A A deep ep l lear arni ning-bas based ed p pipel peline ne f for e error de detecti tion and and qua quality ity c contr trol of of br brai ain M MRI segm egmentatio tion r resu esults ts I. Brusini 1,2 , D. F. Padilla 2 , J. Barroso 3 , I. Skoog 4 , Ö. Smedby 1 , E. Westman 2 , C. Wang 1 1 KTH Royal Institute of Technology (Stockholm, Sweden), 2 Karolinska Institute (Stockholm, Sweden), 3 Universidad de La Laguna (La Laguna, Spain), 4 University of Gothenburg (Gothenburg, Sweden)
Oute ter g grey ey m matte ter Brai ain M MRI segm egmentati tion Wh Whit ite ma matter (GM) l ) layer er (WM) l ) layer er • Useful to study different properties of brain structures (e.g. cortical thickness, ventricular volume, etc.) • Many tools available for automatic segmentation, but prone to errors • Quality control (QC) is needed, but it can be very slow and subjective if performed visually Cereb ebrospinal f fluid ( (CSF) MIDL Conference 2020 Brusini et al., Automatic QC and error detection in brain MRI segmentation
1. 1. Crea eatio tion o of er error m map aps Data from ADNI [1], GENIC [2] and H70 [3] cohorts. s G ( s ) Input segmentations s from FreeSurfer [4]. Generator G ERROR MAP SLICE Post-processing: Error removal outside a dilated brain mask Thresholding > 0.5 Gaussian smoothing Discriminator D ( σ = 1 mm) Fake or real? Original MRI slice One pix2pix model [5] for each view (axial, coronal, sagittal). All trained using random slices from 1600 [1] Jack et al. (2008), J Magn Reson Imaging ; [2] Ferreira et al. (2015), Psichotema ; [3] segmentations that had been visually rated as good. Thorvaldsson et al. (2008), Neurology; [4] Fischl et al. (2012), NeuroImage; [5] Isola et al. (2017), Proc IEEE Comput Soc Conf Comput Vis Pattern Recognit. MIDL Conference 2020 Brusini et al., Automatic QC and error detection in brain MRI segmentation
Enlarged ventricles from an Alzheimer’s patient had been misclassified and correctly highlighted in the error map. The error map shows a false positive that is related to subject-specific anatomical features. Other areas of the brain are dark and error-free. Identification of a small CSF misclassification in a segmentation that had been visually rated as good. not visible in the error map! Several cortical overestimations are visible, but the present method highlighted only some of them in the error map. Promising results for checking segmentation quality, evaluating its reliability and speeding the error correction up. Limitations: frequent false positives, and some false negatives in correspondence of cortical overestimations. Future work: better post-processing and testing other methods to compare generated and original MRI slices. MIDL Conference 2020 Brusini et al., Automatic QC and error detection in brain MRI segmentation
2. 2. C Classi assificati tion of of s segm egmentati tion qua quality ty Input layer Convolutional layers (128 × 128 × 128 × 2) 1 × 128 1 × 128 GAP p bad … … 128 128 64 64 32 32 Three convolutional blocks made of: 3D Conv (3×3×3) + BatchNorm + ReLU + Binary classification using 10-fold cross-validation on 600 3D Conv (3×3×3, strides=2) + BatchNorm + ReLU randomly selected subjects: 300 with segmentations rated as bad and 300 rated as good. MIDL Conference 2020 Brusini et al., Automatic QC and error detection in brain MRI segmentation
Classi sificati tion perf perform rmance ce High sensitivity (0.96 with a classification threshold of 0.3) is achieved with the proposed classifier, which can be used together with the error maps. Limitation: high presence of false positives (i.e. segmentations wrongly classified as bad). However, some errors were missed by the visual raters, so high sensitivity is here preferable to high specificity. AUC = 0.85 MIDL Conference 2020 Brusini et al., Automatic QC and error detection in brain MRI segmentation
Recommend
More recommend
