Sacral tumours represent a diagnostically complex subset of bone neoplasms, with management decisions highly dependent on accurate preoperative classification. Tumour type influences surgical strategy, the use of chemotherapy or radiotherapy and long-term prognosis. However, sacral lesions are often detected late due to nonspecific symptoms, and histological confirmation commonly relies on biopsy, an invasive procedure associated with sampling error and complications. Noncontrast computed tomography (NCCT) is routinely used for initial assessment, particularly for evaluating bone destruction, calcification and ossification. Despite its widespread availability, visual interpretation of NCCT images remains challenging, especially given the rarity and heterogeneity of sacral tumours. Against this background, automated image analysis using artificial intelligence has emerged as a potential means to support radiological decision-making, reduce variability and improve diagnostic accuracy in preoperative workflows.
Automated Segmentation Using a Hip Bone Reference
The proposed pipeline integrates a fully automated segmentation process as its first step, addressing one of the main barriers to clinical adoption of artificial intelligence in musculoskeletal imaging. Manual segmentation of tumours is time-consuming and prone to interobserver variability, limiting feasibility in routine practice. In this approach, deep convolutional neural networks automatically segment both sacral tumours and hip bones from preoperative NCCT images. Postprocessing is applied to preserve the largest connected region, reducing artefacts and ensuring that a single tumour or hip bone is retained.
A distinctive feature of the method is the use of the hip bone as a reference frame for tumour localisation. Rather than relying on voxel-based coordinates, which may vary with scan coverage and acquisition protocols, tumour position is defined relative to the segmented hip bone. Tumour centroids are calculated within a standardised coordinate system, allowing spatial information to be incorporated consistently across patients and centres. This strategy improves the robustness of spatial features and enables reliable integration of location data into subsequent classification tasks.
Segmentation performance demonstrated high reproducibility across validation, internal test and external test sets, with Dice coefficients remaining stable after postprocessing. Agreement between radiologists was also high, supporting the reliability of the reference annotations used during model development. Together, these findings indicate that automated segmentation can provide a dependable foundation for downstream classification without the need for manual intervention.
Must Read:CT Score Uses Preoperative Features to Flag Recurrence Risk
Multimodal Classification of Sacral Tumour Types
Following segmentation, the pipeline applies a six-class classification model designed to differentiate metastatic tumours, chondrosarcoma, osteosarcoma, chordoma, neurogenic tumours and giant cell tumours. The classifier integrates three categories of input: cropped NCCT images, clinical variables and tumour location relative to the hip bone. Clinical data include age, sex and tumour volume, all of which are standardised to ensure consistent weighting during training.
The classification architecture is based on a deep learning framework that fuses imaging features extracted from NCCT with non-imaging inputs. By combining these data streams, the model addresses limitations observed in image-only approaches, which often struggle to distinguish tumours with overlapping radiological appearances. Comparative experiments showed that models incorporating both clinical and location information outperformed those relying on a single data type.
Performance remained consistent across validation, internal and external test cohorts, demonstrating generalisability beyond the development dataset. The model achieved its strongest results in identifying neurogenic tumours and chordomas, while performance for chondrosarcoma was comparatively lower, reflecting the known heterogeneity and imaging similarity of its pathological subtypes. Importantly, the model maintained balanced performance across all six categories, despite inherent class imbalance associated with rare tumours.
Comparison With Radiologist Interpretation
To assess clinical relevance, classification results were compared directly with independent radiologist readings on an external test cohort. Radiologists were provided with NCCT images and basic clinical information but were blinded to histopathological outcomes. Under these conditions, the automated model demonstrated higher overall discriminatory performance than expert readers, particularly in the identification of metastatic tumours.
Metastatic lesions often present diagnostic difficulty due to variable imaging characteristics and their role as an exclusion diagnosis in clinical practice. The automated approach showed improved sensitivity and precision for this category, suggesting potential value as a decision-support tool in challenging cases. While radiologist performance remained comparable for several tumour types, the automated system provided more consistent results across the full classification spectrum.
The comparison highlights the complementary role of artificial intelligence rather than replacement of clinical expertise. By offering reproducible, noninvasive classification based on routinely acquired NCCT images, the pipeline may support radiologists in complex diagnostic scenarios and reduce reliance on invasive biopsy in selected cases.
A fully automated NCCT-based pipeline combining segmentation and multimodal classification demonstrated robust performance in differentiating six sacral tumour types across multiple centres. By using the hip bone as a spatial reference and integrating imaging, clinical and location data, the approach addresses key limitations of existing image-only models and manual workflows. Performance exceeded that of radiologists in overall classification accuracy, with particular gains in identifying metastatic tumours. These findings underline the potential of automated, noninvasive tools to enhance preoperative assessment, support personalised treatment planning and improve efficiency in the management of rare and complex sacral tumours.
Source: Radiology: Imaging Cancer
Image Credit: iStock
