A method for detecting slit lamp image quality

Abstract

The application discloses a kind of slit lamp image quality detection methods.The method uses image quality filter to carry out image quality detection to the diffuse light image of real-time shooting, when detection is unqualified, image quality analysis is carried out to the diffuse light image by real-time guidance feedback system, and the operation guidance of rephotographing is generated according to image quality analysis result;When the detection result of the diffuse light image is qualified, whether the diffuse light image meets the shooting condition of slit light image and red light image is detected;If it is satisfied, slit light image shooting and red light image shooting are carried out, if it is not satisfied, shooting is ended and shooting result is output;Image quality analysis is carried out to the slit light image and red light image of unqualified by the real-time guidance feedback system, and the operation guidance of rephotographing is generated according to image quality analysis result.The technical scheme of the application improves the collection quality and efficiency of clinical slit lamp image.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image quality detection, in particular to a slit lamp image quality detection method. BACKGROUND

[0002] Slit lamp is the most commonly used optical instrument in ophthalmology clinic, which is applied to the diagnosis of anterior segment diseases such as cataract, corneal disease, pterygium, etc. Taking age-related cataract as an example, this disease is the most important eye disease causing blindness and visual impairment in the elderly. About 50% of people over 60 years old have different degrees of cataract, and 95% of people over 70 years old have cataract. If these anterior segment diseases can be detected early and timely intervention, the disease process can be effectively improved, the quality of life of patients can be improved, and the social disease burden can be reduced. Therefore, population screening based on slit lamp anterior segment images is crucial in the management of anterior segment diseases.

[0003] However, slit light photography sometimes results in poor imaging quality of the anterior segment due to defocusing, improper exposure, movement of the examinee, etc. during actual photography, making it difficult for doctors to interpret. At this time, the patient often needs to requeue for examination and return to the examining doctor for analysis, causing waste of manpower and material resources.

[0004] The following problems exist in the process of slit light photography:

[0005] (1) Whether the slit lamp image quality affects the interpretation of the picture requires professional ophthalmologists to make accurate judgments. Some images have a large blurred range, but do not affect the observation of the lesion, and should be considered as acceptable pictures in quality; some images have a small blurred range, but block the important structures of the anterior segment, and need to be judged as poor quality pictures. Non-ophthalmic disease professionals cannot make correct judgments.

[0006] (2) Slit lamp image quality affects the process of disease judgment by doctors. When the slit lamp image quality is poor, doctors have difficulty in making accurate assessments and need to rephotograph.

[0007] (3) The photography of slit lamp images is generally completed by professional technicians, and it takes a long time to accumulate experience for subsequent adjustment of different image quality problems. Defocusing, improper exposure, movement of the examinee, etc. during the process of slit lamp image photography can all lead to poor imaging of the anterior segment, but the corresponding adjustment methods are not the same. Moreover, if the examinee has turbidity in the anterior segment, the examinee needs to be referred to an ophthalmologist. When the photographer lacks experience, it often takes a long time to take a slit lamp image of qualified quality. SUMMARY

[0008] In order to overcome the above-mentioned deficiencies in the existing slit lamp image photography, the present application provides a slit lamp image quality detection method, which improves the collection quality and efficiency of clinical slit lamp images.

[0009] The embodiment of the present application provides a slit lamp image quality detection method, comprising the following steps:

[0010] The image quality of the real-time photographed diffuse light image is detected by using an image quality filter, when the detection result is unqualified, the image quality of the diffuse light image is analyzed by using a real-time guidance feedback system, and operation guidance for rephotographing is generated according to the image quality analysis result;

[0011] When the detection result of the diffuse light image is qualified, whether the diffuse light image meets the photographing condition of the slit light image is detected, if yes, the slit light image is photographed, and if not, the photographing is ended and a photographing result is output;

[0012] When the detection result of the diffuse light image is qualified, whether the diffuse light image meets the photographing condition of the red light reflection image is detected, if yes, the red light reflection image is photographed, and if not, the photographing is ended and a photographing result is output;

[0013] The image quality of the photographed slit light image and red light reflection image is detected by using an image quality filter, when the detection result is unqualified, the image quality of the slit light image and red light reflection image is analyzed by using the real-time guidance feedback system, and operation guidance for rephotographing is generated according to the image quality analysis result.

[0014] Further, the image quality analysis comprises image brightness analysis, image integrity analysis and image definition analysis.

[0015] Further, whether the diffuse light image meets the photographing condition of the slit light image is detected, specifically comprising: whether the pupil area of the diffuse light image is shielded is detected, if not, the slit light image is photographed.

