Lung Nodule Detection Using Machine Learning

Abstract

Lung cancer remains to be one of the most critical types cancer, with a lifetime prevalence of just 18%. Efficient computer-aided diagnostic systems are required to diagnose lung cancer before time for better treatment planning. The variety of lung nodules and their visual similarity with surrounding regions make their detection difficult. Traditional image processing and machine learning methods cannot usually handle all types of nodules with a single method. In this study, we propose efficient end-to-end segmentation algorithms with an improved feature learning mechanism based on densely connected dilated convolutions. We applied dense feature extraction and incorporated multi-dilated context learning by using dilated convolutions at different rates for better nodule segmentation. In the first algorithm, lung ROIs are extracted from the CT scans using k-mean clustering and morphological operators to reduce the model’s search space instead of using full CT scan images or nodule patches. These ROIs are then used by our proposed architecture for nodule segmentation and efficiently handle different types of lung nodules. The efficiency of the first algorithm is asscessed on a publicly available dataset LIDC-IDRI and attained a dice score of 81.1% and a Jaccard score of 72.5%. For lung nodule segmentation, we presented two improved UNET models leading to better receptive fields and atrous convolutions. To retrieve rich information, our first framework is built on dual branches and deep residual learning. It utilises several pooling sizes to gather information at both the local and global levels of contexts. The second model employs naive inception blocks, which are composed of parallel convolution layers wherein different kernel scales are employed to highlight features of varying size nodules. Furthermore, by applying K-means clustering and morphological operators, the computational time of both models is restricted to the lung area of interest (ROI). Both suggested models employ atrous convolution operation to expand the visual field of the filters. The suggested models get 80 % and 86.2 % dice v scores on the publicly accessible benchmark LIDC-IDRI database, respectively, and performs better in all previously existing literature.