y*(confidence score) -z > 0 wherein x, y, and z represents any real number value except ‘0’.

6. The system (202) as claimed in claim 1 , wherein on ascertaining none of the scores being greater than the predefined value, the analysis engine (210) is to: designate the patient’s eye as diabetic retinopathy negative when following condition satisfies: x*(reliance score) + y*(confidence score) -z <= 0 wherein x, y, and z represents any real number value except ‘0’.

7. The system (202) as claimed in claim 1 , wherein while operating the first detection model (114) based on the processed information (214) to generate the confidence score(218) and operating the second detection model (116) based on the unprocessed information (216) to generate the reliance score (220), the analysis engine (210) is to: derive the processed information (214) from the processed image and the unprocessed information (216) from the unprocessed image using a feature extraction technique, wherein the processed information (214) and the unprocessed information (216) comprises an attribute value corresponding to a plurality of eye characteristics; process the attribute values comprised in the processed information (214) based on the first detection model (114) to assign a weight for each of the plurality of eye characteristics to generate a first weighted information; process the attribute values comprised in the unprocessed information (216) based on the second detection model (116) to assign a weight for each of the plurality of eye characteristics to generate a second weighted information; generate, based on the first weighted information, the confidence score (218) and the reliance score (220) based on the second weighted information.

8. The system (202) as claimed in claim 7, wherein the plurality of eye characteristics comprises size, location, and color of different types of lesions present in the retinal image, wherein the types of lesions comprises one of microaneurysms, retinal hemorrhages, soft exudates, hard exudates, intraretinal microvascular anomalies, neovascularization, fibrous proliferation, and preretinal and vitreous hemorrhage, fractional retinal detachment, or combination thereof.

9. A method (300) comprising: obtaining (302) a training information (108) comprising a processed training image and an unprocessed training image of a patient’s eye; deriving (304) a processed training information corresponding to the processed training image and an unprocessed training information corresponding to the unprocessed training image; training (306) a first detection model (114) based on the processed training information; training (308) a second detection model (116) based on the unprocessed training information;wherein the first detection model (114) and the second detection model (116), when trained, are to generate a confidence score (218) and a reliance score (220), respectively, corresponding to an input eye image (212) of a patient’s eye.

10. The method (300) as claimed in claim 9, wherein the processed training information and the unprocessed training information comprises a plurality of training eye characteristics with corresponding training attribute value, wherein the plurality of training eye characteristics comprises size, location, and color of different types of lesions present in the retinal image, wherein the types of lesions comprises one of microaneurysms, retinal hemorrhages, soft exudates, hard exudates, intraretinal microvascular anomalies, neovascularization, fibrous proliferation, and preretinal and vitreous hemorrhage, fractional retinal detachment or combination thereof..

11. The method (300) as claimed in claim 9 to 10, wherein the training a first detection model (114) based on the processed training information comprises: classifying the training attribute values corresponding to the processed training information as a first processed training eye characteristic, wherein the first processed training eye characteristic belongs to a first range of attribute values of the training eye characteristics; and associating the first processed training eye characteristic with a first confidence score based on the training attribute values corresponding to the processed training information.

12. The method (300) as claimed in claim 9 to 10, wherein the training a second detection model (116) based on the unprocessed training information comprises: classifying the training attribute values corresponding to the unprocessed training information as a first unprocessed training eye characteristic, wherein the first unprocessed training eye characteristic belongs to a first range of attribute values of the training eye characteristics; and associating the first unprocessed training eye characteristic with a first reliance score based on the training attribute values corresponding to the unprocessed training information.

13. The method (300) as claimed in claim 9, further comprising:obtaining (310) a subsequent processed training image comprised in a subsequent training information; ascertaining whether training attribute values corresponding to the subsequent processed training image lies in the range of training attribute values of first processed training eye characteristic; on ascertaining training attribute values does not correspond to first processed training eye characteristics, classifying the training attribute values corresponding to the subsequent processed training image as a second processed training eye characteristic in the first detection model (114) which belongs to a second range of attribute values; and associating second processed training eye characteristic with a second confidence score based on the training attribute values.

14. The method (300) as claimed in claim 9 and 13, further comprising: obtaining a subsequent unprocessed training image comprised in the subsequent training information; ascertaining whether training attribute values corresponding to the subsequent unprocessed training image lies in the range of training attribute values of first unprocessed training eye characteristics; on ascertaining training attribute values does not correspond to first unprocessed training eye characteristics, classifying the training attribute values corresponding to the subsequent unprocessed training image as a second unprocessed training eye characteristics in the second detection model (116) which belongs to a second range of attribute values; and associating second unprocessed training eye characteristic with a second reliance score based on the training attribute values.

15. The method (300) as claimed in claim 9, wherein the processed training image is obtained by performing a number of pre-processing steps, wherein the pre-processing steps comprises: processing a training retinal image to crop black border to obtain a cropped image; and processing the cropped image to subtract background color to obtain the processed training image.STATEMENT UNDER ARTICLE 19Without prejudice to the objection raised in the International Searching Report (ISR), the Applicant has amended the independent claim and dependent claims by correcting dependency in claim 13 and has made other clarificatory amendments along with inclusion of reference numerals to bring further clarity. The Applicant submits that the scope of the amended claim set is within the scope of the originally filed subject matter, and no new subject matter is added.The present subject matter provides approaches for detecting presence of diabetic retinopathy (DR) in a retinal image of patient’s eye. To detect the presence of referable DR, a first detection model and a second detection model are trained to identify a confidence score and a reliance score, respectively. In an example, an analysis engine of a system may generate a confidence score by operating a first detection model. The confidence score may indicate the probability of diabetic retinopathy being present in a processed image of the patient’s eye. In an example, the first detection model may operate based on processed information extracted from the processed image which may relate to the structural configuration of lesions forming in the retina of the patient’s eye.Additionally, the analysis engine may generate a reliance score by operating a second detection model. The reliance score may indicate the probability of diabetic retinopathy being present in an unprocessed image of thepatient's eye. The second detection model may operate based on unprocessed information extracted from the unprocessed image. This unprocessed information may also relate to the structural configuration of retinal lesions.Thereafter, both scores, i.e., the confidence score and the reliance score, are individually compared with a predefined value. In an example, the comparison may be performed to determine whether either the confidence score or the reliance score exceeds the predefined value. Based on the comparison, if the analysis engine ascertains that either the confidence score or the reliance score is greater than the predefined value, it may designate the patient’s eye as diabetic retinopathy positive.By utilizing both processed and unprocessed images, the system may potentially improve the accuracy and robustness of diabetic retinopathy detection. Further, relying on scores to identify the presence of diabetic retinopathy, rather than determining specific image features, may offer several advantages. By generating confidence and reliance scores, the analysis engine may quantify the likelihood of diabetic retinopathy presence on a continuous scale, which may provide more granular information than binary feature-based detection methods. This score-based method may also be more adaptable to variations in image quality, patient demographics, and different manifestations of the disease. Additionally, using scores may simplify the decision-making process for healthcare providers, offering a clear, quantitative measure of risk.One or more of these above discussed claimed features are not described in the cited references D1 -D2. The Applicant therefore submits that the amended claims 1 -15 involve an inventive step with respect to the prior art documents D1 -D2 cited in the ISR.

Images