| Malignant tumors have become one of the major public health issues in the world.And China leads the world in the number of new cancer cases and deaths,which conspicuously affects China’s national health.What’s more,early screening,diagnosis,and treatment of cancer have become current research hotspots and been incorporated into the national health strategic plan.At present,although malignant tumors can be detected by immunology,genetics,molecular biology,and other techniques,the popularization of such examination methods is too expensive to promote.Imaging pathophysiology have become the first choice for clinical malignant tumor detection because of their intuitive image analysis.However,this kind of technology weightily depends on the knowledge of clinicians,resulting in it being difficult to carry out accurately during the early screening of cancer.In addition,if cancer cells are infiltrated,clinical diagnosis and prognosis treatment will be severely affected due to its low proportion and atypical cell morphology after infiltration.Therefore,how to develop effective high-throughput malignant tumor diagnostic and use intelligence analysis to compensate for the shortcomings of the current technology are not only crucial to the realization of precision medicine,but also is the researching emphasis of the 14 th Five-year Plan of China,which is theoretically and practically significant.This thesis combined with artificial intelligence technology and microscopic imaging technology is based on the advantages of the medical and industrial crossplatform,which aims at the current limitations in malignant tumor detection.The research focuses on key technologies of intelligent analysis and high-throughput detection for hematological malignancies and exudate cytology in cancer infiltration.The main research contents and conclusions of this thesis are as follows:(1)Existing intelligent analysis methods for multiple myeloma(MM)mainly focus on multi-stage two classifications,which mainly detect monoclonal plasma cells(PCs)and others.However,the clinic also pays close attention to the proliferation of other lineages of cells in the bone marrow(BM),which is conducive to the discovery of suppressed cell lineage as evidence for more targeted treatment.Additionally,existing methods require manual separation of single-cell images from bone marrow smears,leading to extremely time-consuming and laborious diagnoses which are difficult to apply effectively in the clinical setting.As a result,we present an automated diagnostic MM synthetic diagnosis method on the whole BM slide.First,we introduce the preparation and standardization of the cell collection method.Secondly,we consider the clinical need to combine multiple cell ratios as therapeutic evidence;we therefore model MM as a multi-object detection task.Moreover,we utilize multi-scale image geometry broadening and color enhancement methods to improve cell staining morphological differences,which efficiently improved the accuracy of cell detection.And We construct the lightweight neural network to deep mine features of MM cells,realizing high-throughput real-time automated detection of the common six lineages of cell on the whole slide.In the end,we have solved the problem of high-accuracy detection and identification of six lineages of cell lines in bone marrow,which plays a significant role in assisting clinicians in making accurate and comprehensive judgments according to patients’ disease status.Notably our results showed that this newly developed model successfully detected most cells in BM microscopic images,especially monoclonal PCs,with an accuracy of 92.45%.In the meantime,it’s worth noting that the erythroid,granulocytic,lymphoid,and monocytic cell lineage and others likewise achieved a high accuracy rate of 91.23%.91.67%,81.82%,86.21% and 89.91%,respectively.The detection speed can reach 80 frames per second,which resoundingly meets the requirements of low complexity and efficient cell detection and analysis in clinical.This method can provide a computeraided diagnosis for cytologists in clinical practice,thus saving labor costs and providing new ideas for similar cytological analysis applications.(2)Current clinical Acute Leukemia(AL)typing methods are often stain-based and highly limited by its specificity and availability of staining which probes to meet the demand for high-throughput clinical detection.Moreover,existing cell image intelligent processing methods require manual labeling of single cells,which is conspicuously limited by sample purity and difficult to obtain a fully supervised dataset,thereby hindering effective modeling of AL typing,and rather than specific cell identification,AL typing should also take full account of the overall state of the entire sample.Therefore,we propose an attentional neural network to collaborate label-free AL comprehensive classification method for an optofluidic time-stretch(OTS)microscopy imaging system.Firstly,we use OTS microscopic imaging system to obtain many label-free cell images with a throughput of 100,000 cells/s and a spatial resolution of better than 780 μm.Secondly,we utilize machine learning to roughly screen many cells,aiming at obtaining valid cell images.Accordingly,we design a multi-channel attention neural network to transform AL sample typing into cell image group classification,that is,to judge whether the cell image group contains relevant AL features.What’s more,in order to overcome feature mining difficulties caused by sample diversity,virtual adversarial training and entropy minimization techniques are used to enhance model robustness and improve network decisionmaking ability.Eventually,an ensemble learning voting mechanism is proposed to aggregate scores for all cell image groups to obtain the final AL type.In this thesis,30 clinical samples were used to obtain a total of 227620 cell images,which proved the high-throughput detection capability of the system,18 samples were trained and 12 samples were used for testing.Ultimately,the results revealed that this method achieved AL high-precision and comprehensive typing,and proved the potential of the high-throughput labeling-free AL typing method proposed in this thesis.(3)Since malignant tumors may infiltrate and metastasize,in which would cause the formation of pleural and abdominal effusions,it is extremely considerably to distinguish benign effusions from those caused by malignant hyperplasia in clinical diagnosis,and the variability of their diagnostic results can affect disease staging and treatment choices.Existing diagnostic techniques usually use cytology slice image analysis,which relies heavily on human expertise,and the analysis efficiency is low,making it difficult to meet the needs of efficient cancer screening.As for low proportions of malignant samples,cancer cells may not be identified resulting in missed or false detections.Therefore,we propose to apply an optofluidic time-stretch(OTS)microscopic imaging system and construct a multi-stage lightweight analysis framework suitable for high-throughput detection.First,according to the prior knowledge of the pathological morphology of pleural effusion in clinical,we purify the samples by using machine learning.,Then,we use a lightweight multi-head selfattention deep neural network to fuse local features with global features to achieve efficient deep characterization of cancer cells combined with effective regularization training strategies to improve the detection performance of malignant cells.In the end,we use 40 clinical samples for experimentation and validation,of which 18 clinical samples were used to test the validity of the method,including 3 malignant,4 benign,and 11 mixed samples.The results revealed that the accuracy of this method for detecting malignant pleural effusion cells was 90.53%,which was consistent with clinical diagnosis,and the efficient cell analysis was achieved at 754 frames per second.The high throughput,high accuracy,and high convenience of this method make it a promising solution for the diagnosis of malignant serous effusions in various conditions. |
