CN-121985899-A - Systems and methods for cervical anomaly detection and classification

Abstract

Cervical inspection systems and methods are based on the use of a dedicated image capture device that is used in conjunction with a machine learning model as a diagnostic tool for use at the inspection site. By evaluating acetic acid and lugol's iodine images using a pair of combined models, a good balance of sensitivity and specificity can be achieved. Furthermore, the combined model showed higher accuracy when distinguishing between normal/LSIL and cancer (AUC 0.98) and HSIL and cancer (AUC 0.89 for combination-1 and AUC0.88 for combination-2).

Inventors

• Ankita shastri
• Rao jeramati
• Soham Patak

Assignees

• 纳瑟维有限公司

Dates

Publication Date

20260505

Application Date

20241030

Priority Date

20231031

Claims (19)

1. 1. A system for detecting and grading cervical abnormalities, comprising: a dedicated image capture system configured to capture an image of the cervix at a predefined point in the examination procedure in response to the guiding procedure; A computing device coupled to the dedicated image capture system and including a set of guidance instructions for prompting a user to complete a defined set of steps to perform a cervical examination, and An image evaluation and assessment module for receiving as input an image captured by the dedicated image capture system, the image evaluation and assessment module comprising a processor component, a memory component, and one or more trained neural network models for evaluating the received image and providing as output a report comprising ranking the image based on a set of categories used in training of the neural network model.
2. 2. The system of claim 1, wherein the computing device includes a graphical user interface for displaying the various steps of the guidance instruction set, the guidance instruction set including prompts for initiating image collection.
3. 3. The system of claim 1, wherein the guidance instruction set includes a step of instructing to apply acetic acid and prompting image capture for a predetermined period of time after application.
4. 4. A system according to claim 3, wherein the prompt for image capture is arranged to occur within a time frame of 60-120 seconds after application of acetic acid.
5. 5. The system of claim 1, wherein the guidance instruction set includes a step of instructing application of lugol's iodine and prompting image capture after application of lugol's iodine.
6. 6. The system of claim 5, wherein the prompting of image capture is performed after confirmation that lugol's iodine has adequately stained the cervix.
7. 7. The system of claim 6, wherein whether cervical staining is adequate is confirmed by automated image evaluation.
8. 8. The system of claim 1, wherein the image evaluation and assessment module comprises a first trained neural network model for assessing acetic acid images and a second trained neural network model for assessing images related to lugol's iodine application.
9. 9. The system of claim 8, wherein the image evaluation and assessment module further comprises an additional trained neural network model for assessing the captured images associated with additional steps of the guided procedure of the included hint definition.
10. 10. The system of claim 1, wherein the image evaluation and assessment module comprises one or more training neural network models for assessing images captured under different lighting conditions.
11. 11. The system of claim 1, wherein the computing device is configured to provide a prompt to a system user to manually perform image capture.
12. 12. The system of claim 1, wherein the computing device is configured to include an automatic prompt for image capture by the system itself.
13. 13. A method of providing cervical anomaly detection and classification during a cervical examination, comprising: Providing one or more trained neural network models for examining images of the cervix and providing a diagnostic assessment of the presence of pre-cancerous or cancerous conditions; Installing one or more training models within a local computing system of the inspection location; Performing a guided cervical inspection procedure, the guided cervical inspection procedure comprising image capturing at a predetermined point in time during an inspection; submitting the cervical exam captured image to an image assessment and evaluation module; Screening the submitted images to delete defective, poor quality images; selecting an appropriate training model from the one or more training models; Submitting the screened images to a selected training model; evaluating the screened image according to the selected training model, and An output report including a diagnostic grade is generated based on the assessment of the submitted image by the model.
14. 14. The method of claim 13, wherein the one or more trained neural network models are configured to perform multistage classification of at least three diagnostic levels from normal to IL to invasive cancer.
15. 15. The method of claim 13, wherein after screening, images captured during the cervical examination are further pre-processed to create images that are similar in form in one or more image features selected from the group consisting of focus, brightness, blood volume, specular reflection, a clipping image closer to the cervix, and normalized color space.
16. 16. The method of claim 13, wherein the image selected for submission to the selected model is manually selected by the user.
17. 17. The method of claim 13, wherein the generated output report further comprises a superposition over the input image to display an image region used by the selected model when reaching a diagnostic level.
18. 18. The method of claim 13, wherein the step of providing one or more trained neural network models comprises the steps of: collecting a plurality of existing cervical images of known histopathology; confirming that each existing image was collected during a period of time consistent with the prompts of the inspection program, deleting any images that failed in the confirmation; Retaining a set of known histopathologically confirmed images for later testing and defining the remaining images as a set of input images; Selecting a neural network model; Submitting the set of input images to a selected neural network for training and validation in an iterative process to generate a final training model set; Testing the final training model set with a set of retained histopathological confirmation images; Evaluating the performance of each training model during the test, and One model with the highest correct diagnostic grade of known histopathology is selected as the final version of the training model.
19. 19. The method of claim 18, wherein prior to submitting the set of training images to the selected neural network model, the method further comprises the steps of: sorting the set of input images with respect to each other according to an image quality parameter selected from the group consisting of focus, brightness, blood volume, and specular reflection; determining an image quality threshold according to the ranking to be applied to the input image for reasoning; filtering the set of input images according to the determined threshold; Preprocessing the input image using techniques such as clipping closer to the cervix, normalizing color space, and the like The input image is enhanced by applying one or more functions selected from the group consisting of random cropping, flipping, limited rotation, saturation, and brightness variation to the original input image while mitigating class imbalance.

