Deep learning-based extracranial carotid plaque detection on CBCT images

To evaluate the performance of a deep learning-based model for extracranial carotid plaque detection on cone-beam computed tomography (CBCT) images. CBCT data from 78 patients were retrospectively selected and assessed by three observers. The anonymized image sequences containing extracranial carotid plaques were manually segmented using 3D Slicer software (version 5.6.2) and saved as NIFTI files. The CBCT image sequences containing extracranial carotid plaques and the NIFTI files were imported into a nnU-Net v2-based algorithm, from which 71 cases were selected into the training group and 7 into the test group. In 46 cases, the identified extracranial plaque was unilateral (59%), whereas in 32 cases, it was bilateral (41%). A total of 110 calcifications were identified as distinct entities, with 59 located on the left side (53.6%) and 51 on the right side (46.4%). The performance of the deep learning model was evaluated using the Dice coefficient with a value of 0.56, a sensitivity of 0.53, a Jaccard index of 0.41, and an area under curve (AUC) of 0.76. The trained convolutional neural network achieved the highest Dice coefficient, with a value of 0.78 in the test group on a CBCT image sequence presenting a unilaterally localized extracranial carotid plaque. In terms of the number of samples processed, the algorithm successfully ran on the selected image sequences. However, to increase its performance, further development of the nnU-Net v2-based model is necessary to make it suitable as a decision support system for everyday clinical use.