Assessment of consistency detection in mandibular first premolars root canal morphology using U-NET model: Evaluation with dice coefficient and intersection over union – a pilot study / Melissa Wong Li Zheng

Melissa Wong , Li Zheng (2024) Assessment of consistency detection in mandibular first premolars root canal morphology using U-NET model: Evaluation with dice coefficient and intersection over union – a pilot study / Melissa Wong Li Zheng. Masters thesis, Universiti Malaya.

	PDF (The Candidate's Agreement) Restricted to Repository staff only Download (138Kb)
	PDF (Thesis M.A.) Download (1176Kb)

Abstract

This study assesses the consistency of the U-Net model in pulp space segmentation of extracted mandibular first premolars (MFPs) using CBCT images. Five training, validation, and testing datasets were randomly generated from a pool of 130 CBCT images. Mean dice coefficient (DC) and Intersection over Union (IoU) scores were measured and compared across these datasets. The CBCT images were loaded into ITK-SNAP software (Version 4.0.2) for semi-automatic segmentation, performed by a postgraduate student (MW) with an ICC of 0.95. The datasets were split into five different training (70%), validation (20%), and testing (10%) sets from the same data pool. Image augmentation was applied, and the files were resized for the U-Net model. The mean DC and IoU scores for the training, validation, and testing datasets were compared using one-way ANOVA. This was followed by a post-hoc LSD test for multiple comparisons between the groups. A p-value < 0.05 was considered statistically significant. The mean DC from datasets 1 and 5 showed no significant difference during training, validation and testing phase with the p value of 0.32, 0.20, and 0.06 respectively. Similarly, the mean IoU from datasets 1 and 5 showed no significant difference during training, validation, and testing phase with the p value of 0.14, 0.23 and 0.12 respectively. In the training phase, the mean DC for dataset 2 (0.39  0.04) was significantly higher than dataset 1, 3, 4 and 5 (p = 0.00). Similarly, during the testing phase, the mean DC for dataset 2 (0.34  0.05) was significantly higher than dataset 1, 4 and 5 (p = 0.00), and dataset 3 (p = 0.02). The similar p-values for datasets 2 compared to datasets 1, 4 and 5 suggest that the performance differences between these datasets are consistent and significant in both the training and testing phases. The difference in p-values between datasets 2 and 3 during training and testing indicates that the performance gap was more pronounced during training than testing phases. The mean IoU for dataset 2 (0.40  0.07) performed significantly higher than dataset 1 (0.17  0.08, p = 0.00), 3 (0.26  0.10, p = 0.00), 4 (0.32  0.06, p = 0.04) and 5 (0.22  0.09, p = 0.00) during the training phase. Comparably, the mean IoU during the testing phase for dataset 2 (0.35  0.05) was significantly higher than dataset 1 (0.13  0.05, p = 0.00), 3 (0.27  0.07, p = 0.01), 4 (0.17  0.08, p = 0.00) and 5 (0.18  0.06, p = 0.00). The findings indicate that the consistency of the U-Net model in pulpal space segmentation of MFPs was affected despite the five datasets being randomly generated from the same data pool. When the predicted and ground truth images overlap completely, the DC and IoU value is 1. In the present study, the overall mean DC and IoU values across all phases were below the ideal value of 1, which indicates that the U-Net model performance in pulp segmentation is less consistent in the assessment of MFP.

Actions (For repository staff only : Login required)