| Objectives:With the development of surgical technology and the increase of surgical volume,surgical-related anemia shows an increasing trend.Anemia can lead to related complications such as decreased tissue oxygen carrying capacity,cognitive function and so on.Extensive researches have shown that perioperative anemia increases the incidence of surgical complications and perioperative mortality.Early identification and intervention of anemia patients,and assessment of the severity degree of anemia can not only provide support for clinical decision-making,but also avoid the occurrence of related complications,through which can save medical resources and costs.The existing hemoglobin(Hb)monitoring methods are limited in the application of anemia screening,emergency transport,field first aid,perioperative monitoring and other environments due to the long waiting time,high cost,the need for professional blood collection,and the non-removable equipment.In recent years,the non-invasive anemia monitoring method based on machine learning has shown great potential.This study intends to provide a novel method for rapid,dynamic,and non-invasive Hb monitoring by analyzing the conjunctiva images of surgical patients captured by smartphone,associating the images with the patient’s actual Hb concentration,establishing Hb monitoring model using machine learning algorithm.Methods:The conjunctival images,laboratory Hb concentration and general data of 284 patients undergoing elective surgery in the Department of Anesthesia,the First Affiliated Hospital of the Army Military Medical University from March 18 to April 26,2021,were prospectively collected and analyzed.The conjunctiva area was exposed fully and the eye images was captured by smartphones in an operating room environment.Mark the image according to the corresponding Hb concentration.This study was approved by the Committee(KY2021060)and registered in the Chinese Clinical Trial Registration Center with the registration number of Chi CTR210044138.1.The establishment of anemia monitoring technology based on deep learning algorithm and conjunctiva imagesTwo tasks were applied for performance evaluation.The objective of Task 1 was to predict patients’anemic states(Hb<12 g/dl);The objective of Task 2 was to predict mild anemia(Hb<10 g/dl).Images were labeled as normal and anemia based on different Hb threshold.Using the manual selected conjunctival image as input,four deep learning algorithms,including InceptionV3,ResNet50V2,EfficientNetV2B0 and DenseNet121,were fitted to construct anemia classification prediction models,and the application value of the four models for anemia monitoring was compared.The performance of the model was evaluated by the receiver operating characteristic curve(ROC curve),accuracy,sensitivity,specificity,positive predictive value,negative predictive value.2.The establishment of hemoglobin concentration monitoring technology based on semantic segmentation and deep learningDue to the limitations of manually selected conjunctiva image and classification research in the previous chapter,we have used eye images as input,applied mask R-CNN deep learning algorithm to automatically semantic segmentation of conjunctiva images,and MobileNetV3 deep learning algorithm to predict Hb concentration,and establishes a joint algorithm model in this chapter.At the same time,using the eye image,the manually selected conjunctiva image and the features extracted from the above two images as input,the Hb concentration monitoring model is established based on the traditional machine learning algorithm and MobileNetV3 algorithm alone.Compare the performance differences between the joint algorithm model and other algorithm models.The model performance was evaluated by R~2,explained variance score(EVS)and mean absolute error(MAE).Result:1.The establishment of anemia monitoring technology based on deep learning algorithm and conjunctiva imagesIn task 1,the area under curve(AUC)of the four models,named InceptionV3、ResNet50V2、EfficientNetV2B0、DenseNet121,were 0.709(95%CI,0.643-0.769),0.661(95%CI,0.594-0.725),0.670(95%CI,0.603-0.733),and 0.695(95%CI,0.628-0.756).The AUC,accuracy,sensitivity,specificity,PPV and NPV of InceptionV3 in task 1 were 0.709,69.48%,75.00%,41.21%,70.73%and 62.89%respectively.In Task 2,the AUC of each model is 0.729(95%CI,0.664-0.788),0.758(95%CI,0.695-0.814),0.769(95%CI,0.707-0.824),0.770(95%CI,0.708-0.825);The AUC,accuracy,sensitivity,specificity,PPV and NPV of EfficientNetV2B0 were 0.769,80.28%,32.08%,90.06%,73.91%and 96.25%respectively.Based on the optimal algorithm,a network service application is developed(http://150.158.58.4).2.The establishment of hemoglobin concentration monitoring technology based on semantic segmentation and deep learningModel performance based on combination of mask R-CNN and MobileNetV3 using the eye images achieved an R~2,EVS and MAE of 0.503(95%CI,0.499-0.507),0.518(95%CI,0.515-0.522)and 1.6 g/d L(95%CI,1.6-1.6 g/d L),which was similar to that based on MobileNetV3 using the manually selected palpebral conjunctiva images(R~2:0.509,EVS:0.516,MAE:1.6 g/d L).Conclusion:1.Models,which established based on the deep learning algorithm using conjunctiva image as input,have a good effect on fast and automatic prediction of anemia.The comprehensive prediction performance of InceptionV3 model is excellent in Task 1,and the EfficientNetV2B0 model performs better in Task 2.2.We have developed a non-invasive and dynamic hemoglobin concentration monitoring model based on deep learning,which can rapidly,non-invasive and dynamic monitor Hb in the environment of the surgical period,anemia screening,the treatment site,and the transport of patients.3.The image acquisition tool is a smart phone with strong mobility,eliminates the need to collect blood samples,achieving non-invasive Hb monitoring.Directly using the image as model input,the model doesn’t need to selected conjunctiva image manually and extract image features by manual,and achieved dynamic Hb monitoring. |
PDF Full Text Request