[0016] Further, whether the diffuse light image meets the photographing condition of the red light reflection image is detected, specifically comprising: whether the pupil area of the diffuse light image is shielded and whether the mydriasis of the diffuse light image is qualified are detected, if the pupil area of the diffuse light image is not shielded and the mydriasis of the diffuse light image is qualified, the red light reflection image is photographed.

[0017] Further, when the real-time guidance feedback system detects that the number of rephotographing times exceeds 3, the photographing is stopped, and it is prompted that the photographing result is only for reference.

[0018] Further, the operation guidance specifically comprises: if the brightness is unqualified, please adjust the brightness of the background light, the brightness of the slit light or the brightness of the indoor light;

[0019] if the structure is uncompleted, please adjust the sitting posture, eye position or auxiliary repeated exposure photographing area;

[0020] Defocus, check and clean the lens, or adjust the slit beam angle.

[0021] Further, the image quality filter is specifically an Inception-ResNet network model, the Inception-ResNet network model is obtained by introducing a residual structure of ResNet in an Inception module, and comprises a Stem module, an Inception-resnet-A module, an Inception-resnet-B module and an Inception-resnet-C module.

[0022] Embodiments of the present application have the following beneficial effects:

[0023] The present application provides a slit lamp image quality detection method, which detects the image quality of a diffuse light image, a slit light image and a red light image by using an image quality filter, to determine whether the image quality is qualified, and further analyzes the unqualified image through a real-time guidance feedback system, and generates an operation guidance for re-shooting according to the analysis result, thereby improving the collection quality and efficiency of the clinical slit lamp image. BRIEF DESCRIPTION OF DRAWINGS

[0024] Figure 1 Fig. 1 is a flowchart of a slit lamp image quality detection method according to an embodiment of the present application;

[0025] Figure 2 Fig. 5 is a structural diagram of an Inception-ResNet network model according to an embodiment of the present application;

[0026] Figure 3 Fig. 6 is a structural diagram of a Stem module of the Inception-ResNet network model according to an embodiment of the present application;

[0027] Figure 4 Fig. 7 is a structural diagram of an Inception-ResNet-A module of the Inception-ResNet network model according to an embodiment of the present application;

[0028] Figure 5 Fig. 8 is a structural diagram of an Inception-ResNet-B module of the Inception-ResNet network model according to an embodiment of the present application;

[0029] Figure 6Fig. 1 is a structural schematic diagram of an I nception-ResNet-C module of an I nception-ResNet network model provided by an embodiment of the present application. DETAILED DESCRIPTION

[0030] The technical solutions in the present application will be clearly and completely described below with reference to the drawings in the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present application.

[0031] As shown in Fig. 1, an image quality detection method of a slit lamp provided by an embodiment of the present application comprises the following steps: Figure 1

[0032] Step S101: image quality detection of a real-time photographed diffuse light image is performed by using an image quality filter, when the detection result is unqualified, image quality analysis of the diffuse light image is performed by using a real-time guidance feedback system, and operation guidance for re-photographing is generated according to the image quality analysis result. The image quality analysis comprises image brightness analysis, image integrity analysis and image definition analysis, and the image quality analysis result comprises results of three dimensions of brightness, integrity and definition.

[0033] Step S102: when the detection result of the diffuse light image is qualified, whether the diffuse light image meets the photographing condition of a slit light image is detected; if yes, slit light image photographing is performed, and if no, photographing is ended and a photographing result is output.

[0034] As one of the embodiments, whether the diffuse light image meets the photographing condition of the slit light image is detected, specifically comprising: whether the pupil area of the diffuse light image is blocked is detected, and if no, slit light image photographing is performed.

[0035] Step S103: when the detection result of the diffuse light image is qualified, whether the diffuse light image meets the photographing condition of a red reflex light image is detected; if yes, red reflex light image photographing is performed, and if no, photographing is ended and a photographing result is output.

[0036] As one of the embodiments, whether the diffuse light image meets the photographing condition of the red reflex light image is detected, specifically comprising: whether the pupil area of the diffuse light image is blocked is detected, and whether the mydriasis of the diffuse light image is qualified is detected; if the pupil area of the diffuse light image is not blocked and the mydriasis of the diffuse light image is qualified, red reflex light image photographing is performed.

[0037] ​Step S104: image quality detection is performed on the photographed slit light image and red light image using an image quality filter; when the detection result is unqualified, image quality analysis is performed on the slit light image and red light image by the real-time guidance feedback system, and operation guidance for rephotographing is generated according to the image quality analysis result; when the detection result is qualified, the photographing is completed.

[0038] When the real-time guidance feedback system detects that the number of rephotographing times exceeds 3, the photographing is stopped, and it is prompted that the photographing result is only for reference.

