Survival Prediction Model For Pancreatic Cancer Patients Based On Pathological Whole-slice Image Analysi

Pancreatic cancer is a highly malignant digestive system tumor with a very low five-year survival rate after diagnosis.Accurately predicting the patient’s survival risk and exploring factors closely related to survival is crucial for treatment planning.Histopathology is the gold standard for disease diagnosis and subtyping.It can conduct in-depth analysis of tumor microenvironment under high-resolution microscopy and is an important method for evaluating risk.However,manual evaluation of tumor microenvironment indicators is labor-intensive and subjective,making it difficult to apply in clinical practice.Therefore,this thesis aims to quantitatively analyze multiple tissues and cells in pancreatic cancer whole slide images(WSIs)through computer analysis,explore tumor microenvironment factors closely related to survival,and accurately predict the survival risk of pancreatic cancer patients.This thesis proposes a multi-task and attention-based approach for pancreatic tissue segmentation in pancreatic cancer WSIs,addressing the challenges of tissue complexity and limited annotated samples.The proposed model includes auxiliary tasks to enhance the generalization performance and achieves promising results in identifying eight tissue types,including tumor,stroma,blood vessels,nerves,lymphocytes,normal acini,and background.Then,considering the importance of multi-scale contextual information for cell segmentation tasks,the thesis proposes a modified UNet architecture by incorporating densely connected Transformer modules instead of simple skip connections.This strategy allows for effective fusion of features from different levels of the encoder,enabling accurate segmentation of tumor cells,lymphocytes,and connective tissue cells.Additionally,a post-processing strategy based on concavity detection is proposed to address cell adhesion issues,further improving the segmentation performance.Based on the two topics below,a series of histopathological features are extracted based on various factors such as the number,distribution,and morphology of different tissues and cells.Feature selection methods and machine learning techniques are then employed to predict the survival of pancreatic cancer patients.On an independent test set,the model achieved an AUC value of 0.751 for classifying high and low survival risk patients,and Kaplan-Meier survival analysis showed significant differences in patient prognosis based on the model score.Furthermore,the feature selection results highlighted the importance of lymphocyte-related features,which is in accordance with clinical knowledge.The proposed pipeline not only achieves accurate prediction of patient survival risk,but also helps doctors better understand the progression mechanism of tumors,thereby assisting in developing more appropriate treatment plans for patients and improving their quality of life.