Author: Nuraddeen Nasiru Garba, Kalpana M Kanal, Abdullahi Mohammed, Rabiu Nasiru, Muhammad SHAFIU Shehu, Daniel Vergara, Joseph Everett Wishart 👨🔬
Affiliation: AHMADU BELLO UNIVERSITY, ZARIA, University of Washington 🌍
Purpose: To establish local Diagnostic Reference Levels (DRLs) for head and neck computed tomography (CT) exams in Abuja, Nigeria, and to investigate the performance of brain metastasis (BM) and brain lesion (BL) segmentation techniques using U-Net models.
Methods: This study analyzed head and neck CT images from 374 oncology patients (303 adults and 71 pediatrics) at National Hospital Abuja, Nigeria. Images were acquired with a Canon Aquilion Lightning 32-slice CT scanner. The volume CTDI (CTDIvol), size-specific dose estimate (SSDE) and water-equivalent diameter (WED) were calculated using IndoseCT (Indonesia). Calculated local DRLs for both adult and pediatric patients were compared with previously published DRLs. Additionally, BM and BL segmentation algorithms were applied to adult CT images using established U-Net models with varying numbers of layers and optimized contouring thicknesses. To create a reliable ground truth dataset, contouring of brain regions in DICOM images was performed under radiologist supervision. The Dice Similarity Coefficient (DSC) was used to determine the optimal contouring thickness for image quality and model detectability. A custom-designed U-Net model was trained with a 4-to-8 encoder decoder layer, tailoring the architecture to our needs to enhance the model's performance.
Results: Adult and Pediatric DRLs compared well with other published DRLs. Pediatric DRLs generally increased with patient age. The current model struggles to accurately predict tumors alone, often incorrectly identifying bone structures as tumors. Future improvements are needed to enhance tumor segmentation accuracy.
Conclusion: Preliminary results suggest that local pediatric protocols were modified according to patient habitus, while adult protocols were fixed. Both adult and pediatric protocols yielded similar DRLs to previously established metrics. Future work would extend this analysis to include more CT scanners and regional data. The 4-layer U-Net models with 0.5 mm contouring thickness demonstrated a suitable BM segmentation, while BL datasets require further optimization.