Artificial Intelligence Pipelines For Lung Cancer And Esophageal Cancer Diagnosis

With the development of artificial intelligence（AI）in the field of medical imaging,it is increasingly becoming the mainstream of computer-aided diagnosis（CAD）research.With the help of CAD system,radiologists can efficiently obtain the diagnostic results of the computer as second opinion,and make the final diagnosis and decision more accurately.Fully automated pipeline is one of the future trends of CAD development,for it enables clinicians to use the system more conveniently.Therefore,the research and development of medical imaging AI models covers a wide range of applications,including segmentation and detection of organs and lesions,diagnosis,prognosis assessment etc.Furthermore,model interpretability is also essential for AI model in high-stake decision-making,such as healthcare and criminal justice.CAD systems are not only expected to help clinicians improve diagnostic accuracy,but also expected to provide evidences and clues of the predictions,which will make the system more reliable.In this paper,we explored several major processes of the automated CAD pipelines,including segmentation,diagnosis,survival analysis and model interpretability.We also constructed models for esophageal cancer（EC）and lung cancer diagnosis,which enabled us to evaluate the application of automatic AI pipelines in clinical settings.In the study of EC diagnosis based on magnetic resonance imaging（MRI）,we built models for EC lesion segmentation,survival analysis and tumor regression grade（TRG）classification and achieved state-of-the-art performance.First,due to the small size and low contrast of EC lesions,we proposed a 3D U-Net segmentation network with a spatial attention mechanism.We coarsely segmented low-resolution MRI images scanned with Star VIBE sequence,and used the coarse segmentation as the attention map in another U-Net model for fine segmentation of EC lesions on highresolution MRI images scanned separately with Star VIBE sequence.This method can effectively alleviate the problem of failing to locate EC lesion.In the survival analysis of EC,to avoid overfitting caused by the numerous radiomics features,we proposed a feature screening method to obtain features the most relevant to survival.We combined radiomics features with clinical variables to build a combined survival analysis model,which outperformed the model using only clinical features and the model using both clinical features and T-grading determined by MRI reading.For TRG classification of EC,we used the difference of radiomic features,named delta-model,extracted from MRI images before and after chemotherapy to improve the low accuracy of classic radiomics models using images from single modality.The test AUC（Area Under Curve）of the delta-model was 0.842,while the models using only pre-chemotherapy or postchemotherapy MRI images only archived 0.408 and 0.788,respectively.In the study of CT-based lung cancer diagnosis,we proposed a multi-task framework,named interpretable multi-task Lung Nodule Diagnosis model（im LND）,to build new interpretable deep learning models,which increased both performance and interpretability of the neural model.We firstly segmented the lung nodules and input the segmented lesion together with segmentation probability map into the im LND for classification.Different from most related diagnosis networks,our im LND model not only achieved high accurate lung cancer diagnosis performance,but also tried to solve the “black box” problem of AI model by pointing out the manifestations relevant to the diagnosis.We validated our method on both public and in-house datasets,each with different diagnosis task.It was demonstrated that the proposed method increased not only the interpretability of the model,but also the diagnostic performance,which may be attributed to the fact that learning features of relevant manifestations is also beneficial to extraction of features more relevant to the diagnosis.Our study has demonstrated the value of the AI-empowered CAD models in clinical diagnosis and can potentially provide inspirations for other research working on AI’s interpretability,which will facilitate the application of CAD models in clinical settings.