Key Techniques For Pulmonary Nodule Detection And AI Diagnosis In Chest CT Scans

Lung cancer is a kind of malignant tumor,which has the highest morbidity and mortality among all caners of human beings.Early discover and treatment of lung cancer can greatly increase the survival rate of patients.In clinical practice,low dose spiral computed tomography(CT)is one of the most used tools for screening pulmonary nodules(the suspicious early lesions of lung cancer).Nodule screening entails high work to radiologists:they must find nodules of various types across a large number of CT images generated from thin-sliced reconstructions.It is relatively easy for radiologists to locate big nodules,but some small ones,especially the ground-glasses and the solids surrounded by other tissues,are difficult to find even by experienced radiologists.These small nodules are most of seen in clinical practice.Fatigue plus to the subjectivity of human beings can lead to omission of nodules.Besides,the doctors need to analyze the detected nodules to establish the treatment plan.Analysis of the nodules relies on some features of the detected nodules,such as size,intensity and location.Manual extraction of these features demands relatively high workload and the results may vary greatly from individual to individual.Therefore,computer-aided systems(CAD)has been one of the research hotspots in both academic and industrial fields.An enormous challenge faced in the design of CAD systems is that nodules have a broad spectrum of appearance and various distributions inside the lung regions,which is unable to cope just using traditional image processing techniques.Based on the study of relevant literatures and existing CAD systems,we have performed a systematic research work on the automatic detection and diagnosis of pulmonary nodules.Our major contributions can be concluded as:(1)We propose a framework for nodule detection from volumetric chest CT scans using CNNs-based nodule-size-adaptive detection and classification.The framework can detect all kinds of nodules,including ground-glasses,part-solids and solids,and nodules of size ranging from 3mm to 70mm(in clinical practice,nodule-like lesions with size larger than 30mm are usually called masses).Evaluations on independent and publicly available dataset demonstrates that the performance of the proposed framework is superior to existing CAD systems and comparable to manual screenings,especially in the detection of nodules with size smaller than 5 mm.(2)Based on the nodule detection,we designed a nodule segmentation method which combines traditional image segmentation techniques and a so-called ray method.After segmentation of a nodule,we compute the nodule's size,volume and average intensity.Comparison to golden-standard from doctors demonstrates that the proposed segmentation can achieve promising results and can satisfy the requirements in clinical applications.(3)We propose a 3DCNN-based landmark classifier for determining the lung segment in which a detected nodule is located.Training a deep learning model usually requires a large number of annotated samples.In our study,we avoided this issue using a simple yet effective strategy.Experiments on two hundred of scans reveal that the proposed classifier,even trained from just eight scans,can achieve promising result which is comparable to human observers.(4)We propose a 3DCNN-based classifier to estimate the malignancy of nodules.Each nodule is classified into one of three categories:benign,preinvasive and malignant.The research work in this study was conducted on special clinical data,which was collected from targeted scans.Comparison to conventional scans shows that targeted scans can significantly improve the performance of the classification.Additionally,the proposed classification is compared to the diagnostic results from doctors,which also demonstrates that the proposed classifier,even trained from just a few samples,can achieve results that comparable to the doctors.