Renal cell carcinoma (RCC) encompasses three major histologic subtypes: clear cell RCC (KIRC), papillary cell RCC (KIRP), and chromophobe RCC (KICH), which account for 70 to 80 percent, 14 to 17 percent, and 4 to 8 percent of cases respectively. Each subtype has a distinct prognosis, molecular profile, and response to targeted therapies, making accurate classification essential for treatment planning.
Manual examination of hematoxylin and eosin (H&E) stained tissue slides by pathologists is the standard method for RCC subtype classification, but it is time-consuming, prone to interobserver variability, and critically dependent on the expertise of the individual pathologist. In high-volume centers or resource-limited settings, this creates diagnostic bottlenecks.
Each RCC subtype has distinctive histological features visible at the microscopic level. Clear cell RCC is characterized by abundant clear cytoplasm, staghorn-shaped vasculature, and cells organized in acinar or alveolar patterns. Papillary RCC shows finger-like projections with fibrovascular cores and bilateral or multifocal origin. Chromophobe RCC features pale eosinophilic cells with prominent cell membranes, irregular nuclei, and perinuclear halos.
Deep learning has demonstrated strong performance across many histopathology classification tasks, but prior approaches for RCC subtyping often used transfer learning from non-medical image datasets, which may not optimally capture the unique nuclear and textural features of renal tumor cells. A purpose-built architecture trained from scratch could better exploit domain-specific morphological patterns.
RenalNet is a novel convolutional neural network (CNN) developed specifically for classifying RCC subtypes from H&E stained histopathology patches. Unlike prior approaches that fine-tune ImageNet-pretrained models, RenalNet is trained from scratch to capture the morphological and clinical features unique to renal cell carcinoma without relying on natural image priors.
The architecture is organized around two core modules: the Multiple Channel Residual Transformation (MCRT) module for cross-channel feature extraction and the Group Convolution Deep Localization (GCDL) module for integrating multi-scale spatial features. Together, these modules enable the network to extract features that are simultaneously sensitive to nuclear texture, cellular organization, and tissue architecture.
A key design principle of RenalNet is computational efficiency. The model uses only 0.2131 million trainable parameters and 2.71 billion floating-point operations, which is dramatically fewer than comparable models such as the Vision Transformer at 33.5 million parameters and Hybrid Deep Feature Fusion at 34.7 million parameters. This compactness reduces training time and hardware requirements.
RenalNet was validated on three different histopathology datasets: a newly introduced TCGA kidney dataset created specifically for this study, and two liver histopathology datasets from Kasturba Medical College. Testing on liver data assessed the model's ability to generalize to different tissue types, verifying that its design is not overfitted to one organ or staining pattern.
The Multiple Channel Residual Transformation (MCRT) module consists of three parallel convolutional branches, each followed by batch normalization, ReLU activation, and a Squeeze-and-Excitation (SE) channel attention block. Each branch uses a different squeeze ratio (r = 2, 4, and 8), producing channel descriptors at three different levels of compression.
The SE mechanism first squeezes spatial information into a channel-wise descriptor through global average pooling, then excites the channels using a sigmoid-gated fully connected layer that weights each channel by its relevance to the current input. This allows the network to focus on the most informative feature channels for each specific region of the histopathology image.
By using three parallel SE blocks with different squeeze ratios, MCRT captures channel-wise dependencies at multiple compression levels and fuses them to strengthen local semantic features. The enriched multi-ratio channel representation is combined across parallel paths and added to the input through an identity connection, forming a residual module that preserves gradient flow during backpropagation.
The MCRT module is placed at three successive stages within the encoder, progressively enriching features from coarse low-level texture patterns to fine nuclear-level morphological detail. This staged application allows the network to build a hierarchical representation of renal tissue characteristics at increasing levels of abstraction.
The Group Convolution Deep Localization (GCDL) module integrates features from three different stages of the MCRT encoder using three parallel strided convolution branches with strides of 4, 2, and 1 respectively. This multi-stride design aligns features from different spatial resolutions into a common feature space through concatenation.
