Automatic Identification Of Lung And Colon Cancer Cells Based On CNN And YOLO Algorithms

Can deep learning help in providing reliable early cancer detection? Expert systems based on Deep Learning can be used to make an early diagnosis,offering a second opinion or a preliminary diagnosis,hence reducing the mortality rate of cancer patients.Even though cancer is considered one of the most serious health problems,its cause still remains unknown,making it a major problem worldwide.When access to specialized health services is not affordable or easy to obtain and regular medical check-ups are not frequent,disease detection is prone to happen in the advanced stages where the symptoms have become noticeable and severe.That’s why creating and developing new technologies have become crucial keys in reducing the mortality rate of patients.This study proposes a model based on Depthwise Convolutional Networks for the classification of Lung and Colon Cancerous Cells,tested on 3 different datasets using ADAM,ADAGRAD,“SGD+Nesterov”,and“SGD+MOMENTUM” optimizers for the main model,and a secondary expert model who will verify the predictions in cases where the calculation doesn’t reach a minimum threshold,thus achieving high and competitive results.The outcome presented by the Classification model was analysed,and processed to create masks that will isolate the nuclei of these cells by applying OTSU algorithm,where the samples will be transferred from the original RGB colour space into HSV colour space,finding the optimal threshold values and computed by the Otsu algorithm,translating these im ages into monochrome pictures.The nuclei-isolated cancerous samples were labelled and went through a set of Data Augmentation algorithms for expanding this new dataset and improving its quality,to be then then fed into a YOLOV5 model in order to detect if any Carcinomatous Pattern are present inside of the cancerous cell elements.In the end,this paper verified the effectiveness of combining the previously mentioned techniques with the presented models from a subjective and objective evaluation through comparative experiments,thus offering a method for improving the cancer detection accuracy.