Using CNN for Brain Tumor Diagnosis: An Overview

1.1 What Is CNN?

DL a subfield of AI has caught the attention in recent years owing to its remarkable capabilities of solving complex problems. It processes vast amounts of data through layers of algorithms to recognize patterns, and features and make decisions. CNN is quite well-known DL algorithms. Convnet is another name for CNN. CNNs are a specific category under neural network that are used in analyzing data that has a topology resembling a grid, such as pictures and time series data (1D and 2D). CNN's capacity to comprehend and interpret visual data, like photos or videos, like human vision, has made it extremely popular in a variety of fields, including health. They are made to resemble human brain to interpret, process, and use visual information to make decisions (Indolia et al., 2018; Qin et al., 2018; Wu, 2017).

1.2 Basic CNN Architecture

It consists of many layers (e.g., convolutional, pooling, and fully connected layers). They work collectively to extract features or patterns from the input/given data and make decisions (predictions) on the basis of those features. Figure 1 shows the basic architecture of CNN.

Convolutional layers are the basic structures of CNN. The network in a convolutional layer applies filters (also known as kernels) to input data. A filter is a number matrix used to extract specific features (Mrazova et al., 2012), like edges or textures, from given (input) data in a convolutional layer. These filters slide over the input image, performing mathematical operations (such as element-wise matrix multiplication and summation) between the filter and small (sub-matrix) regions of the input, which is termed as convolution operation. The process extracts various features from input data. Also, there are multiple convolutional layers present in a model so multiple filters are used in each of these layers to detect different features. Output of the layer is referred to as feature maps. Feature maps demonstrate the presence of distinctive features in given data after applying filters and performing convolution operations (Alzubaidi et al., 2021).

Figure 1.

Basic architecture of CNN

Pooling layers appear after convolutional layers and help to make feature maps smaller. Max, min, average, and global are four types of pooling operations. Max pooling: summary of features of a defined region is represented by the maximum value of that region. It is utilized when an image has a dark background since it will select brighter pixels. Figure 2 (a) represents the max pooling operation. Average pooling: features in a particular area are summarized by calculating average value of that defined region. Figure 2 (b) represents the average pooling operation. Min pooling: It represents features within an area by taking minimum (lowest) value found in that domain (region). Figure 2 (c) represents the min pooling operation. Global pooling: each channel (column) in feature map gets condensed into a single number (value). This number is determined by the type of pooling method used, like average pooling or maximum pooling. Figure 2 (d) represents a global pooling operation where c1, c2, c3, and c4 are column-wise channels. The concept of strides is discussed in the later section.

Figure 2.

Pooling operations: (a) max, (b) average, (c) min, and (d) global

In the fully connected (dense) layer), feature maps (2D) from previous layers are flattened into a single long vector (1D). Each neuron of these layers is linked to every neuron of flattened vector, thus allowing the network to understand complex relationships between features and predictions (output). This is where the final decision-making happens, such as identifying pieces in an image or classifying various data types.

CNN learns to recognize patterns and objects in images by repeatedly processing the data through these layers. They can identify various features which help them understand what is present in the image. The working of these layers is to be discussed in the later section (Bhatt et al., 2021; Nguyen et al., 2022).

CNN has revolutionized medical image analysis by successfully detecting complex details that may escape human observation. This capability helps in the early detection of potential health issues (discussed in the upcoming section). CNN shows great efficiency in rapidly processing large datasets. This efficiency is critical as it enables healthcare professionals to analyze numerous images rapidly, resulting in advanced, quick, and precise diagnoses. Ultimately, CNN empowers healthcare providers to deliver timely remedies and solutions, thereby enhancing patient outcomes and potentially saving lives.