A Review on Medical Image Analysis Using Deep Learning ^†

Abstract

The objective of the medical image analysis is to increase the effectiveness of the diagnosis options. The Coevolution Neural Network (CNN) is the predominant neural network architecture used in Deep Learning (DL) for medical image analysis. Recently, various innovative technics of DL such as different activation functions, optimization technics, and loss functions have enhanced the performance of CNNs. The Deep Learning CNN (DL-CNN) assists as valuable tool to assist radiologist in diagnosis and improves efficiency and accuracy. Numerous DL-CNN methods have been published to analyze medical images. This paper compiles the performance metrics of DL-CNN, as presented by various authors. This paper reviews the image analysis of six different diseases, viz., lung cancer, colorectal cancer, liver cancer, stomach cancer, breast cancer, and brain tumors.

1. Introduction

DL is subfield of machine learning that focuses on training artificial neural networks with multiple layers to learn and make predictions from available data set. DL models are built using artificial neural network, which are inspired by the structure and functioning of the human brain. These networks consist of interconnected neurons organized into layers. The connections between neurons have associated weights that are adjusted during the training process to optimize network performance. DL systems have been shown to be effective for analyzing medical images [1]. Some familiar image modalities used to detect the abnormalities in human organs are MRI (Magnetic Resonance Imaging) [2], X-ray, CT (Computed Tomography) scan [3], Ultrasound, and Mammography [4]. Enhancing diagnostic capacity is the goal of medical image processing [5]. Image enhancement, segmentation, feature extraction, and classification are the main components of medical image analysis [3,6,7,8]. This paper reviews the image analyses of six different diseases, viz., lung cancer, colorectal cancer, liver cancer, stomach cancer, breast cancer, and brain tumors. In medical image analysis, the data set is collected from authorized websites or real-time data from reputed medical institutes. A total of 70 percent of the data set is used for training and 30 percent of the data set is used for testing the designed model. The designed model could be based on machine learning (ML) or DL algorithms. The performance of the DL algorithm is superior when compared to ML if a very large data set is available. In order to enhance the performance of the DL further, preprocessing of the images is conducted before they are subjected to the DL network, as discussed in Section 2.

2. Architecture of DL-CNN

The architecture of the DL-CNN is portrayed as follows:

Convolution layers: The DL-CNN consists of several convolutional layers. In this layer, convolution operations are performed on the input images with various filter kernels.

Activation function: after each convolutional layer, Rectified Linear Unit (ReLU), an activation function is applied.

Pooling layers: These layers are used to reduce the spatial dimensions of the feature maps while retaining the important information. The common pooling operation is max. pooling, which selects the maximum value within a pooling region and discards the rest.

Fully connected layers: integrate all information from previous layer, crucial for learning complex patterns and making decisions in neural networks.

Output layer: this layer represents the final layer of the network, providing the predicted probabilities for each class.

3. Review on the Literature of DL-CNN on Medical Image Analysis

In this review paper, the performance of the DL-CNN, as published by various authors, is investigated. The performance of the DL-CNN is presented in

Table 1. Various abnormalities and corresponding techniques.

Reference	Disease	Modality	Performance
[9,10,11,12]	Lung Cancer	CT scan	AUC—0.64 to 0.93 Accuracy—75–99.5%
[13,14,15,16,17,18]	Colorectal Cancer	Endoscopy [13,17]; Whole slide images (WSIs) [14,18]; slide images [15]; histopathology images [16]	AUC—0.96 to 0.97 Accuracy—81.2–95%
[19,20,21]	Liver Cancer	CT scan [19]; MRI [20]; CEUS data [21]	Accuracy 89% to 94.5%
[22,23,24]	Stomach cancer	Pathological images [22]; tongue images [23]; endoscopic images [24]	Accuracy—61.1–99.6%
[25,26,27,28,29]	Breast cancer	Infrared images [25,26,28]; mammogram images [27,29]	Accuracy—90% to 97.53%
[30,31,32,33,34,35]	Brain tumor	MRI	Accuracy is 92.3% to 97.5%

for various abnormalities.

4. Conclusions

This review paper compares the performance of various DL-CNNs on life-threatening cancers. Significant accuracy was achieved in all papers as the proposed DL-CNN has been trained on a very large data set. The authors used various types of kernels and activation functions to enhance image quality and extract features. This review helps researchers by providing a benchmark to develop DL algorithms with more accuracy using the most efficient programming languages.