Lung And Colon Cancer Detection Using Weighted Average Ensemble Transfer Learning

Cancer is a deadly disease that is caused by metabolic abnormalities or a confluence of inherited diseases. Lung and colon cancer are considered as two of the most widely spread causes of disability and death in humans today. In the process of determining the best course of therapy for patients, the histological diagnosis of these tumors is the most important aspect. Diagnosis of the cancer in an early stage before spreading even more in the body will reduce the risk of death greatly on either front. Through utilizing the machine learning and deep learning models, this type of cancer diagnosis can be sped, providing researchers a cost-effective way to analyze a larger number of patients in much less time. In this study, we propose an ensemble transfer learning model to rapidly diagnose lung and colon cancer. Through utilizing multiple transfer learning models and ensemble them for a better performance. The proposed model uses the lung and colon histology (LC25000) dataset, Our models has an accuracy for each of, MobileNet V1, Inception V3, and VGG16 98.32%, 98%, and 96.93% for lung and colon cancer detection, respectively, while our ensemble model has an accuracy of 99.44%. This study’s findings indicate that our proposed method outperforms existing models therefor it could be used in clinics to assist medical personnel with the detection of lung and colon cancer.

I. Introduction

Cancer is a group of diseases characterized by random cell mutations in the human body which results the formation of abnormal cells; when cells reach a certain age or level of damage, they die and are replaced by new cells [1]; these cells divide in an uncontrolled way during their creation as well as spreading throughout the organs. Prediction is mostly dependent on clinical information, such as determining the type of cancer, molecular profile, tumor grades, etc. In recent years, more types of data have been made available to better determine the condition of the disease [2]. Biopsies, PET scans, CT scans, MRI scans, and ultrasonography have been utilized extensively for early cancer detection, monitoring, and follow-up after treatment [3]. Lung cancer is responsible for 18.4% of all cancer-related deaths worldwide, whereas colon cancer is responsible for 9.2% [4]. Based on the study of Koich Kurishima et al. Between April 2009 and July 2016, they examined all the pathology and medical reports of all lung cancer patients reports that were treated in the Respiratory Medicine Departments of all their four specialized hospitals. During the study period, out of 3,102 a percentage of 17 (0.54%) patients were diagnosed with synchronous colon cancer with the initial lung cancer. Although this does not happen often, but a late diagnosis can lead to the spread of cancer cells which is fairly common across the two organs [5]. This study focuses on automating the diagnoses of lung and colon cancers from histological scans. Medical experts often use histopathological images which aids the diagnosing process and play a crucial role in assessing a patient's survival odds. Examining histopathological images was traditionally a time-consuming process for diagnosing cancer by medical specialists. Nonetheless, with today's technology tools, this process can be automated which can lead to spending less time and effort [6]. for long-term survival and cure early detection and conventional therapy are considered to be alternatives, intensified surveillance for persons with synchronous lung and colon cancers is reasonable to consider, more genetic and epidemiological study is required to discover a probable relationship between these two malignancies [7].