Description

Systems and methods for cervical anomaly detection and classification Cross Reference to Related Applications The present application claims priority from U.S. provisional application No. 63/546601, filed on day 31, 10, 2023, which is incorporated herein by reference. Technical Field The present invention relates to improvements in conventional medical examination procedures for assessing cervical tissue abnormalities, and more particularly to a system and method for rapidly and accurately providing examination results based on the use of a dedicated image capture system and associated computer-aided image analysis. Background In many developing countries, screening for pre-cervical lesions and cancers involves the use of acetic acid and Lugol's iodine (Lugol's iodine) to help observe the extent of the cancer or its precursor, which is known in the art as visual acetic acid inspection (VIA) and Lugol's iodine visual inspection (VILI). Similar to VIA, in VILI, the procedure involves performing a vaginal speculum examination, during which a lugol's iodine solution is applied to the cervix. The cervix is then typically visually inspected to determine the particular color change that occurs due to glycogen content on the cervix. Eventually, the results appear positive or negative for potential pre-cancerous lesions or cancers. Administration involves immediate treatment, or biopsy to confirm a pre-cancerous lesion, if there is a concern that the patient may not return to follow-up. VILI is an attractive method because it is simple and easy to learn, low in cost, and the test results are immediate, reducing subsequent potential revisits. Although the application of VIA and acetic acid are used in combination to evaluate the cervix, the outcome of VIA depends on the consistency of its performance. VIA relies on assessment of the cervix just 60 seconds after administration of acetic acid, after which staining begins to resolve and the abnormalities seen at the peak begin to resolve. However, the use of acetic acid and its assessment are often diverse throughout the world, and the presence of mucus, blood, etc. prevents complete exposure of the cervix to acetic acid, and some practices wait less than or more than 60 seconds, thus affecting accurate time. This results in greater variability in assessing the cervix using acetic acid and makes lugol's iodine assessment more attractive because it is simple and easy to implement and does not require a specific time to assess the cervix after application. Researchers studying the efficacy of VILI report that the high sensitivity of this test is at the expense of moderate specificity, which may lead to over-treatment of a single visit. The accuracy of the test also depends on the experience of the healthcare provider performing the examination and his or her ability to evaluate the characteristics of the lesions. In order to find a highly accurate objective technique for detection of pre-cervical lesions/cancers, some researchers have trained image recognition models based on artificial intelligence. Since these models were developed for screening environments (rather than triage environments), they were developed for binary diagnosis of the presence or absence of cervical abnormalities. The accuracy estimates in these models are subject to validation bias (i.e., tend to be exaggerated) because no biopsies are obtained when cytology or colposcopy is normal. Furthermore, consistency or repeatability of the inspection conditions is not utilized, and in many cases, image captures of different quality are found. For example, in many cases, the cervix is taken during the examination using a cell phone, and the digital photograph of the cervix taken by the cell phone is used as the source of the digital image. Obviously, it is difficult to continuously take high quality cervical photographs with a hand-held mobile phone. Thus, there remains a need to improve the accuracy and specificity of cervical visual inspection while utilizing artificial intelligence based image recognition capabilities. Furthermore, increasing the ability to make finer diagnostic decisions (e.g., being able to distinguish between low-grade lesions and high-grade lesions) would increase the value of such simple tests, enabling them to be used more frequently and in environments where the level of care is non-optimal. Disclosure of Invention The present invention relates to improvements in cervical examination procedures that are particularly useful in "screening-as-you-go" and triage protocols, and more particularly to systems and methods that utilize neural network models trained to discover patterns of pre-cervical cancer/cancer images at defined points in time during the examination (e.g., after administration of acetic acid and/or after staining with lugol's iodine). According to the present invention, the use of a dedicated image capture system to collect multiple sets of images at specific defined po