Region-Specific Masked Loss for Preserving Bone HU in CECT-to-VNC CT Translation

Juhyeong KI, Hyeongmin Jin, Seongmoon Jung, Jimin Lee

Keywords: Dual-energy computed tomography, virtual non-contrast CT, radiotherapy planning

Registration Requirement: Yes

Abstract: Deep learning-based virtual non-contrast computed tomography (VNC CT) can enable accurate dose calculation in radiotherapy planning without additional radiation exposure. However, when dual-energy CT (DECT) derived VNC images are used as training targets, limitations of the DECT-based VNC algorithm can lead to reduced HU values in bone regions. To address this issue, a Region-Specific Masked Loss is proposed using bone masks extracted through a SAM-assisted semi-automated pipeline. A U-Net based bVNC-Net model was trained on DECT data from 24 radiotherapy patients (5,146 slices). The proposed method effectively removed soft-tissue contrast enhancement while preserving bone HU values close to those of the original images.

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 52