Implementation of Convolutional Neural Networks for Lung Cancer Detection from CT Scans

Abstract

Lung cancer is still the foremost common sort of cancer that slaughters individuals around the world, and early finding is key to raising survivor rates. Later advance in profound learning has appeared that Convolutional Neural Systems (CNNs) can be utilized in restorative imaging to rapidly and precisely analyze illnesses. The objective of this consider is to assist specialists make early and adjust analyze by looking into how CNNs can be utilized to discover lung cancer on computed tomography (CT) pictures. Our strategy incorporates a full workflow that begins with getting CT check information and altering it. We utilized freely available datasets with thousands of labeled pictures to create beyond any doubt that there was an rise to number of cases of cancer and cases that were not cancer. Normalization, boosting, and commotion diminishment were a few of the preprocessing steps utilized to progress picture clarity and model steadiness. Information improvement strategies, counting turning, scaling, and flipping, were utilized to form the preparing information more shifted. This cut down on overfitting and made the CNN show way better at generalization. The most important part of our method is designing and teaching a CNN system that can find lung cancer. We looked at a number of cutting-edge CNN designs, such as VGG, ResNet, and DenseNet, to find the best model for feature extraction and classification. We used transfer learning by fine-tuning models that had already been trained on our CT scan dataset. This took advantage of the models' ability to spot complex patterns and features in medical pictures. To get the best model performance, hyperparameter optimization was used to fine-tune the learning rate, batch size, and network depth. Our CNN model was trained and tested on a subset of the data. It was then put through a lot of tests on a separate test set to see how accurate, sensitive, specific, and high its F1-score was. The results showed that the model was very good at telling the difference between images of cancer and those that were not. It had high sensitivity and specificity.