Artificial Intelligence for Data-Driven Medical Diagnosis

Deepak Gupta, Utku Kose, Bao Le Nguyen, Siddhartha Bhattacharyya, Deepak Gupta, Utku Kose, Bao Le Nguyen, Siddhartha Bhattacharyya

• Publisher: De Gruyter
• Year: 2021
• eBook ISBN: 9783110668384
• Edition: 1
• Topic: Informatica
• Subtopic: Informatica generale

This book collects research works of data-driven medical diagnosis done via Artificial Intelligence based solutions, such as Machine Learning, Deep Learning and Intelligent Optimization. Physical devices powered with Artificial Intelligence are gaining importance in diagnosis and healthcare. Medical data from different sources can also be analyzed via Artificial Intelligence techniques for more effective results.

1 Performance of CNN for predicting cancerous lung nodules using LightGBM

Subrato Bharati

Ranada Prasad Shaha University, Narayanganj, Bangladesh

Institute of Information and Communication Technology, Bangladesh University of Engineering and Technology, 723, Polashi, 1205, Dhaka, Bangladesh

Prajoy Podder

Institute of Information and Communication Technology, Bangladesh University of Engineering and Technology, 723, Polashi, 1205, Dhaka, Bangladesh

Abstract

Lung cancer is a common type of cancer. The objective of this chapter is to predict lung cancer in a patient using chest computerized tomography scans. Convolutional neural network (CNN) has also been used in this chapter. CNN has been pretrained on ImageNet for the purpose of generating features from the dataset. Extracted features have been fed into the proposed classifiers in order to train a good classifier. A boosted tree LightGBM (light gradient boosting machine) has been used to perform image classification on the validation dataset. Receiver operating characteristics (ROC) curve and log loss have been evaluated on the training and validation dataset using LightGBM on top of two models VGG19 (Visual Geometry Group 19) and ResNet50 generated features. Log loss is less in the residual network on the validation dataset compared to VGG. ROC curves between VGG19 and ResNet50 architecture have been compared. This research work tries to emphasize the multiple instance nature of detecting the nodules in many scanned images. The area under the ROC curve value is comparatively larger in CNN compared to the recurrent neural network in the simulation result.

Keywords: VGG, ResNet, LightGBM, CNN, LSTM,

1.1 Introduction

Lung cancer is the second-highest leading cause of death worldwide. According to the report of World Health Organization, there were approximately 150,781 cancer patients in 2018 around Bangladesh. About 55.52% of them were males and 44.48% were females. According to the analysis of the International Agency for Research on Cancer, 12,374 (8.2%) people were affected by lung cancer. Therefore, lung cancer is common in Bangladesh. Recovery and survival rates can be improved if early detection of cancer is possible (Bharati et al., 2020a, Bharati et al., 2020c). At present, chest scans have been performed to diagnose lung cancer. A lung scan can generate a few dozen to a few hundred cross-sectional two-dimensional (2D) images of the chest for each patient. These images must be examined by the radiologist. The radiologist can notice the shape changes between neighboring slides and recognize the lung nodules. Then, he/she can decide whether the lung nodule is malignant (cancerous) or not (benign). A lung nodule is a small growth on the lungs. Normally, its size is less than 3 cm. If the size of the nodule is larger than 3 cm, then it is suspicious (Wei et al., 2015).

There are already many proposed machine learning methods available in order to identify the cancerous nodules. There are many methods where a two-step procedure is used. These steps are region generation and classification. The region generation system recognizes the desired regions that might carry pulmonary nodules. The classification system decides the possibility of cancerous nodules (Awais et al., 2015, Rushil et al., 2016).

To address limited data, time and resources, networks ideally take advantage of transfer learning. While lung cancer scanning seems to be highly specialized, there have been successful projects that use pretrained convolutional neural networks (CNNs) and fine-tune them to detect lung cancer nodules (Bharati et al., 2020d). For example, Ramaswamy et al. (2016) demonstrated the efficacy of transfer learning from general image classification. They used AlexNet and GoogLeNet pretrained on ImageNet, and by fine-tuning, they were able to achieve good scores on predicting whether a nodule was cancerous (Ramaswamy et al., 2016).

More pertinent to this project, the paper of Tripathy et al., (2018) showed that by using residual network (ResNet) (pretrained on ImageNet) as the feature generator, they were able to train a boosted tree that achieved reasonable results in predicting whether the complete set of a patient’s scans show any cancer (with no specific nodule detection).

These models cannot often take advantage of the three-dimensional (3D) information across sections (e.g., simply average acrosssections; Ramaswamy et al., 2016). This is the problem of these models. Hence, 3D CNNs have been used on chest scans (Ramaswamy et al., 2016). Ypsilantis et al. (2016) achieved good results with a model that added recurrent neural network (RNN) with a traditional 2D CNN. They used their network on patches that might contain cancer nodules (the size of the recurrence was always seven slices, so this created a sort of fixed-size voxel). Bharati et al. (2020a) proposed vanilla neural network, CNN, modified VGG (Visual Geometry Group) and capsule network for detecting several lung diseases from the lung X-ray images.

The purpose of our research is to find the probability that a given set of scans shows any cancer. We have taken the key part of transfer learning and added RNN long short-term memory (LSTM) features. RNN is normally an unsupervised learning model because RNN LSTM features will allow the transfer learning model to learn 3D information. RNN LSTM not only deals with the various instances of nature of lung nodule detection but also works with the variable number of slices per patient.

Full images have been used in our experiment. There are no patches. RNN has been used over all slice...

Table of contents

• Title Page
• Copyright
• Contents
• Dedication
• 1 Performance of CNN for predicting cancerous lung nodules using LightGBM
• 2 Deep learning-based cellular image analysis for intelligent medical diagnosis
• 3 Deep learning approaches in metastatic breast cancer detection
• 4 Machine learning: an ultimate solution for diagnosis and treatment of cancer
• 5 Artificial intelligence in medicine (AIM): machine learning in cancer diagnosis, prognosis and therapy
• 6 Diagnosis disease from medical databases using neural networks: a review
• 7 A novel neutrosophic approach-based filtering and Gaussian mixture modeling clustering for CT/MR images
• 8 Decentralized solutions for data collection and privacy in healthcare
• 9 Navigation from conventional to intelligent healthcare: adoption of Internet of health things for noncommunicable disease screening, diagnosis, monitoring and treatment in community settings
• 10 Automated gastric cancer detection and classification using machine learning
• 11 Artificial intelligence applications for medical diagnosis and production with 3D printing technologies
• 12 Detection of breast cancer using deep neural networks with transfer learning on histopathological images
• 13 A machine vision technique-based tongue diagnosis system in Ayurveda
• 14 Vine copula and artificial neural network models to analyze breast cancer data
• Index
• Computational Intelligence for Machine Learning and Healthcare Informatics Already published in the series