| Part 1 The reliabilities of radiographic measurements of cervical curvature in cervical spondylosis[Objective]To validating the reliability of manual measurement of cervical curvature using Cobb,Gore and Centroid methods in cervical spondylosis[Methods and Materials]Seventy-five cervical lateral radiographs of local hospital were collected retrospectively.Two spine surgeons(10 years and 8 years of clinical experience)as examiners measured twice with a delay using the three methods,respectively.Cobb method was defined as angle formed by two lines that parallel with the bottom endplate of each C2 and C7 vertebrae.Gore method was defined as angle formed by a line in the posterior body margin of C2 and a line in the posterior body margin of C7.Centroid method was defined as angle formed by a line that passes through the midpoint of the inferior surface of C2 and the centroid of C3,and a line that through the centroid of C6 and C7.Negative value denotes lordotic posture and positive value denotes kyphotic.Measurements were implemented on a developed Image J software(Version 1.53h).The measuring error of the three methods was analyzed using the metrics of mean absolute difference(MAD)and intraclass correlation coefficient(ICC).[Results]The dataset comprised of seventy-five X-ray radiographs.The mean age of the subjects was 56±22 years.Intraclass ICC of Cobb method was higher than Gore and Centroid method[0.991(CI,0.986~0.994)vs 0.985(CI,0.973~0.989),0.987(CI,0.979~0.992)].MAD of Cobb method was the smallest(1.51°±0.62° vs 2.00°±1.06°,1.63°±0.94°),whereas,the intraclass ICC of Cobb and Centroid method was not statistically significant.Interclass ICC of Cobb method was also higher than Gore and Centroid method [0.982(CI,0.971~0.988)vs 0.970(CI,0.954~0.981),0.980(CI,0.968~0.987)].MAD of Cobb method was the smallest(1.94°±1.17° vs 2.73°±1.45°,3.04°±1.34°),whereas,the interclass ICC of Gore and Centroid method was not statistically significant.[Conclusion]Cobb method gained higher reliability than Gore and Centroid method in the radiographic measurements of cervical curvature in cervical spondylosis.Part 2 Construction and preliminary validation of the effectiveness of deep learning model in cervical radiography[Objective]To propose a lightweight convolutional neural network(CNN)model with Res Net as the core and embedded with Transformer model in the measurement of cervical radiograph,and evaluate the effectiveness of the model.[Methods and Materials]Dataset A consists of 1474 cervical lateral radiographs from 1036 subjects obtained between February 2020 and September 2021 from Picture Archiving and Communication Systems(PACS)database of our hospital.Dataset B consists of 669 film-version images from 396 subjects selected randomly from the above dataset A.The ground-truth annotations was gotten form sophisticated senior surgeons.Each of four corners of C2 to C7 vertebra were annotated by rules of Cobb method,and the Cobb angle was calculated.Evaluation of cervical instability was conducted in the joint effort of two surgeons.The training: validating sets were demarcated by a ratio of 8:2.We proposed an encoder-decoder CNN model: 1.Encoding process was performed using the Res Net 34 pretrained on Image Net database,five layers of which went downsampling and extracted feature map;2.Transformer module with its intrinsic encoder and decoder was embedded to output one-dimensional features.Then features were transformed to vector,went into the decoder for deconvolution,and to extract the heatmap of the centroid and corner offset;3.We introduced a Rotating Attention module and integrated it between the encoder and decoder to extract channel and spatial information;4.We proposed a Vector Loss module to address deviation and loss problems of landmark localization in conditions of sever abnormally curved,S-shaped profile,or obscureness of the cervical region.Euclidean distance was used as metric of the landmark-localization error.Error of Cobb angle measurement was evaluated by mean absolute difference(MAD)and symmetric mean absolute percentage error(SMAPE).Our model was compared with stateof-the-art models as landmark-based MVE-Net and segmentation-based RU-Net.Finally,the ablation study was conducted to validate effectiveness of all modules.