Papillary thyroid carcinoma (PTC) is the most common thyroid cancer, yet its behaviour varies widely from patient to patient. Risk is often determined after surgery, which constrains preoperative planning and can lead to unnecessary procedures. A multicentre effort spanning four hospitals proposes a preoperative approach that combines computed tomography (CT), ultrasound, clinical characteristics and immunological markers within a single framework. By mapping tumour subregions as distinct “habitats” and integrating their signatures with key patient data, the method aims to distinguish low-risk disease suitable for active surveillance from cases that are more likely to require active treatment. The study draws on a large cohort and evaluates performance across multiple external sites, with an emphasis on practical implementation in clinical settings.
Must Read: ESR Practice Recommendations for Thyroid Imaging
Cohorts, Imaging Habitats and Key Predictors
The programme enrolled more than a thousand adults with pathologically confirmed PTC, dividing participants into a training group, an internal validation group and several external validation cohorts from separate centres. All underwent preoperative CT and ultrasound under harmonised quality protocols. Tumours were segmented by experienced radiologists, with agreement checked on a subset to ensure reproducibility.
Ultrasound images were over-segmented into superpixels and clustered to reveal four habitats with distinct echogenic patterns. One habitat, characterised by heterogeneous hypoechogenicity, showed the strongest link to higher-risk classification when its radiomic features were used for modelling. CT images were decomposed into five habitats using supervoxel segmentation and two-stage clustering, then quantified with a multi-scale index (MSI) set that captured spatial heterogeneity, boundary interactions and texture. Together, these modalities portrayed the internal architecture of each tumour rather than treating it as uniform, reflecting differences that correspond to biological heterogeneity.
Clinical and immunological variables were reviewed for independent association with higher-risk status. Chronic lymphocytic thyroiditis, tumour size and platelet-to-lymphocyte ratio were retained as key predictors. These factors complement the imaging-derived habitats by capturing background thyroid inflammation, tumour burden and systemic immune-inflammation balance, offering a broader view of disease behaviour before any operative step.
Fusion Modelling with Consistent Validation
Single-modality models based on ultrasound or CT habitats each achieved solid discrimination. The ultrasound habitat linked to aggressive behaviour performed well across training and validation, and the CT MSI habitat model also maintained strong results across centres. To exploit complementary strengths, three multimodal fusion strategies were tested: early fusion by concatenating features, late fusion by combining model outputs, and an ensemble approach using a soft-voting classifier.
Across internal and external cohorts, the ensemble fusion model repeatedly delivered the highest discrimination, with area under the curve around the mid to high 0.9 range and accuracy consistently above the high eighties. Decision curve analysis indicated clinical benefit across a broad span of threshold probabilities, supporting use in settings that weigh false positives and false negatives differently. These results were achieved not only in the originating centre but also across multiple independent hospitals, underscoring generalisability within the tested population and imaging protocols.
Interpretability was addressed using SHapley Additive exPlanations. Feature importance plots highlighted contributions from both CT and ultrasound habitats, alongside clinical and immunological markers. A top CT MSI feature that summarised multi-scale spatial complexity, a texture descriptor from ultrasound indicating disordered tissue organisation, and platelet-to-lymphocyte ratio emerged as leading drivers of risk classification. Case-level decision plots illustrated how specific features pushed individual predictions towards a low-risk or intermediate/high-risk label, providing transparency for multidisciplinary teams.
Clinical Utility, Transparency and Limitations
To support adoption, a web-based interface was developed using a lightweight framework. Clinicians can upload ultrasound and CT images, enter clinical parameters and receive an immediate risk estimate that stratifies cases into low-risk or intermediate/high-risk groups. Example outputs demonstrated clear separation of probabilities with accompanying summaries. The platform aligns with preoperative pathways that aim to avoid unnecessary surgery in patients suitable for active surveillance, while flagging cases that are more likely to need surgery and, where relevant, postoperative radioactive iodine based on subsequent pathology.
Several practical limitations are acknowledged. Tumour segmentation currently requires manual delineation, which introduces operator dependence despite quality controls. Imaging was acquired using standardised but not identical protocols, and the patient population was ethnically homogeneous, which may limit transferability to other regions, healthcare systems and imaging technologies. Although performance remained stable across sites in the study network, broader validation would be required to assess robustness in multi-ethnic cohorts and to confirm stability across different scanners and acquisition parameters. The link between imaging habitats and underlying pathology would also benefit from correlative studies using histopathology, immunohistochemistry and spatial molecular methods. Automation of segmentation with deep learning approaches is a planned step to streamline workflow and reduce variability.
Combining habitat imaging with multimodal analysis offers a robust preoperative method to stratify PTC risk using routinely obtainable ultrasound, CT and simple blood-derived markers. By focusing on tumour subregions and integrating them with clinical context, the approach improves discrimination between low-risk disease appropriate for monitoring and cases more likely to require intervention. Consistent performance across multiple centres, transparent model explanations and an accessible web interface support potential clinical use, with future work aimed at automation, pathological correlation and validation in diverse populations.
Source: Insights into Imaging
Image Credit: iStock
