Lung Nodule Classification on CT Images Using Deep Convolutional Neural Network Based on Geometric Feature Extraction

Lung cancer detection in the earlier stage is essential to improve the survival rate of the cancer patient. Computed Tomography [CT] is a first and preferred modality of imaging for detecting cancer with an enhanced rate of diagnosis accuracy owing to its function as a single scan process. Visual inspections of the CT images are prone to error, as it is more complex to distinguish lung nodules from the background tissues which are subjective to intra and interobserver variability. Hence, computer-aided diagnosis is essential to support radiologists for accurate lung nodule prediction. To overcome this issue, we propose a deep learning approach for automatic lung cancer detection from a low dose CT images. We also propose image pre-processing using Efficient Adaptive Histogram Equalization based Region of Interest [EAHE-ROl] to enhance the CT scan and to eliminate artefacts which occur due to noise and variations of the image. The ROI is extracted from CT scans using morphological operators, thus reducing the number of false positives. We chose geometric features as they extract more geometric elements like curves, lines and points of cancer nodules. Our Non-Gaussian Convolutional Neural Networks [NG-CNN] architecture contains feature extractor and classifier, which has been applied on training, validation and test dataset. Our proposed methodology offers better-classified outcome and effectual cancer detection by outperforming the other competing methods and gives a test accuracy of 94.97% and AUC 0.896.