End-to-End Fully Automated Lung Cancer Screening System

The computer aided diagnosis of lung cancer is majorly focused on detection and segmentation with very less work reported on volume estimation and grading of cancerous nodule. Further, lung cancer segmentation systems are semi automatic in nature requiring radiologists to demarcate cancerous portions on every slice. This leads to subjectivity and delayed diagnosis. Further, these techniques are based on standard convolution leading to inaccurate segmentation in terms of actual boundary retention of the cancerous nodule. Also, there is a need of automatic system that not only grades the lung cancer based on actual parameters but also enables early warning for flagging of anomalies in periodic screening. This research work reports the design of a fully automated end-to-end screening system that consists of 5 major models with an improved performance on cancer detection, segmentation, volume estimation, grading, and an early warning system. The traditional convolutional technique is modified to allow for retention of actual shape of cancerous nodule. The simultaneous segmentation of cancer, lymph nodes and trachea is also achieved through a focus module and a modified loss function to remove redundancy and achieve an accuracy of 92.09%. The volume estimation model is developed using GPR interpolation to give an improved accuracy of 94.18%. A grading model based on the TNM classification standard is developed to grade the detected cancerous nodule to one of the six grades with an accuracy of 96.4%. The grading model is further extended to develop an early warning system for changes in the CT scans of lung cancer patients under treatment. The research is undertaken in collaboration with Nanavati Hospital, Mumbai, and all the models are validated on a real dataset obtained from the hospital.