A proof-of-concept study of multitask learning for cranial synthetic CT generation across heterogeneous MRI field strengths.

BACKGROUND

Accurate synthesis of computed tomography (CT) images from magnetic resonance imaging (MRI) is clinically valuable for cranial applications such as attenuation correction, radiotherapy planning, and image-guided interventions. However, the heterogeneity of MRI data across magnetic field strengths and pulse sequences limits the generalizability of existing methods, posing a barrier to clinical translation.

PURPOSE: This study aims to reformulate cranial CT synthesis as a modular, structurally coupled problem using a deep learning approach, enhancing adaptability across heterogeneous MRI conditions, including different field strengths and sequence protocols.

METHODS

We implemented a cascaded multitask pipeline that jointly models skull segmentation and Hounsfield Unit (HU) regression in anatomically targeted regions. A 3D patch-based training paradigm shifts the modeling focus from global image translation to localized bone feature extraction. The backbone leverages a residual Mamba-based state space model within a 3D U-Net structure to improve spatial representation, while a Transformer U-Net serves as a widely recognized reference baseline comparison. The model was trained and evaluated in terms of multi-modal (T1-weighted and T2-FLAIR) and cross-domain transferability from a public 1.5 Tesla brain dataset (n = 37) to an independent 7 Tesla clinical brain dataset (n = 44). Performance was assessed using Dice and Jaccard indices for segmentation accuracy and mean absolute error (MAE) for HU regression.

RESULTS

Quantitative analysis showed that our multitask pipeline, incorporating both morphological and HU map prediction stages, significantly outperformed conventional direct MRI-to-CT mapping across all metrics on the public 1.5T dataset (p < 0.05). Consistent performance on an external 7T clinical dataset (p < 0.001) further demonstrated the method's robustness and adaptability across field strengths.

CONCLUSION

These findings support that task-structured training for modality transformation markedly improves both accuracy and generalizability of cranial CT synthesis across heterogeneous MRI conditions. The observed consistency across field strengths validates the robustness of the proposed methodology.