WO2022153100A1

WO2022153100A1 - A method for detecting breast cancer using artificial neural network

Info

Publication number: WO2022153100A1
Application number: PCT/IB2021/055971
Authority: WO
Inventors: Kamal Upreti; Binu Kuriakose Vargis; Nishant Kumar; Mohammad Haider Syed; Sanjay Sharma; Meena Agrawal; Karamjeet Kaur; Kamal Kant Sharma
Original assignee: Kamal Upreti; Binu Kuriakose Vargis; Nishant Kumar; Mohammad Haider Syed; Sanjay Sharma; Meena Agrawal; Karamjeet Kaur; Kamal Kant Sharma
Priority date: 2021-04-26
Filing date: 2021-07-02
Publication date: 2022-07-21

Abstract

The present invention pertains to detection of breast cancer in the early stages using artificial neural network (ANN). The invention uses digital mammographic images (DMIs) obtained from public databases (101), eliminates the image artifacts like background, noise (102), removes the pectoral region in the binary image (103) and extracts the fundamental features (105) in the DMI using digital image processing (DIP) and Artificial intelligence (AI) techniques to in order to create biomarker for breast cancer diagnosis. With the help of this information, a generalized regression artificial neural network is trained (106) to detect breast lesions or tumor.

Description

A METHOD FOR DETECTING BREAST CANCER USING ARTIFICIAL NEURAL NETWORK

TECHNICAL FIELD

[0001] The present disclosure relates to a detection method for breast cancer in early stages and separating malignant and benign tumors with high accuracy using artificial neural network.

BACKGROUND

[0002] Cancer falls in the category of diseases characterised by uncontrolled cell proliferation, and is medically described as a malignant neoplasm. Cancer occurs as cells differentiate and multiply uncontrollably, resulting in malignant tumours that infiltrate surrounding body sections. Via the lymphatic or blood systems, cancer can spread to all areas of the body. Tumors are classified as malignant or benign, which helps doctors determine if they are cancerous. Benign tumours are an abnormal development that seldom causes death; however, certain benign tumours may raise the risk of cancer. Malignant tumours, on the other hand, are more severe, and early detection aids in clinical progress. As a result, cancer prediction and detection will improve treatment odds while lowering the typically high costs of surgical treatments for patients.

[0003] Breast cancer (BC) is the most often diagnosed cancer in women worldwide, as well as the leading cause of death. With the exception of skin cancer, BC is the second most prevalent cancer in women. Furthermore, as opposed to other cancers, BC has a very high mortality rate. BC, like most cancers, begins with an unregulated outgrowth and proliferation of a portion of the breast tissue, which is classified as benign or malignant based on the potential for damage. The diagnosis of illness based on different examinations conducted on the patient is a major class of problems in medical research. The most significant considerations of diagnosis are the assessment of data obtained from patients and

SUBSTITUTE SHEET (RULE 26) expert decisions. One of the main issues of medicine is determining the proper diagnosis of BC.

[0004] PET, MRI, CT scan, X-ray, ultrasound, photoacoustic imaging, tomography, diffuse optical tomography, elastography, electrical impedance tomography, opto-acoustic imaging, ophthalmology, mammography, and other approaches are some of the most popular methods used for breast cancer diagnosis (BCD). However, Manual image classification is a difficult and time-consuming process that is particularly vulnerable to interobserver uncertainty and human mistakes, resulting in extraordinarily low vital results and significantly raising the workload of radiologists due to their scarcity. Furthermore, medical care costs related to imaging are rapidly increasing.

[0005] Bioimaging quantification is a new method in the area of radiology that is becoming more widely used in hospitals. It contains useful knowledge that isn't visible to the naked eye in traditional radiological reading. It entails extracting quantitative information from images, primarily of high resolution, in order to facilitate a clinical assessment. Bioimagen markers allow the characterization and analysis of various diseases using various types of data, such as genetic, histological, clinical imaging, and so on. This biomarker can be used to diagnose anomalies in detection such as genetic mutations that cause certain disorders to identify patients with specific diseases.

