An Optimized Cancer Classification Approach Using Deep Learning

Abstract

Cancer, being a genetic disease, is engendered when the body’s cells grow uncontrollably and disseminate to other parts of the body. Early diagnosis and classification of cancer predominantly focus on recognizing symptomatic and indicatory patients as soon as practicable so that successful and well-timed treatment can be determined. Among many kinds of cancers: skin, breast, and brain cancer are the most common and deadliest ones. According to the statistics of 2020, more than 1.5 million new cases of melanoma have been reported which is one of the deadliest forms of skin cancer. An estimated number of 325,000 cases were diagnosed worldwide and 57,000 people died in the year 2020. A recent study conducted on March 2022 by scientists from IARC predicts that the number of cases of melanoma will increase by more than 50% from 2020 to 2040. Similarly, Breast cancer on the other hand is the second most commonly occurring cancer in women after lung cancer. Most frequent deaths in women are from breast cancer. The chance that a woman will die from breast cancer is about 1 in 39 (about 2.5%). In 2020, nearly 684,996 women (about half the population of Hawaii) died from breast cancer. As for brain cancer, the most common form of malignant brain cancer called glioblastoma is the deadliest human cancer. In 2020, around 251,329 people worldwide died from primary cancerous brain and CNS tumours. The death ratios due to these cancers call for an alarming situation. Delaying cancer care and holding up to proper treatment can expedite to lower the chance of survival, inducing greater complications associated with treatment and ex travagant costs of care. To classify such cancers, Artificial Intelligence has played a significant role. Various machine learning and deep learning models have been applied in this domain. However, there remain some areas of research that have been left unex plored. xv List of Tables For a deep learning model to perform better, it should be trained on colossal amounts of data. Most of the deep learning models that predict and classify different cancers are trained on a limited amount of data. Therefore, there is a need for diverse datasets. Furthermore, a comprehensive classification for detecting different kinds of cancer to the best of our knowledge is not available. Therefore, there is a need for an optimal classification approach that classifies different kinds of organ cancers simultaneously and further identifies their associated tumour type. This thesis proposes a consolidated dataset that has been prepared using different pub licly available datasets, which is then used for training different deep learning models. The proposed model achieves higher accuracy as compared to other state-of-the-art models.