Introduction
Papillary thyroid microcarcinoma is frequently detected and often managed with active surveillance to avoid overtreatment. Early identification of thyroid capsule invasion signals progression that may precede extrathyroidal spread and can influence surgical planning. Ultrasound is the primary imaging tool for thyroid assessment yet reporting of capsule involvement lacks standardised criteria and can be inconsistent. A deep learning approach that fuses peri-tumoural radiomics with intra-tumoural ultrasound features has been developed to predict capsule invasion. Evaluation on retrospectively collected data from multiple centres, with performance benchmarked against radiologists at different experience levels, demonstrates high diagnostic accuracy and consistent improvements when used as an assistive tool. These capabilities support earlier image-based risk stratification in patients considered for surveillance or surgery.
Integrated Ultrasound Modelling
The dataset included 964 papillary thyroid microcarcinomas in 964 patients drawn from three centres. Within the internal cohort of 880 cases, lesions were partitioned 7:1:2 into training (n = 615), validation (n = 88) and internal test (n = 177) sets, with a separate external test set of 84 cases. Peri-tumoural radiomics were extracted from concentric rings surrounding each nodule. Among these, features from a 30% expansion zone achieved the best standalone performance with a support vector machine, yielding an area under the curve of 0.795 for classifying capsule invasion.
These peri-tumoural features were then fused with intra-tumoural features from ultrasound images and fed into deep learning backbones including DenseNet-121, InceptionV3, ResNet-50 and SwinTransformer. Across architectures, combined peri- plus intra-tumoural inputs outperformed intra-tumoural inputs alone, underscoring the diagnostic value of peri-capsular context. SwinTransformer delivered the strongest results. On the internal test set it reached an area under the curve of 0.923 with sensitivity 0.847, specificity 0.771 and accuracy 0.802. On the external test set it achieved an area under the curve of 0.892 with accuracy 0.795. Incorporating peri-tumoural radiomics improved SwinTransformer area under the curve from 0.804 to 0.923 internally and from 0.792 to 0.892 externally, highlighting consistent gains across cohorts. Class activation mapping showed attention concentrated along peri-capsular regions in invasion-positive cases, offering visual support for model interpretability.
Must Read: ESR Practice Recommendations for Thyroid Imaging
Performance Against Radiologists
Six radiologists—two senior, two attending and two junior—reviewed static ultrasound images without clinical context to classify capsule invasion, first unaided then with model assistance after a washout period. On the external set, unaided area under the curve values were 0.720 and 0.725 for senior readers, 0.685 and 0.665 for attending readers and 0.643 and 0.635 for junior readers. All were lower than the SwinTransformer’s 0.892 on the same images.
With model assistance, reader performance improved. Area under the curve increased to 0.796 and 0.790 for senior readers, 0.758 and 0.788 for attending readers and 0.752 and 0.676 for junior readers. One attending reader showed a notable improvement from 0.665 to 0.788. These findings indicate that the tool can raise diagnostic accuracy across experience levels and reduce variability in assessment when applied to static ultrasound images. The assisted gains, alongside the model’s standalone performance, suggest a role as an adjunct that standardises capsule invasion evaluation and supports more consistent reporting.
Clinical Signals and Study Boundaries
Baseline comparisons revealed clinical and imaging associations relevant to risk assessment. In the internal cohort, the proportion aged 55 years or older was higher among capsule invasion cases than non-invasion cases. Capsular contact on ultrasound appeared more frequently when invasion was present. Lymph node metastasis was also more common in the invasion group. In the external cohort, nodules larger than 5 mm were more prevalent among invasion-positive cases. By contrast, TI-RADS distributions were broadly similar between invasion and non-invasion groups across cohorts. Together, these signals reinforce the clinical value of detecting capsule invasion early, including for planning intraoperative lymph node evaluation given the known low sensitivity of ultrasound for central compartment disease.
The authors acknowledged several constraints. Images were static rather than cine, and the analysis drew on three centres, which may limit generalisability. Lesion segmentation was manual, and there was no stratification by ultrasound equipment. Depth of invasion was not graded, and background thyroid pathology was incompletely captured. These factors point to the need for prospective, multi-device validation and exploration of workflow integration to understand performance in routine practice.
An integrated deep learning approach that fuses peri-tumoural radiomics with intra-tumoural ultrasound features can identify thyroid capsule invasion in papillary thyroid microcarcinoma with high accuracy and can improve radiologist performance when used as an assistive tool. By providing a more standardised, image-based assessment of capsule involvement, the method supports earlier risk stratification and more tailored intervention planning for patients under surveillance or being considered for surgery. The reported limitations define a clear agenda for prospective validation, device-level testing and automation of segmentation to consolidate clinical utility without adding burdens to imaging workflows.
Source: Insights into Imaging
Image Credit: iStock