[Results]Ablation study showed a progressive enhancement of the accuracy of localization while adding on Rotating Attention module,Transformer module,and Vector Loss module,with Euclidean distance dropped from 6.64 to 5.23,MAD dropped form 5.32° to 2.38° and SMAPE from 10.71% to 8.66%.Visualization results display a remarkable elimination of misrecognition and missed-recognition.In addition,transfer-learning with pretraining elevated the accuracy of the intact model,with Euclidean distance dropped from 5.94 to5.32,MAD dropped form 3.33° to 2.38° and SMAPE from 9.24% to 8.66%.Our model performed better with a MAD of 2.38°compared with 6.43° of MVE-Net and 4.32° of RUNet.The model was efficient even in the interference of ossified posterior longitudinal ligament,spondylolisthesis,and osteophytes.[Conclusion]The proposed lightweight convolutional neural network detection model embedded with Transformer yielded better accuracy of landmarks reorganization and angle measurements than other CNN models,and showed high-level robustness in images with paraspinal ossification and malalignment.Part 3 Study on the accuracy of cervical radiograph deep learning model[Objective]To validate the accuracy of cervical curvature measurement and instability evaluation of the proposed convolutional neural network(CNN)model embedded with Transformer on digital and film-transformed version images,and make comparison with human examiners.[Methods and Materials]The testing datasets were arranged as annotated 353 digital images from dataset A and150 film-transformed images from dataset B that mentioned in Part 2 section.The ground truth(i.e.,gold standard)of Cobb angle was acquired from surgeon 1 and cervical instability from mutual discussion by surgeon 1 and surgeon 2.Differences between our CNN,MVENet,and RU-Net prediction and the ground truth were calculated.Future,other three surgeons(5 years,3 years,and 2 years of vocational experience)were invited as human examiners to evaluate cervical instability.Error of curvature was calculated from two dimensions,i.e.,C2-7 Cobb angle and unstable-segment interspace angle.Cases with or more than two unstable segments were calculated as mean error of all segments.Evaluation metrics of instability was defined as accuracy,sensibility(recall),specificity,precision and F1-score.The Kruskal-Wallis test was used to compare error between groups.The BlantAltman graph,and the Cohen’s Kappa analysis were used to analyze the consistency between models and human examiners.[Results]On Cobb angle measurement,our model outperformed the other two MVE-Net and RU-Net,whereas,was slightly inferior than human counterparts.In detail,on dataset A,the median(MAD)and interquartile range(IQR)of C2-7 Cobb measurement yielded by our model,MVE-Net,RU-Net,and human were 2.06°(2.00°),2.44°(1.97°),2.32°(2.01°),2.08°(1.82°),respectively.There was no statistically significant difference between MAD of our model and of human(P=0.40),while SMAPE of our model was larger than human(16.63%,15.97%).On dataset B,the MAD(IQR)was 2.21°(1.84°),2.73°(1.66°),2.32°(1.92°),2.18°(1.45°),and MAD of our model was larger than human(P<0.01),while SMAPE was larger than human too(10.30%,9.83%).On unstable segment Cobb angle measurement in dataset A,the MAD(IQR)was 1.54°(2.13°),1.97°(2.12°),1.88°(1.97°),1.47°(2.20°).MAD of our model was larger than human(P<0.01),while SMAPE was larger than human too(19.35%,17.69%).On dataset B,the MAD(IQR)was 1.85°(2.22°),2.73°(2.24°),2.12°(1.99°),1.92°(2.32°).MAD of our model was smaller than human(P<0.01),while SMAPE was smaller than human too(19.01%,19.22%).As for instability evaluation,our model outperformed MVE-Net and RU-Net on accuracy(Dataset A: 89.80% vs 88.39%,88.7%;Dataset B: 90.00% vs 88.00%,86.67%).In gross,our model was inferior than the three surgeons on accuracy(Dataset A:89.90% vs 90.65%,92.35%,1.22%;Dataset B: 90.00%vs 90.67%,93.33%,92.00%).Correlations among all models and human examiners were moderately high.Kappa coefficient between each model was larger than that between human,while Kappa coefficient between each model and the ground truth was smaller than that between human and the ground truth.[Conclusion]The proposed deep learning CNN model for cervical curvature measurement and instability evaluation displayed substantial effectiveness and narrow gap with human examiners,and kept its robustness on digital and film-version radiographs. |
PDF Full Text Request