Improving Segmentation of the Inferior Alveolar Nerve through Deep Label Propagation

Abstract

Many recent works in dentistry and maxillofacial imagery focused on the Inferior Alveolar Nerve (IAN) canal detection. Unfortunately, the small extent of available 3D maxillofacial datasets has strongly limited the performance of deep learning-based techniques. On the other hand, a huge amount of sparsely annotated data is produced every day from the regular procedures in the maxillofacial practice. Despite the amount of sparsely labeled images being significant, the adoption of those data still raises an open problem. Indeed, the deep learning approach frames the presence of dense annotations as a crucial factor. Recent efforts in literature have hence focused on developing label propagation techniques to expand sparse annotations into dense labels. However, the proposed methods proved only marginally effective for the purpose of segmenting the alveolar nerve in CBCT scans. This paper exploits and publicly releases a new 3D densely annotated dataset, through which we are able to train a deep label propagation model which obtains better results than those available in literature. By combining a segmentation model trained on the 3D annotated data and label propagation, we significantly improve the state of the art in the Inferior Alveolar Nerve segmentation.

SISTR: Sinus and Inferior alveolar nerve Segmentation with Targeted Refinement on Cone Beam Computed Tomography images

“…Sinus segmentation achieved a strong precision, as evidenced by a 0.972 Dice score and a Chamfer boundary distance of 0.29 mm, which is lower than voxel precision of 0.4 mm. The IAN, a structure known for its challenging segmentation due to subjective annotation 25 , was segmented with a 0.855 Dice score and a Chamfer boundary distance of 0.63 mm.…”

“…This was based on 256 synthetic annotations, 68 actual annotations, and an internal test set of 15 CBCTs, although limited by the use of a single data source. In contrast, Usman et al 24 reported Dice scores of 0.751 and 0.77 on an internal test set of 500 CBCTs from a unique center and on Cipriano et al’s dataset 25 , respectively, employing a development set of 500 densely annotated CBCTs from a single center. While these results are noteworthy, they are in a different range of magnitude compared to our achieved Dice score of 0.843 (range: 0.810-0.856), underscoring the potential superiority of our method.…”

“…The literature acknowledges the complexity of IAN segmentation, largely due to considerable inter-observer and intra-observer variability; for instance, Järnstedt et al 14 reported an inter-observer variability of 0.77 mm, benchmarked against a gold standard set by experienced radiologists. Comparative studies like Cipriano et al’s 25 and Usman et al’s 24 used internal test sets with limited diversity, primarily from one or two centers, often overlapping with their development sets. This lack of external validation and diversity in test datasets limits the generalizability of their findings.…”

Dual-Stage Deeply Supervised Attention-Based Convolutional Neural Networks for Mandibular Canal Segmentation in CBCT Scans

“…They exploited their developed dataset to achieve state-of-the-art performance by employing a deep learning-based method proposed by Jaskari et al [20]. The same authors extended their work in [28] by leveraging their dataset with 3D dense annotations to train a deep label propagation model which outperformed the previous techniques.…”

“…We only used densely annotated scans for training. The data distribution was 76 scans for training and 15 for testing, which is the same distribution as used in [28]. Table 6 summarizes the results from our study and two previously published studies.…”