An Automated Workflow for Lung Nodule Follow-Up Recommendation Using Deep Learning

Early detection of lung cancer increases a patient’s survival rate and provides healthcare professionals, valuable time, and information to administer effective treatment. Lung nodules are early signs of lung cancer. Computer-aided diagnostic systems that can identify pulmonary nodules improve early detection as well as provide an independent second opinion. We propose an automated workflow for follow-up recommendation based on low-dose computed tomography (LDCT) images using deep learning, as per 2017 Fleischner Society guidelines. As per guidelines, follow-up is based on size, volume and texture of nodules. In this paper, we present a 5 stage approach for automated follow-up recommendation. The 5 stages are Lung segmentation, Nodule detection and False Positive Reduction (FPR), Texture classification, Nodule segmentation and Follow-up recommendation. Our nodule detection has a sensitivity of 94% @ 1 false positive per scan. The FPR network improves the specificity of detection to 90% without changing sensitivity. Nodule segmentation has a Jaccard index of 0.77 on 768 nodules from Lung Nodule Database (LNDb) [1]. Texture classification has a sensitivity of 97% on solid nodules and a Fleiss-Cohen’s Kappa of 0.37 on LNDb data with most errors between sub-solid and solid nodules. Our rule-based follow-up recommendation has a Fleiss-Cohen’s Kappa of 0.53 on 236 patients from LNDb. In conclusion, we found that rule-based approach for follow-up alongside deep learning models is the best approach in achieving best results. As we improve the first 4 stages, we foresee that recommendation from AI will become closer to radiologists recommendation.