2023 International Conference on Electrical, Computer and Communication Engineering (ECCE) Detection of Colon Cancer Using Inception V3 and Ensembled CNN Model Ishrat Jahan Swarna Department of Computer Science and Engineering Rajshahi University of Engineering and Technology Rajshahi, Bangladesh ishratswarna887@gmail.com Emrana Kabir Hashi Department of Computer Science and Engineering Rajshahi University of Engineering and Technology Rajshahi, Bangladesh emranakabir@gmail.com Abstract—Colon cancer is one of the most prevalent types of cancer. Early diagnosis of colon cancer can lead to an increased chance of successful treatment with less cost. To speed up this process deep learning can provide very useful and effective approaches. In this thesis work, two types of models were developed to classify colon cells from image data - one is the transfer learning model where a deep network Inception V3 is used as the pre-trained model and the other one is an Ensembled model which combines predictions of three simple sequential CNN models. To develop these models, 10k images were used from the LC25000 dataset and a very small Warwick-QU dataset having only 165 images was used to provide new data for retraining and testing purposes. Both models achieved a high result for the first dataset with 99.4% and 99.95% accuracy respectively, where Inception V3 showed 94.545% accuracy on new data from Warwick-QU after retraining and Ensembled model showed 78.182% accuracy. This approach can be used in research in the field of early and effective detection of colon cancer with a larger amount of varying images and more preprocessing methods to reduce overfitting and to make the model perform well in various types of images. Index Terms—colon cancer, Inception V3, ensemble, deep learning, image classification. I. I NTRODUCTION Colon cancer, also known as colorectal cancer is a type of cancer where cells of colons in our large intestine get affected. In the United States in 2022, there will be 106,180 new instances of colon cancer and 44,850 new cases of rectal cancer, as reported by the American Cancer Society. Colorectal cancer is the second most prevalent cause of cancer-caused fatalities in the United States and is in third place overall. It is predicted to result in 52,580 fatalities in 2022 [1]. Certain col- orectal tumors may exist without showing any symptoms. To identify issues early, it is crucial to conduct routine colorectal screenings (examinations). The most effective screening test is a colonoscopy. Fecal occult blood tests, fecal DNA tests, flexible sigmoidoscopy, barium enema, and CT colonography are further screening methods (virtual colonoscopy). Your risk factors, particularly a genetic link of colon and rectal cancers, will determine when such screening tests start and at what age [2]. In medicinal practice, pathologists visually examine changes in cells/tissue underneath a microscope and define the level of colon cancer. However, expert pathologists frequently dispute network classifications. So we can conclude that a histopathological appraisal’s performance by expert pathologists alone is not sufficient [3]. Presently AI has a significant impact on disease diagnosis. A supervised Deep Learning model can make decisions based on results from previous experiences. Although extensive data/information is needed to get an accurate result, these techniques have created an effective alternative to all conventional methods. So a well-performed image classification model will introduce an alternate path in the diagnosis of colon cancer which will lead to the efficient management of colon cancer. Therefore to make early detection of colon cancer more effective, in our work, we tried to develop a deep learning model that detects cancer from image data of colon cells by classifying them into 2 classes –– (i) Colon Adenocarcinoma (cancerous) and (ii) Colon Benign Tissue (non-cancerous). Fig. 1: Sample images of colon cancerous cells(on left) and colon non-cancerous cells(on right) We made an effort to introduce two distinct models: one that uses transfer learning and the other that uses ensemble. Three CNN models were chosen for ensembling to make the process simple. In order to expose the models to a variety of data, they were trained on two different datasets. 979-8-3503-4536-0/23/$31.00 ©2023 IEEE 2023 International Conference on Electrical, Computer and Communication Engineering (ECCE) | 979-8-3503-4536-0/23/$31.00 ©2023 IEEE | DOI: 10.1109/ECCE57851.2023.10101654 Authorized licensed use limited to: Charles Darwin University. Downloaded on April 21,2023 at 05:48:51 UTC from IEEE Xplore. Restrictions apply.