Multiphase contrast-enhanced CT (CECT) is widely used to evaluate renal tumours, yet several treatment-relevant factors remain difficult to determine before surgery. Histological subtype, clinical stage and anatomical complexity can influence surgical planning and clinical decision-making, but these elements are not always straightforward to establish from imaging alone. Pathology and TNM staging are typically obtained through invasive assessment, while anatomical scoring systems such as the R.E.N.A.L. nephrometry score rely on reader interpretation and can vary between observers. A multi-task deep learning approach has been developed to provide a more integrated preoperative assessment from routine CECT, aiming to generate multiple clinically relevant outputs with a single model rather than separate tools for each task.
Must read:Deep Learning US Model Predicts Thyroid Capsule Invasion
One Model Connecting Histology, Stage and Complexity
A two-centre retrospective dataset included several hundred patients with solid malignant renal tumours who underwent nephrectomy and had preoperative multiphase CECT imaging available. Clear cell renal cell carcinoma (ccRCC) represented the largest proportion of cases. The dataset also included a smaller subset of patients with more advanced clinical stages and another subset with high anatomical complexity. Clinical stage groupings were aligned with TNM-based categories, with localised disease separated from more advanced disease. Anatomical complexity was defined using the R.E.N.A.L. nephrometry score and categorised into low, moderate and high groups, with a focus on separating high complexity from lower grades due to its relevance for surgical risk and operative planning.
Multi-Task Architecture and External Validation
The multi-task system used a progressive layered extraction (PLE) framework, allowing shared feature learning while preserving task-specific outputs. A 3D ResNet backbone extracted features from three contrast phases of CT imaging, and these were combined before entering the multi-task module. Tumour segmentation formed part of the workflow, using an automated nnU-Net approach adapted from an existing dataset, followed by radiologist review and manual corrections. Segmentation accuracy was reported as high, with performance differing between the two centres.
The model was trained and internally validated on data from one hospital, while an independent external test set from a second hospital was used to evaluate generalisability. Across the three outputs, performance was reported as strong, with AUC values in the high 0.8 range for distinguishing ccRCC from non-ccRCC, separating localised from advanced clinical stage and classifying low-to-intermediate versus high anatomical complexity. A comparison was also made with three separate single-task deep learning models using the same backbone. The most notable difference was seen in clinical staging in the external test cohort, where the multi-task framework achieved higher overall performance than the single-task approach.
Comparing Model Outputs with Radiologists and Runtime Efficiency
Model performance was compared with radiologist assessment using a multi-reader observer study. Five radiologists with varying abdominal imaging experience evaluated the external test cases and estimated probabilities for histological subtype, advanced stage and high complexity. The multi-task model outperformed the least experienced readers for histology classification and clinical staging. For anatomical complexity grading, the most experienced radiologists performed better than the model, highlighting that surgical complexity scoring remained a challenging target for automated prediction in this setting.
Operational efficiency was presented as an additional advantage. Running three single-task models required substantially more memory than a single multi-task model producing all outputs. The multi-task model also achieved faster inference per case. Decision curve analysis suggested that the multi-task approach delivered higher net benefit across a broader range of thresholds for clinical staging compared with the single-task alternative. Several limitations were noted, including the retrospective design, reliance on surgical cohorts and the need for broader validation beyond one external centre, particularly for less common groups such as advanced-stage disease or high complexity tumours.
A multi-task deep learning model based on multiphase CECT provided a combined preoperative assessment of renal tumour histology, clinical stage and anatomical complexity within a single framework. Performance remained consistent in external testing, with particularly strong results for clinical staging when compared with a matched single-task approach. The model showed advantages over less experienced radiologists for certain tasks, while senior radiologists remained stronger for anatomical complexity scoring. Lower resource requirements and faster runtime supported the feasibility of deploying an integrated tool for preoperative decision support, while the reported limitations underlined the need for further validation in broader patient cohorts and clinical settings.
Source: European Radiology
