Renal masses are increasingly detected incidentally on CT scans performed for unrelated indications. Most are benign or indolent, but a meaningful fraction are malignant and aggressive, requiring intervention. The clinical challenge is distinguishing which masses warrant surgery, active surveillance, or observation, as overtreatment of benign masses carries operative morbidity while undertreatment of aggressive malignancies risks metastatic progression.
Current imaging interpretation relies on radiologist judgment applied to CT enhancement patterns, morphological features, and size metrics. While experienced radiologists perform reasonably well, diagnostic accuracy is imperfect and subject to inter-observer variability. For smaller or atypical masses, the confidence gap between expert and community radiologist assessment can be substantial, leading to inconsistent clinical management decisions across different healthcare settings.
Beyond malignancy detection, the biological aggressiveness of confirmed RCC tumors is critical for surgical planning and post-operative management. Aggressive tumors with high metastatic potential may warrant more radical surgery, lymph node dissection, or enrollment in adjuvant therapy trials, while indolent tumors may be safely managed with nephron-sparing approaches and standard surveillance.
AI-based image analysis systems that can simultaneously assess malignancy likelihood and biological aggressiveness from routine CT scans would provide a transformative clinical tool. This study developed and prospectively validated two such CNN models using data from over 4,500 patients and 13,000 CT volumes from six Chinese centers and TCIA.
The dataset comprised 4,557 patients with renal masses and 13,261 CT volume acquisitions from six Chinese tertiary hospitals and the TCIA (The Cancer Imaging Archive). Patients had pathologically confirmed diagnoses following surgery, providing ground truth labels for training and evaluation. The dataset was split into development, internal validation, and prospective validation sets.
Tumor segmentation was performed using nnU-Net, a self-configuring semantic segmentation network that automatically adapts its architecture and preprocessing to the characteristics of the input data. nnU-Net achieved a Dice Similarity Coefficient (DSC) of 0.852 for renal mass segmentation, enabling automated delineation of tumor regions without manual contouring for the majority of cases.
Two separate CNN models were built using ResNet-18 as the backbone architecture. ResNet-18's residual connections facilitate learning deep feature representations while avoiding the vanishing gradient problem. Both models were trained to analyze multi-phase CT data (unenhanced, corticomedullary, nephrographic, and excretory phases), enabling the networks to capture dynamic contrast enhancement patterns that are among the most biologically informative features of renal masses.
Model 1 was trained to classify malignancy (malignant vs. benign). Model 2 was trained to classify aggressiveness, distinguishing pathologically aggressive tumors (high ISUP grade, advanced stage, metastatic behavior) from indolent ones. Training labels for aggressiveness were derived from surgical pathology reports, with aggressiveness defined by histological grade, lymphovascular invasion, coagulative necrosis, and sarcomatoid differentiation.
Model 1 (malignancy classification) achieved an AUC of 0.871 in the prospective validation set. When compared head-to-head with seven radiologists of varying experience levels, the AI model outperformed all seven in terms of AUC. This performance advantage was most pronounced over community radiologists, while narrowing to near-equivalence with senior subspecialty radiologists, suggesting the model encodes expert-level pattern recognition in malignancy assessment.
Model 2 (aggressiveness classification) achieved an AUC of 0.783 in prospective validation. This represents a more difficult classification task, as aggressiveness is a continuous biological property being dichotomized, and its imaging correlates are more subtle than the malignancy vs. benign distinction. Nonetheless, AUC 0.783 demonstrates clinically meaningful discriminative ability for pre-surgical aggressiveness assessment from CT alone.
The AI-derived aggressiveness score outperformed TNM staging and WHO/ISUP grade in C-index for survival prediction. This is a particularly striking finding because TNM and ISUP grade are the current gold standards for post-operative risk stratification. An imaging-based AI score that surpasses these post-surgical pathological metrics when assessed pre-operatively would be transformative for treatment planning.
Five-year disease-specific survival (DSS) rates differed dramatically between AI-classified groups: 97.9% for AI-indolent tumors vs. 86.0% for AI-aggressive tumors. This survival difference, defined pre-operatively from CT images, provides clinicians with actionable information before the surgical decision is made, rather than after pathology returns.
