1
Analysis of Medical Image Using a Multimodal Approach for Precise Cancer Detection

Soumen Santra1*, Mouparna De1, Dipankar Majumdar2 and Surajit Mandal3

1Department of MCA, Techno International New Town, Kolkata, West Bengal, India

2Department of CSE, RCC Institute of Information Technology, Kolkata, India

3Department of ECE, B.P. Poddar Institute of Management & Technology, Kolkata, India

Abstract

Cancer is one of the most serious medical conditions is cancer, and patients' conditions worsen every day. The early identification of cancer is essential for successful treatment and better patient outcomes. Fast-acting image processing technology should be used for medical diagnosis. To achieve this, we provide a unique method for cancer identification that uses the Gaussian Mixture Model (GMM) algorithm in combination with a machine learning technique that aids in correct analysis. This method involves resizing and altering the color of the image to determine the precise location of cancer. This method makes the operation considerably simpler because it provides a better understanding of the cancer picture. This method's main goal is to create a strong and trustworthy cancer detection system that will aid medical professionals in the early identification of cancer and has the potential to save many lives. This technique can act as a bridge between conventional diagnostic methods and cutting-edge technologies.

Keywords: Resizing, image preprocessing, gaussian mixture model (GMM) algorithm, conventional diagnostic methods, cutting-edge technology

1.1 Introduction

Cancer, a complex and common disease, poses a major challenge to public health worldwide. Cancer has become a serious global health problem because of its high mortality rate and insidious nature. Early and accurate cancer detection is key to improving patient outcomes and survival rates. Although effective, traditional cancer detection techniques, such as biopsy and medical imaging, are often invasive, expensive, and open to human interpretation, which can result in inconsistent diagnoses. Machine learning is a branch of artificial intelligence, has emerged as an effective strategy to improve the efficiency and accuracy of cancer detection in response to these difficulties [1]. Cancer, with its multifaceted etiology and complex pathway physiology, remains a formidable opponent. Its diversity among different types, subtypes, and individual cases requires precision and sophistication in diagnosis [2]. Traditional methods often fail to provide the precision and efficiency required for modern treatment [2, 3]. Cancer is characterized by uncontrolled proliferation of abnormal cells and their ability to invade neighboring tissues.

After viewing images from an X-ray, magnetic resonance imaging (MRI), or computed tomography (CT) scan, medical professionals present their ideas and make decisions [4]. These images are essential for identifying and describing anomalies and tumors, which makes them crucial for cancer surveillance and diagnosis. Several image-processing methods can be applied to this model to extract the characteristics of the impacted area of an image, which provides insight into their nature [5].

Digitally altered images are referred to as raw images. Raw photos are subjected to a variety of image-processing algorithms to identify their attributes [6]. In this study, we explored the incorporation of Gaussian mixture models (GMM) as a pioneering tool in cancer image analysis as per Figure 1.1. The GMM algorithm is a form of unsupervised learning that can be used to model complex data distributions in medical images. Deep learning, a branch of artificial intelligence, has emerged as a potent technique for feature extraction and identification of patterns, structures, and anomalies in pictures, complementing GMM [7, 8]. Machine learning integration enabled our diagnostic methodology to adapt, change, and improve over time. By integrating GMM, we aim to provide a more accurate and precise approach for classifying and characterizing cancer regions in these images. Typically, a medical image is of excellent quality.

Cancer detection through the utilization of cutting-edge technologies, such as image recognition and machine learning, can be challenging as it requires a huge, labeled training dataset, and there are several sites present on the Internet where a dataset can be gathered. This model collected different cancer datasets from Kaggle.com (https://data.mendeley.com/datasets/p2r42nm2ty/1).

Figure 1.1 The framework of the proposed medical image detection.

The use of radiological imaging, genetic information, electronic medical records, and patient demographics contribute to the development of a full understanding of the illness. With the use of machine learning models and the unification of these disparate data sources, we hope to obtain insightful knowledge that will enable earlier and more precise cancer diagnoses [9].

1.2 Methodology

1.2.1 Understanding the Objective and Scope

• Defining the goal of cancer image analysis, whether it is related to disease detection or medical image classification [10].
• Identification of a dataset of medical images for analysis.

1.2.2 Data Collection and Importation

• A dataset of cancer images relevant to our research and analysis was collected to ensure that the images were in a consistent format, such as JPG, JPEG, or PNG [11].
• Utilizing the 'glob' library to efficiently import medical (cancer) images from a specified directory, making it easier to process multiple images simultaneously.

1.2.3 Image Preprocessing

• Importing Libraries: Import the necessary libraries, including NumPy, cv2, and sklearn. mixture and glob.
• Load each cancer image using OpenCV to create image objects.
• Applying preprocessing techniques to standardize the images for analysis. The common preprocessing steps that we have included are [12]:
  • Resizing images to uniform dimensions.
  • Converting color spaces if necessary (e.g., from RGB to grayscale).
  • Applying image thresholding to enhance cancer features.
  • Removing noise or artifacts that might affect the analysis.

1.2.4 Feature Extraction

• Derivation of relevant features to be preprocessed for cancer imaging. These features include:
  • Shape Descriptors: Extract shape features to characterize cancer cell shapes [13].
  • Texture features: Texture features are extracted to characterize cancer image textures.

1.2.5 Segmentation Using Gaussian Mixture Model (GMM)

• Preprocess the images to enhance the features relevant to cancer image segmentation.
• Initialize a GMM using scikit-learn (sklearn. mixture) library to fit the GMM to the preprocessed image data, cluster the cancer and background pixels [14], and segment the cancer region based on the GMM clusters.

(1.1)

where:

• x represents the observed data point (a vector in the feature space).
• K is the number of Gaussian components in the mixture.
• pk represents the mixture coefficient or weight associated with the kth Gaussian component. It represents the probability of selecting the kth component.
• µk is the mean vector of the kth Gaussian component.
• Sk is the covariance matrix of the kth Gaussian component.
• N(x|µk, Sk) denotes the multivariate Gaussian distribution with mean µk and covariance Sk.

Different types of models are employed for cancer cell detection, each with its own balance between accuracy and efficiency. Convolutional neural networks (CNNs) [15], a type of deep learning model, are used in image classification tasks, such as cancer cell detection, because of their ability to learn intricate features from images. They typically offer high accuracy, often around 94%, but their efficiency can vary, with predictions taking approximately 0.5 s. Support Vector Machines (SVMs), which are classical machine learning models, also perform well in this domain, with an accuracy of approximately 88%. However, their efficiency tends to be lower, with predictions taking approximately 1.2 s. Random Forest, an ensemble learning method, achieves a high accuracy of approximately 91%, with moderate efficiency, required approximately 0.8 s for predictions. Logistic Regression, a simple linear model [16], offers moderate accuracy at around 85% and relatively better efficiency, with predictions taking around 0.6 s. Decision Trees, another non-parametric method, yield accuracy of about 87% with similar efficiency to Random Forest, taking around 0.7 s for predictions in Table 1.1. These models represent a spectrum of trade-offs between accuracy and efficiency, allowing tailored approaches based on specific requirements.

Table 1.1 Model accuracy and efficiency.

Model Accuracy (%) Efficiency (s) Convolutional Neural Network (CNN) 94 0.5 Support Vector Machine...