UDCT: lung Cancer detection and classification using U-net and DARTS for medical CT images

Lung cancer is the most fatal disease in recent times. Early detection of the same is very crucial and challenging task. Therefore, proper diagnostic and treatment strategies should be employed for the automatic detection of nodules from Computed Tomography (CT) images. In this direction this paper introduces a novel UDCT, a new and effective framework for detecting and classifying lung cancer based on CT images. The proposed framework exploits Modified U-Net architecture that utilizes the traditional U-Net architecture for multi-scale feature extraction with the DARTS (Differentiable Architecture Search) which utilizes the concept of continuous relaxation of the architecture representation with efficient gradient descent search for efficient classification of lung cancer. In addition to this, the proposed framework, incorporates Multilevel Otsu thresholding for image pre-processing and to further segment the lung nodule from the CT scan images. Furthermore, the modified U-Net is capable to detect and classify lung nodule effectively and efficiently. To evaluate the performance of the proposed UDCT framework, extensive experiments are conducted on LIDC-IDRI and IQ-OTH/NCCD CT image datasets that provided state-of-the-art results. The proposed UDCT framework provided an accuracy of 95.01% on LIDC-IDRI dataset and 96.82% on IQ-OTH/NCCD dataset. An ablation study is also conducted that validated the efficiency and efficacy of the proposed UDCT framework.