A multi-scale network for thyroid segmentation and dose assessment in hyperthyroid patients with uncertainty quantification.

BACKGROUND

Accurate absorbed dose assessment is crucial for personalized hyperthyroidism treatment, but challenges remain in calculating the thyroid dose of 131I using the medical internal radiation dose (MIRD) schema. Further exploration is needed to improve image segmentation for personalized thyroid phantoms, assess uncertainty in S value calculations, and compare dose estimates between the MIRD schema and the Marinelli-Quimby formula.

PURPOSE: This work aims to design a multi-scale deep learning model to enhance accuracy and reliability in absorbed dose calculations for hyperthyroid patients based on the MIRD schema.

METHODS

To improve multi-scale semantic learning and integration, multi-scale residual and multi-head attention modules were added to the UNet++ architecture. An information entropy-based uncertainty quantification module was proposed to assess segmentation confidence and enhance reliability. Patient-specific thyroid voxel phantoms were created from clinical segmentation results, S values were calculated, and overall uncertainty was evaluated. Finally, absorbed dose differences between the MIRD schema and Marinelli-Quimby formula were compared.

RESULTS

The comparative experiments on the publicly available head and neck organ-at-risk CT and MR segmentation (HaN-Seg) dataset showed that the MSRA-UNet++ model excelled in segmenting multiple organs. Specifically, for thyroid segmentation, the proposed model achieved a Dice Similarity Coefficient (DSC) of 88.23% and a Hausdorff Distance 95% (HD95) of 9.43 pixels. When applied on clinical dataset, Jaccard Index (JI) improved 6.17%, compared to UNet++. The uncertainty quantification results showed higher uncertainty at the thyroid edges and low uncertainty in the interior. Patient-specific phantoms were created from segmentation results, and the S value was calculated using Monte Carlo method. Time-integral activity was determined by radioiodine kinetics data. The absorbed dose was evaluated using the MIRD schema and compared with the Marinelli-Quimby formula, which overestimated the MIRD schema by an average of 7.74%.

CONCLUSIONS

The proposed MSRA-UNet++ network provides segmentation results with estimated uncertainty and higher accuracy. The reliability of S values was enhanced by combining segmentation uncertainty with Monte Carlo simulations. To the best of our knowledge, this study is the first to introduce an uncertainty quantification module into the CT image segmentation process for hyperthyroid patients. Additionally, it combines patient-specific thyroid voxel phantoms with individualized radioiodine kinetics data to assess the absorbed dose using the MIRD schema, with the results compared to the Marinelli-Quimby formula. The code will be release on https://github.com/410435553.