Renal cell carcinoma (RCC) is among the most common urological malignancies, and its incidence has been rising steadily worldwide. Early and accurate pathological diagnosis is essential for guiding treatment decisions, particularly distinguishing malignant from benign renal tumors before or during surgery.
Traditionally, pathological confirmation requires microscopic examination of biopsy or surgical specimens, a process that is time-consuming and dependent on expert pathologists. Macroscopic cross-section images, captured at the time of surgery, offer a faster and more accessible diagnostic window that has not been fully exploited.
This study aimed to develop and validate deep learning-based classification models that could automatically recognize renal tumor pathology directly from macroscopic gross images, offering a potential tool to streamline intraoperative and postoperative decision-making.
A retrospective cohort of 467 pathologically confirmed patients who underwent radical or partial nephrectomy was assembled. All patients had macroscopic cross-section images captured from their surgical specimens, providing a diverse dataset of both malignant and benign tumors.
The dataset included 9 malignant subtypes, dominated by clear cell RCC (ccRCC) with 288 cases, alongside papillary RCC, chromophobe RCC, and other rarer subtypes. Benign tumors comprised 10 subtypes including oncocytoma and angiomyolipoma.
The study employed a two-tier classification approach: first, binary malignant vs. benign classification, and second, multi-subtype classification within each category. Five-fold cross-validation was used to ensure robust performance estimates across all model architectures.
Three convolutional neural network (CNN) architectures were evaluated as classification backbones: EfficientNet-B4, ResNet-18, and VGG-16. All models were pretrained on ImageNet and fine-tuned on the renal tumor image dataset through transfer learning, enabling the networks to leverage learned visual features even with limited medical data.
Transfer learning was critical in this context because medical imaging datasets are typically smaller than those used in general computer vision tasks. By initializing with ImageNet weights and fine-tuning on tumor images, the models could learn pathology-relevant features without overfitting.
Performance was evaluated using area under the receiver operating characteristic curve (AUC), sensitivity, specificity, and accuracy. Both macro-averaged and micro-averaged metrics were reported to account for class imbalance across tumor subtypes.
For binary malignant vs. benign classification, all three CNN models achieved satisfactory performance. ResNet-18 delivered the highest AUC of 0.9226, demonstrating that macroscopic image features carry sufficient discriminative information to differentiate tumor types reliably.
Statistical comparison revealed no significant difference between EfficientNet-B4 and ResNet-18, but both architectures significantly outperformed VGG-16. This suggests that more modern, efficient architectures provide meaningful improvements over older designs for this specific imaging task.
The strong binary classification results validated the fundamental premise of the study: that gross pathological appearance captured in macroscopic images encodes enough morphological information for automated malignancy detection, even without microscopic histological features.
For malignant subtype classification, VGG-16 achieved a micro-averaged AUC of 0.9398, macro-averaged sensitivity of 0.5774, macro-averaged specificity of 0.8660, and micro-averaged accuracy of 0.7917. Interestingly, EfficientNet-B4 did not surpass VGG-16 on the micro-averaged AUC for malignant subtypes, though it outperformed on other metrics.
For benign subtype classification, EfficientNet-B4 ranked best overall but showed no statistically significant advantage over the other two architectures in terms of micro-averaged AUC. These mixed results suggest that different architectures may have complementary strengths depending on the classification task.
Multi-subtype classification is inherently more challenging than binary classification due to the greater number of categories, class imbalance, and subtle visual differences between subtypes. Despite these challenges, all models showed clinically meaningful discriminative ability.
This study is among the first to systematically apply deep learning to macroscopic cross-section images for renal tumor classification. Unlike CT or MRI-based AI studies, the macroscopic approach uses images taken directly from resected specimens, making it highly relevant to the intraoperative and post-resection pathology workflow.
A key limitation is the retrospective single-center design, which may introduce selection bias and limit generalizability. The dataset size, while reasonable for a deep learning study in this niche, remains relatively small compared to large-scale computer vision benchmarks, which may constrain model generalization.
Despite these limitations, the study provides proof-of-concept evidence that macroscopic image analysis using CNNs can achieve clinically useful accuracy. Future work should address prospective multicenter validation and explore ensemble approaches combining multiple architectures.
Automated classification of renal tumors from macroscopic images could substantially accelerate the pathological reporting process, reducing turnaround time between surgery and diagnosis. This is particularly valuable in settings with limited access to experienced pathologists.
Intraoperative use of such models could help surgeons make real-time decisions about surgical margins and extent of resection, potentially improving oncological outcomes while preserving renal function whenever possible.
The model's ability to distinguish malignant from benign tumors and identify specific subtypes could also guide decisions about adjuvant therapy, follow-up intensity, and genetic counseling, making it a valuable component of a precision oncology workflow for kidney cancer.