Deep Learning for Automated Lung Nodule Detection in Low-Dose CT Scans

One of the biggest causes of cancer-related death globally is still lung cancer. Survival chances are greatly increased by early diagnosis of lung nodules with Low-Dose Computed Tomography (LDCT) screening. However, interpreting a hand LDCT scan is subjective, takes a lot of time, and is easily affected by differences between observers, especially when noise as well small tumours are present. It is the goal of this project to make a reliable as well as automated deep learning-based system for accurately identifying lung nodules in LDCT images. The study's goal is to create a 3-dimensional Convolutional Neural Network (3D-CNN)) structure that uses volumetric spatial data to better detect small lumps and handle noise. This is how the network will be built and tested to make sure it can reliably find lung nodules. We test the suggested system using the open-source LIDC-IDRI dataset. The data is enhanced, and a thorough preparation pipeline is implemented, which includes filtering out Gaussian noise, segmenting the lungs, and normalizing the intensity. The Adam optimizer & binary cross-entropy loss are used to train a 3D-CNN model using$64 \times 64 \times 64$voxel volumetric patches. The F1-score, AUCROC, sensitivity, specificity, and accuracy are the performance measures used. The suggested model has a great diagnostic performance, with an AUC of 0.94 as well as excellent sensitivity, especially for small nodules ($\geqslant 3 ~\text{mm}$). Qualitative study backs up the idea that robust detection works even when LDCT is noisy. The suggested system is a dependable computeraided detection tool that may help radiologists find lung cancer early, cut down on the amount of effort they have to do to get a diagnosis, and help them make better clinical decisions.