[0006] CAD/CADx devices are one of the many main research subjects in diagnostic radiology and medical imaging at the moment. Mammography can be manipulated with CAD systems to highlight characteristics that would otherwise be difficult to see. CAD is now a better tool for primary cancer diagnosis in computed tomography, X-ray, MRI, or mammogram images. The CAD method acts as a bridge between the input images and the radiologist. While the CAD production is not considered a final product, it is used as a point of reference for further research in the relevant area. Medical doctors may use the CAD technique to identify diseases more quickly while reducing test time and expense, as well as preventing needless biopsy procedures. However, CAD systems not only provide for

SUBSTITUTE SHEET (RULE 26) improved representation of mammograms, but they also allow for the preselection of those regions of interests (ROIs) for subsequent study by the radiologist using various digital image processing (DIP), information discovery from data (KDD), and artificial intelligence (Al) techniques such as artificial neural networks (ANN).

[0007] European Patent no. EP1668156B9 relates to a diagnostic method that entails evaluating the extent of expression of a breast cancer-associated gene that distinguishes breast cancer cells from healthy cells. The diagnostic method entails evaluating the expression level of a breast cancer-associated gene that distinguishes between breast cancer cells, DCIS cells, and invasive ductal carcinoma (IDC) cells. The invention discloses an invitro approach for diagnosing invasive ductal carcinoma (IDC) or a predisposition to develop IDC in a subject, which includes evaluating the level of expression of a breast cancer-associated gene in a patient-derived biological sample.

[0008] Therefore, the breast cancer diagnosis methods disclosed in the prior art are time-consuming and are prone to human errors. As a result, the need for development of a Breast cancer diagnosis (BCD) device that is both computationally efficient and accurate in terms of detection is the need of the hour. The present invention overcomes the above mentioned problems by using advanced Digital Image Processing (DIP) to analyse and improve new imaging biomarkers by quantitative mammography, and to use that knowledge to develop technologies focused on advanced Al techniques, with the goal of detecting breast cancer in its early stages to aid detection and prioritisation of high-risk patients.

OBJECTS OF THE INVENTION

[0009] It is an object of the present disclosure to provide a method for detecting cancer in early stages to aid diagnosis and prioritisation of high-risk patients.

[0010] It is another object of the present invention to train and test a generalized regression artificial neural network for classification of malignant and benign tumors.

SUBSTITUTE SHEET (RULE 26) [0011] It is another object of the present invention to provide an efficient method that diagnoses the cancerous cells without human involvement with high accuracies.

SUMMARY

[0012] The present invention relates to detecting breast cancer using Artificial Neural Networks. The technique is broken down into two stages. The first uses advanced Digital image processing (DIP) techniques to remove image features from Digital mammographic image (DMI) in order to develop a Breast cancer diagnosis (BCD) biomarker. With this knowledge, a GRANN is trained in the second level. The Computer Aided Diagnosis (CAD) method disclosed is divided into two stages: the first identifies abnormal regions with high sensitivity, then presents the findings to the radiologist with the aim of reducing false positives. This method is initiated by a preprocessing algorithm based on advanced DIP techniques, which is designed to minimise image noise and increase image quality, and then it executes a segmentation process of various Region of interests (ROIs)/Breast abnormalities, which is designed to detect high suspicion of certain cancer signs. To detect breast lesions using a GRANN, a procedure based on advanced DIP and Al techniques is being used to extract fundamental features in Digital image mammography (DIM).

[0013] In an aspect of the present disclosure relates to using various public mammography databases to create and validate digital image processing algorithms capable of selecting ROIs from mammograms and extracting image features for training a generic regression artificial neural network (GRANN). Image objects such as background, noise, and image labels are removed from all Digital mammographic images (DMIs). A typical threshold is used in the DMIs to establish a breast area and other regions with labels and marks on mammography. The ROI obtained during the segmentation stage is used to construct a binary mask. The ROI's features are extracted as the next stage in the operation of standard CAD systems. The method of inferring and quantifying the parameters that describe the object under consideration is referred to as feature extraction. These features aid in determining the type of tumour found in a mammogram image in BCD. The features

SUBSTITUTE SHEET (RULE 26) chosen have a significant impact on classification accuracy, classification time, the amount of examples required for learning, and classification expense. These image features are used to train a neural network to identify benign and malignant BC for use in BDC decisionmaking.

