基于UNet融合液体神经网络与tanh激活函数的模型分割甲状腺结节的探索性研究

打开文本图片集

ABSTRACTToanalyzetheacuracy UNet integrated with liquid neural networks tanh activation function in segmentingthyroid nodules， toexploreanovelmethodfordynamicalloptimizing UNetsegmenters.Atotal 3493thyroid ultrasoundimages fromtheTN3Kdataset wereromlydivided intoatrainingset（2879 images）atestset（614images）at aratio7：3.The clasical UNetsegmenterservedasthe baseline model.Aliquid neuralnetwork（LNN）module was embedded intotheencoding，botleneck，decoding layers UNettoconstructtheLNN-UNet model.Buildingonthis，thetanh activationfunctionwasintroducedintotheLNNfeedbacklooptoconstructtheLNN-UNet-tanh model.Thesethreemodels were usedto segmentthe 3493 thyroid ultrasound images.The AdamoptimizeramixedBCE-Dicelossfunction wereused for modeloptimizationduringtraining.Theperformanceeach modelinsegmenting thyroidnodules wasevaluatedbyareaunder thecurve（AUC），Dicecoeficient，itersectionoverunion（U），F1sore，acuracy.Resultsshowedtatintheetset，the AUC for thyroidnodule segmentationbytheUNet，LNN-UNet，LNN-UNet-tanh models were 0.9159，0.9736，0.9831， respectively.Dicecoeficients intheestsetwereO.7787，0.8174，0.8417，respectively.IoUinthetestset were0.4871， 0.5118，O.5773，respectively.F1 scores inthetestset were0.6102，.6220，0.6725，theaccuacywere 0.9305， 0.9328， O.9474，respectively.Theresults indicatedthatforthyroidnodulesegmentationtasks，the LNN-UNet-tanhmodel outperformsboththeUNetLNN-UNetmodels.Itcanprovidemorepreciseregion--interestalgorithmsupportforclinical intelligentdagnosticmodels，herebyimprovingthetranslationrateclinicalresouresintomedicalatificialintelligence.

KEY WORDSUltrasonography； ; Image segmentation； UNet； Liquid neural network；Hyperbolic tangentactivation function ; Deep learning

甲状腺结节在普通人群中的患病率高达 67% ，在女性和老年人群中发病率更高，且其检出率随着影像技术的发展和普及持续上升1]。（剩余11079字）