[0039] As one of the embodiments, the operation guidance specifically includes: when the brightness is unqualified, please adjust the brightness of the background light, the brightness of the slit light or the brightness of the indoor light;

[0040] When the structure is incomplete, please adjust the sitting posture, eye position or auxiliary repeated exposure of the photographing area;

[0041] When the definition is unqualified, please adjust the focusing, check and clean the lens or adjust the angle of the slit light band.

[0042] As one of the embodiments, the image quality filter specifically adopts an Inception-ResNet network model for image quality detection, and the Inception-ResNet network model introduces the residual structure of ResNet in the Inception module. The model adopts a deep learning algorithm, and according to the quality evaluation aspects such as definition, brightness and completeness, the input slit lamp image is classified respectively, and a binary classification evaluation result, i.e. qualified or unqualified, is output. In the training process of the model, more than 30000 photos taken by different models of digital slit lamps from multiple central medical institutions are used as the training set. Figure 2As shown, the Inception-ResNet network model includes a Stem module, an Inception-resnet-A module, an Inception-resnet-B module and an Inception-resnet-C module. The image detection process of the Inception-ResNet network model includes: inputting a 512x512x3 image from an input layer, the input layer being linked to a Stem main module, performing image feature extraction on the input image in the Stem module by asymmetric convolution and pooling operations to generate a first feature map, taking the first feature map as the input of the Inception-ResNet-A module, sequentially passing through the stacking of five Inception-ResNet-A modules and the scaling block of Reduct-ion-A to obtain a second feature map; taking the second feature map as the input of the Inception-ResNet-B module, sequentially passing through the stacking of ten Inception-ResNet-B modules and the scaling block of Reduct-ion-B to generate a third feature map, taking the third feature map as the input of the Inception-ResNet-C module, and sequentially passing through the stacking of five Inception-ResNet-C modules and the scaling of the Reduct-ion-C module to obtain a fourth feature map. Finally, performing global average pooling on the fourth feature map, connecting the tensor obtained by the global average pooling to a fully connected layer with a size of 1024 and passing it through a Re lu activation function, mapping to a sample label space, then passing through a Dropout layer with a retention rate of 0.5 to enhance the generalization degree of the model, at this time the output layer size is 1, passing the feature tensor through a Si gmoi d activation function for classification to obtain the final detection result.

[0043] The Inception-ResNet network model selects binary cross entropy (Binary Cross Entropy) as the loss function.

[0044] As Figure 3As shown, the Stem module is mainly a shallow feature extraction module that constitutes the network, adopts asymmetric convolution and pooling to constitute, and after three times of 3x3 convolution processing of different properties, the input 512x512x3 image enters a 3x3 max pooling layer with a step of 2, and then after a 1x1 convolution processing and a 3x3 convolution processing, the extracted features pass through four parallel convolution branches. The first branch will pass through average pooling and a 1x1 convolution; the second branch will pass through a 1x1 convolution; the third branch will pass through a 1x1 convolution and a 5x5 convolution respectively; the fourth branch will pass through a 1x1 convolution and two 3x3 convolutions. After filtering the outputs of the four branches, they are stacked and transmitted to the subsequent module (i.e., the Inception-ResNet-A module).

[0045] As shown in Figure 4 The Inception-ResNet-A module has a total of 5, and the structure of Inception-ResNet-A is composed of four branches, a total of three convolution channels, and uses a 1x1 convolution kernel to reduce the input dimension. One branch is processed by a 1x1 convolution, with an output of 32 channels; the second branch is processed by a 1x1 convolution and a 3x3 convolution, with an output of 32 channels; the third branch is processed by a 1x1 convolution and two 3x3 convolutions, with an output of 64 channels. The three convolution channels extract different feature data for feature stacking, and the original feature map is added for residual stacking and input into the ReLU activation function to obtain higher-level feature mapping.

[0046] As shown in Figure 5 The Inception-ResNet-B module has a total of 10, and its structure is composed of three branches, two of which are convolution channels. One is processed by a 1x1 convolution to obtain a 192-channel output; the second is an asymmetric structure convolution channel, which is first processed by a 1x1 convolution, and then the 7x7 convolution kernel is decomposed into a 1x7 and a 7x1 small convolution kernel, so that the parameters are greatly reduced, and a 192-channel convolution output is obtained. The feature results of the two channels are processed by a 1x1 convolution operation, and are stacked into the ReLU activation function through residual connection and original input features for mapping.

[0047] As shown in Figure 6As shown, the Inception-ResNet-C module has 5 in total, the structure of Inception-ResNet-C and Inception-ResNet-B has similar structure, and the structure block is two convolution channels and three channel branches. The first branch is processed by a 1x1 convolution with an output channel of 192; the second branch is processed by a 1x1 convolution with an output channel of 192 and asymmetric convolution processing of 1x3 and 3x1, two convolutions are stacked through a 1x1 convolution operation and residual connection with the original input branch, and finally mapped through a ReLU activation function.