BRIEF DESCRIPTION OF THE DRAWINGS

[0014] The accompanying drawings are included to provide a further understanding of the present disclosure, and are incorporated in and constitute a part of this specification. The drawings illustrate exemplary embodiments of the present disclosure and, together with the description, serve to explain the principles of the present disclosure.

[0015] FIG. 1 illustrates a flow diagram of the method for detecting breast cancer using artificial neural network.

DETAILED DESCRIPTION

[0016] Aspects of the present disclosure relate to an effective method for detecting cancerous cells without the need for human intervention and with high accuracy. Image processing method is used in the present invention to create imaging biomarkers based on mammography analysis and artificial intelligence technologies, with the aim of detecting breast cancer in its early stages to aid detection and prioritisation of high-risk patients.

[0017] The generation and evaluation of CAD/CADx systems is done with mammograms from research trials or public databases. As a result, various public mammography databases are used in the first stage to build and validate automated image processing algorithms capable of selecting ROIs from mammograms and extracting image features used to train a GRANN capable of diagnosing BC as a radiology help. A mammogram image can be thought of as a reflection of X-ray radiation density reflecting breast tissue. When a white area appears on a mammogram chart, it indicates a high tissue

SUBSTITUTE SHEET (RULE 26) density, which can be called abnormal. This is a risk factor for Breast Cancer (BC) in a patient. ROI is a term used to describe a breast abnormality.

[0018] A method for detecting breast cancer using artificial neural network, the said method comprises the steps of: extracting digitial mammographic images (101) from public mammography databases; eliminating background, noise and image labels (102) from the digital mammographic image to create a logical image; removing the pectoral region (103) in the logical image to obtain a surplus white region representing the region of interest (ROI)/breast abnormality; segmentating the ROI (104), by creating logical image with a very high binarization threshold and creating a binary mask using the ROI; analyzing the ROI by extracting features (105) such as shape, texture, size, border, and other tissue parameters of the ROI to create a biomarker for breast cancer detection; training a generalized regression artificial neural network (106) using the extracted image features to classify benign and malignant breast cancer.

[0019] In an aspect of the present invention, a public database was used to obtain lateral mammographic images. A lesion can be seen in any of the chosen images, and can be classified as benign or malignant. Background, noise, and image labels are removed from all digital mammographic images. In the digital mammographic images, a standard threshold is used to establish a breast region and other mammography regions with labels and artifacts. After generating the logical image using the built automatic programming tool, all small regions (less than 10,000 pixels) are removed to delete the regions deemed unnecessary in Digital mammographic images (DMIs). The logical image (mask) is then used instead of the original image to produce a breast image free of artifacts and labels. A typical role of digital image processing is converting a greyscale image to a digital or logical image. Calculating the threshold value for generating rational representations can be done in a variety of ways. The threshold in the present invention is determined by transforming the values of nonzero pixels to one. The grey tones are transformed to white level to create a logical image that includes the ROI and pectoral muscle. The white area attached to the binary image's

SUBSTITUTE SHEET (RULE 26) boundary is removed in order to exclude the pectoral region in the logical image. The ROI identified in the mammography image is thus represented by the surplus white field. The segmentation procedure begins after the image has been cleaned.

[0020] A binary or logical image with a very high binarization threshold, where low grey levels become white, is generated for ROI segmentation. This method takes into account the majority of the image's grey pixels in order to avoid losing a large number of pixels to ROI. After that, the mammographic image's white logical area attached to the edge is removed. The ROI obtained in the segmentation stage is then used to construct a binary mask. We'll get the ROI in shades of grey by using the mask and combining it with the full picture in shades of grey.

