DSDFF-YOLO: an automated computer-aided diagnostic method for post-therapy radioiodine whole-body scan.
Authors
Affiliations (3)
Affiliations (3)
- School of Computer Engineering and Science, Shanghai University, Shanghai, China.
- Department of Nuclear Medicine, Shanghai Sixth People's Hospital Affiliated to Shanghai Jiao Tong University School of Medicine, Shanghai, China.
- School of Computer Engineering and Science, Shanghai University, Shanghai, China. [email protected].
Abstract
Radioiodine-131 (<sup>131</sup>I) therapy is an important post-thyroidectomy treatment, but lesion diagnosis on post-therapy whole-body scans (RxWBS) is clinically challenging. This study aimed to propose an automated computer-aided diagnostic method for RxWBS that detects four categories of lesions: thyroid remnants, lymph node metastases (mLNs), lung metastases, and other distant metastases. This retrospective study analyzed data from 361 thyroid cancer patients treated with <sup>131</sup>I between June 2022 and February 2024. We proposed an automated diagnostic method to assist physicians, in which the DSDFF-YOLO model automatically detects and diagnoses lesions on RxWBS. Moreover, we designed the depthwise-separable multi-scale fusion and the dual-control frequency-domain self-attention modules and additionally improved the frequency-domain dynamic convolution. The DSDFF-YOLO was compared with other state-of-the-art and conventional models, as well as with nuclear medicine physicians, to evaluate its diagnostic performance. The validation set included 62 patients (mean age 43.90 ± 13.80 years [SD], 36 female patients). The DSDFF-YOLO outperformed others in detecting thyroid remnants and mLNs, achieving a mAP₅₀ of 0.772, mean precision of 0.801, and recall of 0.731. On the independent test set (51 patients, mean age 44.51 ± 15.51 years [SD], 22 female patients), it surpassed nuclear medicine physicians in detecting thyroid remnants and mLNs, with sensitivities of 85.53% and 79.31%, and specificities of 91.11% and 93.52%, respectively. The DSDFF-YOLO enabled effective and accurate automated auxiliary diagnosis based on RxWBS. This method demonstrated excellent diagnostic performance and holds significant potential for clinical application. Question Can Deep Learning effectively assist clinicians in accurately diagnosing whole-body lesions on post-therapy radioiodine whole-body scans, rather than remaining confined to cervical lesions? Findings The proposed model achieved high sensitivity and specificity in diagnosing lesions on post-therapy radioiodine whole-body scans, outperforming nuclear medicine physicians. Clinical relevance The proposed model was a promising automated diagnostic tool that could assist clinicians in accurately diagnosing lesions on post-therapy radioiodine whole-body scans, potentially reducing patients' need for costly SPECT/CT imaging.