[0048] As one of the embodiments, a first Inception-ResNet network model, a second Inception-ResNet network model and a third Inception-ResNet network model are established, the first Inception-ResNet network model is used for image quality detection of the diffuse light image, the second Inception-ResNet network model is used for image quality detection of the slit light image, and the third Inception-ResNet network model is used for image quality detection of the red light image. The first sample set composed of diffuse light images is used as the training set of the first Inception-ResNet network model, and the first Inception-ResNet network model is trained to convergence; the second sample set composed of slit light images is used as the training set of the second Inception-ResNet network model, and the second Inception-ResNet network model is trained to convergence; the third sample set composed of red light images is used as the training set of the third Inception-ResNet network model, and the third Inception-ResNet network model is trained to convergence.

[0049] As one of the embodiments, the real-time guidance feedback system analyzes the image quality of the diffuse light image, the slit light image and the red light image by establishing a plurality of Inception-ResNet network models respectively, and outputs the image quality analysis result. The image quality analysis result includes the results of three dimensions of brightness, integrity and clarity.

[0050] The present application has the following advantages:

[0051] 1. The system covers three common slit lamp shooting modes of diffuse light, slit light and red reflex light, and can specifically identify various quality defects encountered in clinical practice, and can adapt to clinical application scenes more comprehensively.

[0052] 2. The system forms a standardized process capable of providing real-time feedback according to clinical application logic, and in actual application, can provide real-time shooting guidance for the shooter, especially for the shooter with insufficient experience, thereby promoting the collection of high-quality slit lamp images and the early diagnosis and treatment of eye disease patients.

[0053] The above is the preferred embodiment of the present application, and it should be pointed out that for ordinary skilled persons in the art, several improvements and refinements can be made without departing from the principles of the present application, and these improvements and refinements are also considered within the protection scope of the present application.

[0054] Those skilled in the art can understand that all or part of the processes in the above embodiments can be completed by a computer program instructing related hardware, and the program can be stored in a computer readable storage medium. When the program is executed, it can include the processes of the above embodiments. The storage medium can be a magnetic disc, an optical disc, a read-only memory (ROM) or a random access memory (RAM), etc.

Claims

1. A method of slit lamp image quality detection, characterized in that, The method comprises the following steps: The real-time captured diffuse light image is detected by an image quality filter, and when the detection result is unqualified, the real-time guiding feedback system is used to analyze the image quality of the diffuse light image, and the operation guidance for re-capturing is generated according to the image quality analysis result; When the detection result of the diffuse light image is qualified, it is detected whether the diffuse light image meets the capturing condition of the slit light image, and if yes, the slit light image is captured, and if not, the capturing is ended and the capturing result is outputted; When the detection result of the diffuse light image is qualified, it is detected whether the diffuse light image meets the capturing condition of the red light image, and if yes, the red light image is captured, and if not, the capturing is ended and the capturing result is outputted; The captured slit light image and red light image are detected by an image quality filter, and when the detection result is unqualified, the real-time guiding feedback system is used to analyze the image quality of the slit light image and red light image, and the operation guidance for re-capturing is generated according to the image quality analysis result.

2. The method of claim 1, wherein, The image quality analysis comprises image brightness analysis, image integrity analysis and image definition analysis.

3. The method of claim 2, wherein, The detection of whether the diffuse light image meets the capturing condition of the slit light image specifically comprises detecting whether the pupil area of the diffuse light image is blocked, and if not, the slit light image is captured.

4. The method of claim 3, wherein, The detection of whether the diffuse light image meets the capturing condition of the red light image specifically comprises detecting whether the pupil area of the diffuse light image is blocked and detecting whether the mydriasis of the diffuse light image is qualified, and if the pupil area of the diffuse light image is not blocked and the mydriasis of the diffuse light image is qualified, the red light image is captured.

5. The method of claim 4, wherein, When the real-time guiding feedback system detects that the number of re-capturing times exceeds 3, the capturing is stopped, and it is prompted that the capturing result is only for reference.

6. The method of claim 5, wherein, The operation guidance specifically comprises: brightness unqualified, please adjust the brightness of the background light, slit light or indoor light; Incomplete structure, please adjust the sitting posture, eye position or auxiliary repeated exposure of the capturing area; Definition unqualified, please adjust the focus, check and clean the lens or adjust the slit light band angle.

7. The method of claim 6, wherein the step of detecting the image quality of the slit lamp image comprises: The image quality filter is specifically an Inception-ResNet network model, the Inception-ResNet network model is obtained by introducing the residual structure of ResNet into the Inception module, and comprises a Stem module, an Inception-resnet-A module, an Inception-resnet-B module and an Inception-resnet-C module.

Images