After concatenation, the fused multi-scale feature undergoes a further separable convolution to combine the cross-stage information, followed by global average pooling that generates a compact representation encoding the overall effect of all previous network operations. This global pooling step allows the model to integrate information from across the entire input patch.
An attention mechanism within GCDL magnifies the output of the deepest MCRT stage (which contains the richest nuclear and contextual features) by the global pooled representation, producing an attention-modulated feature map that emphasizes granular tumor regions relative to background tissue. This step is specifically designed to improve localization of diagnostic histological structures.
An aggregated depthwise separable convolution block within GCDL reduces computational complexity by factorizing standard convolutions into depthwise and pointwise components. This efficiency is key to keeping the total parameter count low without sacrificing the expressiveness needed for accurate subtype differentiation.
On the newly introduced TCGA kidney dataset with four classes (Normal, KIRC, KIRP, KICH), RenalNet achieved an overall accuracy and F1-score of 0.9167, outperforming all eight comparison models including ShuffleNet, ESPNetV2, BHCNet, Pan-RCC, BreastNet, LiverNet, Vision Transformer, and Hybrid Deep Feature Fusion. The closest competitor, BHCNet, achieved 0.8755.
On the KMC Liver dataset with four HCC grade classes, RenalNet achieved an accuracy of 0.9714 and F1-score of 0.9726, demonstrating strong generalization to a completely different organ and tissue type despite being tested without retraining on liver-specific data. This confirms that the architecture captures generalizable morphological features.
Five-fold cross-validation results on the kidney dataset were consistent with the held-out test set performance, with the proposed model achieving accuracy of 0.9011 and showing similar or better results compared to all reference models in all five folds. The stability across folds indicates the model is not sensitive to the specific training-test partition.
Computational complexity analysis showed that RenalNet required only 1.227 hours for training and 12.73 seconds for inference on the full test set, compared to 3.2 hours and 27.5 seconds for the Vision Transformer. The combination of highest accuracy with the lowest training and inference time across all tested models makes RenalNet particularly attractive for clinical deployment.
Ablation experiments removing each major component confirmed that both the MCRT and GCDL modules contribute independently to classification performance. Removing GCDL reduced accuracy by 2.43 percent, while removing MCRT reduced accuracy by 3.54 percent, indicating that MCRT is the more critical of the two modules but both are necessary for optimal performance.
Activation map comparisons across eight network variants showed that the full RenalNet assigns high probability scores specifically to nuclear regions while correctly suppressing background tissue, while ablated variants showed false-positive activation in non-diagnostic areas. This pattern confirms that MCRT and GCDL together improve the network's spatial specificity for tumor-relevant features.
The decision to train from scratch rather than use ImageNet pretraining was validated by direct comparison against eight standard transfer learning models including VGG-16, ResNet-50, DenseNet, and NasNet. RenalNet trained from scratch achieved the highest accuracy among all tested models including those using ImageNet weights, with an 11.93 percent margin over the worst transfer learning model.
The class-balanced dataset construction, with an equal number of patches from each class extracted under pathologist supervision, is identified as a critical factor in model performance. High class imbalance in prior publicly available datasets led to biased models that underperform on rare subtypes like KICH, a limitation this study explicitly addresses.
RenalNet demonstrates that a lightweight CNN designed specifically for renal histopathology morphology can outperform much larger and more computationally expensive architectures on the task of RCC subtype classification, validating the principle that domain-specific design outperforms generic model scale.
The introduction of a class-balanced, pathologist-curated TCGA kidney dataset addresses a critical gap in the field and provides a reproducible benchmark for future model comparisons. The dataset and code are publicly available, supporting the replication and extension of these results by other research groups.
The validation on liver histopathology data demonstrates that RenalNet's architectural principles generalize across tissue types, suggesting that the MCRT and GCDL module design could be adapted for histopathology classification problems beyond renal cancer.
Future work will explore hardware implementation of RenalNet for real-time pathology support systems, integration with clinical laboratory information systems, and extension to grading tasks within RCC subtypes, which remain important determinants of prognosis and therapy selection beyond subtype classification alone.