A Design of End-to-End CNN Architecture for Colon Cancer Detection and Visualization Using ContinuousData Cleaning Method

Abstract Colon cancer threatens the quality of life and causes deaths of mankind worldwide today. Convolutional Neural Networks(CNNs) are recognized as a robust and prevalent technique in deep learning methodologies for biomedical signals, includinghistopathology imaging and processing. As a diagnostic tool, medical CNNs gradually facilitate the decision-making process inclinical detection. However, there is still potential improvement for an accurate model to achieve millionth-level accuracy ofconfirmation. Beyond that, overfitting misleads performance of a model to properly recognize pattern especially to unseendata. Further, visualized result could facilitate interpretability of a detected results from clinical end.. This research aims tointroduce an end-to-end colon cancer detection using CNN with continuous parametric data cleaning method and interpretativevisualization approaches.Based on the results of the alternative settings, the raw dataset used in Experiment 1 caused overfitting during the trainingprocess, while Experiments 2 and 3 overcame the overfitting after applying Gaussian 95% and 99% rules based on RGBcolor distribution result. Further, there were 10 extensive experiments conducted based on settings of Experiment 3 to ensurethe reliability of the results, which showed 0.99996 averaged accuracy and 0.00044 averaged loss with 1,00 precision, recall,and F1-score. Lastly, a K-Mean clustering and 3D intensity visualized sample imagefor interpretability improvement in clinical diagnostic process. The comparison result indicated the proposedmodel gained superior accuracy compared to the benchmark studies from the past 5 years. This research contributed to anend-to-end CNN model in an accurate colon cancer detection