Automated Analysis Of Multiphoton Microscopic Features And Its Application In Prognosis And Pathologic Diagnosis Of Breast Cancer

Biomarker features observed or extracted from medical images are indispensable for cancer screening,diagnosis,and prognosis prediction.Medical imaging techniques,represented by Magnetic resonance imaging and computed tomography,have been widely used for early screening and diagnosis of cancer,while histopathological images based on hematoxylin and eosin-stained sections are considered the gold standard for tissue examination and pathological diagnosis.Numerous studies have shown that the tumor microenvironment has an important impact on tumor generation,development and metastasis by regulating the polarity,migration and signaling of cancer cells.For heterogeneous tumor diseases,the tumor microenvironment varies greatly among individuals,so its in-depth study can provide pathologists with important insights into tumor progression.Traditional medical imaging techniques,however,are constrained by imaging modalities and are unable to image some important components of the tumor microenvironment in high resolution.Multiphoton microscopy is an advanced optical imaging technique that enables labelfree,low-damage imaging of cells and subcellular structures through a variety of nonlinear optical effects generated by biological tissues under femtosecond laser excitation.In multiphoton imaging,the multiphoton microscopic features obtained through different optical channels are closely related to the biological properties of specific tissue components and contain important information about tumor progression and histopathological changes.Currently,the extraction and analysis of multiphoton microscopic features is still in its infancy,and many research areas are still blank,for example,the complex interactions between heterogeneous tumor-associated collagen during tumor development,which has not been reported in studies on the impact on tumor progression and prognosis.In addition,effective methods for histopathological change assessment and automated detection on large-size multiphoton microscopic images containing tens of millions of pixels are lacking.In this paper,we investigate the multiphoton microscopic features of breast cancer,a highly heterogeneous tumor disease,and explore various methods and models for automated analysis,with the main work including:(1)A intratumor graph neural network model is developed based on graph machine learning technology.The model uses irregular graph data structure to characterize and reason the heterogeneous tumor-related collagen features,revealing and explaining the complex interaction between different collagen features in the tumor process for the first time.Great prognostic value is recovered from this abstract “relational feature”.(2)Multiphoton microimaging analysis of breast cancer sentinel lymph nodes was performed,and an automated histopathological assessment model was developed by measuring morphological features and multifractal spectrum characteristics of a series of collagen structures.A fully convolutional neural network model was also developed,capable of performing fast inference on large size multiphoton microscopic images containing tens of millions of pixels to accurately identify lymphatic metastasis status and locate metastases.(3)Multiphoton microscopic imaging analysis of breast tumor tissues after preoperative chemotherapy was performed,and two multiphoton microscopic features,cell density and collagen content,were measured separately and used to monitor tumor regression response,and an accurate tumor regression grade evaluation system was developed.