[0021] The feature extraction of the RIO is the next step in the operation of standard CAD systems. The method of inferring and quantifying the parameters that describe the object under examination is known as feature extraction. The study of the ROI is aided by feature extraction. The shape, texture, size, boundary, and other tissue parameters that can help in the identification and identification of a cancer risk factor can all be quantified. Shape, intensity, and texture features are collected in order to construct a biomarker for BCD using a CADx system that utilises Al technology in the present invention.

[0022] Both image attributes from the database's Digital Image Mammography (DIM) are collected and used to create a biomarker for training an ANN. The RGB and gray-level optical images are in JPEG format, with a depth of 24 and 8 bits per pixel and a resolution of 3328 x 4084, respectively. The red remarked segment on the RGB mammograms is used by a radiologist to delimit the identified anomaly. To obtain the ROI, the segmentation process uses the red segment indicated in the RGB mediolateral oblique view mammograms. Both red pixels in the graphic, as well as pixels outside the original red field, are removed from the RGB mammograms. Finally, the ROI used for feature estimation is calculated using the gray-level mammogram's remaining pixels.

SUBSTITUTE SHEET (RULE 26) [0023] The measure of physical parameters visualised in a segmented region of an image is the feature extraction process. The aim of feature extraction is to find a mathematical way to express valuable image details in a compact form that can be used to solve a computational problem. These features aid in determining the type of tumour found in a mammogram image in BCD. On a mammography, the variations in mass between benign and malignant breast defects can be differentiated based on their shape, textures, and image intensity.

[0024] A generalized regression artificial neural network (GRANN) is used to separate malignant and benign tumours on DIM for automated classification of BC. GRANN is a probabilistic neural network (PNN) (PNN). GRANN is a one-step-only learning neural network architecture that can solve any problem involving function approximation. Finding a surface in a multidimensional space that gives the best match to the training data is the learning process. It simply saves training data during the training phase and uses it for predictions later. In operation, this neural net is very useful for making predictions and comparing machine results. There are no training parameters in the GRANN architecture; instead, a smoothing factor is added after the network has been trained. Using a data set of mammograms extracted from a public database, a GRANN is trained. An advanced computing tool specially built for this purpose calculates a list of image descriptors for each mammogram. These image features are used to train a neural net to distinguish benign and malignant BC for BDC decision-making. True positive (TP), true negative (TN), false positive (FN), and false negative (FN) are the four judgments that describe these measurements. When malignant instances are correctly expected, TP decision is made.

SUBSTITUTE SHEET (RULE 26)

Claims

We Claim:

1. A method for detecting breast cancer using artificial neural network, the said method comprises the steps of: extracting digitial mammographic images (101) from public mammography databases; eliminating background, noise and image labels (102) from the digital mammographic image to create a logical image; removing the pectoral region (103) in the logical image to obtain a surplus white region representing the region of interest (ROI)/breast abnormality; segmentating the ROI (104), by creating a logical image with a very high binarization threshold and creating a binary mask using the ROI; analyzing the ROI by extracting features (105) such as shape, texture, size, border, and other tissue parameters of the ROI to create a biomarker for breast cancer detection; training a generalized regression artificial neural network (GRANN) (106) using the extracted image features to classify benign and malignant breast cancer.

2. The method for detecting breast cancer using artificial neural network as claimed in claim 1, wherein the threshold is calculated by converting nonzero pixel’s values to one.

3. The method for detecting breast cancer using artificial neural network as claimed in claim 1 , wherein the pectoral region in the logical image is removed by eliminating the white region connected to the border of the logical image.

4. The method for detecting breast cancer using artificial neural network as claimed in claim 1, wherein to obtain the ROI in an RGB mammogram, all red pixels in the

SUBSTITUTE SHEET (RULE 26) image, as well as pixels outside the initial red field, are removed, and the remaining pixels in the grey level mammogram are used to obtain the ROI. The method for detecting breast cancer using artificial neural network as claimed in claim 1, wherein the said method examines breast lesions to detect and classify various stages of breast cancer.

SUBSTITUTE SHEET (RULE 26)