Automated Rating Of Semantic Attributes Based On Multitasking Features In CT Images Of Pulmonary Nodules

Lung cancer,manifested as pulmonary nodule initially,is a malignant lung tumor with fastest growing morbidity and mortality,and early diagnosis is critical to lung cancer treatment.Computed tomography(CT)is a widely used imaging modality for the assessment of pulmonary nodules.However,complicated semantic characteristics of pulmonary nodules in CT images is usually defined and rated in a subjective manner,this process is time-consuming as well as operator-dependent.Computer-aided diagnosis(C AD)is a n assistive software package to provide computational diagnostic references for the clinical image reading and decision support,It has been shown to be effective to lower down the inter-observer variation and save time.We present two pulmonary nodule-assisted diagnostic systems in this paper.We propose a method based on multi-task linear regression(MRLR)model to rate 8 semantic features of pulmonary nodules automatically.Next,we propose a multiple attribute-assisted diagnosis of lung cancer(MAADLC)method to diagnose the malignancy as well as predict the other 8 semantic features.The gap between the computational and semantic features is the one of major factors that bottlenecks the C AD performance from clinical usage.To bridge such gap,we propose to utilize the MRLR scheme that leverages heterogeneous computational features derived from deep learning models of stacked denoising autoencoder(SDAE)and convolutional neural network(CNN)as well as Haar-like features to approach 8 semantic features of lung CT nodules.We proposed MAADLC,an extension of MTLR,to corroborate the effectiveness of cross semantic task relation.First,we exploit more MTL schemes and augment the feature pool with the Ho G features for more thorough study.Second,larger number of 2400 nodules are involved in this study.Meanwhile,we also include the "Mal" semantic feature here for complete investigation.Third,more comprehensive experiments are performed in this study.Specifically,we explored the composite learning scheme to perform the regression for the first 7 tasks and classification for the last two 2 exceptional tasks in a joint-learning fashion.To support the training and testing of CADa scheme,we adopt the Lung Image Dat abase Consortium(LIDC)dataset for its rich annotation resource.The result of MTLR is 0.65± 0.56(absolute error mean± standard deviation)while inter-observer variation among the radiologists' ratings is 0.58 ± 0.78.And the result of MAADLC is 0.69 ± 0.59 while inter-observer variation among 4 groups radiologists is 0.70± 0.79.It will be shown that computerized rating results are comparable to the radiologists.This paper creatively proposed a framework to train 9 semantic tasks jointly based on the current popular machine learning methods,which can solve clinical proble ms effectively.In addition,this rating retrieval framework is generic and extensible to thrust the content-based image retrieval.