Automated Standardization of the Shear-wave Elastography Effective Region Using DeepLabV3 Improves Malignancy Assessment in Thyroid Nodules: A Single-institution Multicampus Study.

Rationale

AND

Objectives

To assess whether a DeepLabV3-based model can standardize the shear-wave elastography (SWE) effective region in thyroid nodules and improve maximum Young's modulus (Emax)-based malignancy assessment.

Materials And Methods

This retrospective single-institution multicampus study included 678 training and 280 intercampus validation nodules. A DeepLabV3 model using registered B-mode and SWE images as two-channel input automatically delineated the SWE effective region. In the validation cohort, model-derived and reader-defined regions were compared with a Simultaneous Truth and Performance Level Estimation-derived expert consensus using intersection over union (IoU), Dice coefficient, and 95th percentile Hausdorff distance (HD95). Emax was extracted from each region, and diagnostic performance was evaluated using pathologic diagnosis (cytology or histopathology) as the reference standard.

Results

DeepLabV3 showed better agreement with the consensus reference than the senior and junior readers, with higher IoU (0.82±0.04 vs. 0.76±0.09 and 0.67±0.14) and dice (0.90±0.15 vs. 0.86±0.11 and 0.80±0.12), and lower HD95 (2.39±0.81 mm vs. 3.61±1.44 mm and 4.23±1.19 mm; all p < 0.01). Emax_DL achieved a higher area under the curve than Emax_R2 and Emax_R1 (0.86 [95% confidence interval: 0.80-0.91] vs. 0.78 [0.74-0.86] and 0.73 [0.67-0.81]), with a likelihood ratio of 3.56 at the optimal cutoff.

Conclusion

Automated standardization of the SWE effective region using DeepLabV3 may provide a more consistent and less operator-dependent basis for Emax measurement, with potential to improve ultrasound-based assessment of thyroid nodules.