To understand the biological basis of AI-classified aggressiveness, the researchers linked CT-based aggressiveness scores to genomic and immune profiling data for a subset of patients with available tissue analysis. AI-aggressive tumors showed significantly higher rates of SETD2 and AHNAK2 mutations compared to AI-indolent tumors. SETD2, a histone methyltransferase, is one of the most commonly mutated genes in high-grade ccRCC and is associated with genomic instability and poor prognosis.
AHNAK2 mutations, while less well-characterized than SETD2, have been associated with poor outcomes in multiple cancer types and may contribute to altered cell adhesion and metastatic dissemination. Their enrichment in AI-aggressive tumors suggests the model is detecting imaging phenotypes that correlate with underlying genomic instability rather than capturing surface-level morphological features alone.
Immune profiling of aggressive tumors revealed a striking immunosuppressive pattern. CD8+ cytotoxic T cells were elevated in AI-aggressive tumors, but so were FOXP3+ regulatory T cells (Tregs), suggesting that increased immune infiltration is counterbalanced by immune suppression mechanisms. The elevated Treg-to-CD8+ ratio in aggressive tumors is consistent with immune exclusion rather than effective anti-tumor immunity.
This co-elevation of CD8+ T cells and Tregs in AI-aggressive tumors may explain why these patients do not benefit from immune infiltration: the immunosuppressive Treg compartment neutralizes cytotoxic T cell function, creating a dysfunctional immune microenvironment that fails to control tumor growth. This pattern is associated with poor immunotherapy response and provides a molecular explanation for the poor DSS in the AI-aggressive group.
The most immediate clinical application of these AI models is pre-surgical decision support. A pre-operative CT scan is already standard of care for renal mass evaluation. Applying the AI malignancy and aggressiveness models to this existing imaging adds no additional patient burden while providing probability estimates that can inform surgical planning discussions.
For malignancy prediction, the AI model could support the clinical decision about whether to proceed with surgery vs. active surveillance for small indeterminate masses. High malignancy probability from AI, combined with clinical factors such as patient age, comorbidities, and tumor size, could help justify intervention in cases where radiologist visual assessment is equivocal.
For aggressiveness prediction, patients classified as AI-aggressive prior to surgery could be considered for more extensive lymph node dissection, discussed for adjuvant therapy clinical trial enrollment, or counseled differently regarding expected outcomes. The ability to have this conversation before surgery, using the AI aggressiveness score, enables more personalized pre-operative informed consent and treatment planning.
Patients whose masses are classified as AI-indolent but are still recommended for surgery (e.g., due to size or growth on surveillance) could be prioritized for nephron-sparing partial nephrectomy, confident that the biological risk of local recurrence from incomplete margins is low. This reduces overtreatment of biologically quiet tumors with radical nephrectomy.
This study represents one of the largest and most comprehensive prospective validations of CT-based AI for renal mass characterization to date. The two-model framework simultaneously addresses malignancy detection and aggressiveness grading, providing a complete pre-operative biological profile from a single CT acquisition.
The finding that the AI aggressiveness score outperforms post-surgical pathological staging in C-index for survival is a remarkable claim that, if confirmed in additional independent cohorts, would redefine the role of pre-operative imaging in RCC management. It suggests that CT harbors prognostic information that pathological staging cannot fully capture, possibly because CT integrates information about the entire three-dimensional tumor volume rather than the sampled regions analyzed in pathology.
The molecular and immune validation connecting CT-derived aggressiveness scores to SETD2 mutations and Treg infiltration provides mechanistic credibility to the AI predictions. This radiogenomic linkage demonstrates that the CNN is not pattern-matching arbitrary visual features but is detecting imaging phenotypes that reflect real biological processes of genomic instability and immune evasion.
Future directions include international multicenter validation across diverse imaging systems and patient populations, integration with liquid biopsy or genomic data to create composite pre-operative risk scores, and prospective clinical trials testing whether AI-guided treatment selection improves survival outcomes in patients with renal masses requiring management decisions.