R-Net: Lightweight CNN for CRC Detection

• R-Net is a lightweight and explainable CNN designed for multiclass colorectal cancer detection and classification from histopathological images.
• It employs a staged convolutional design with alternating Conv2D and max pooling layers, achieving over 99% accuracy on the EBHI dataset.
• The model integrates XAI techniques like SHAP, LIME, and Grad-CAM to provide visual explanations for its predictions while maintaining resource efficiency.

Searching arXiv for the specified R-Net paper and closely related similarly named models for accurate disambiguation. Reasonable Network (R-Net) is a lightweight convolutional neural network proposed for colorectal cancer (CRC) detection and classification from histopathological images, specifically on the Enteroscope Biopsy Histopathological Hematoxylin and Eosin Image Dataset (EBHI). It was introduced as “a reliable and resource-efficient CNN” intended to reduce computational power, training time, and dataset demands relative to several state-of-the-art convolutional neural networks while incorporating explainable AI (XAI) methods, including SHAP, LIME, and Grad-CAM, to visualize the image regions contributing to its predictions (Ayon et al., 17 Sep 2025). In the source formulation, the term “reasonable network” denotes a design that emphasizes a lightweight architecture, strong prediction results, and integrated interpretability rather than architectural depth or ensemble complexity alone (Ayon et al., 17 Sep 2025).

1. Definition and scope

R-Net is a custom-designed CNN for multiclass CRC histopathology classification. The model is presented as addressing a setting in which state-of-the-art CNNs such as ResNet, DenseNet, and MobileNet often require substantial computational resources, large datasets, and longer training times, which can hinder deployment in resource-constrained clinical environments (Ayon et al., 17 Sep 2025).

The network is explicitly framed around three design principles. It is described as lightweight, with few layers and lower parameter count compared to state-of-the-art CNNs; reliable, with high generalization and minimal misclassifications across multiclass CRC histopathology images; and explainable, through native integration of XAI techniques for transparent decision-making (Ayon et al., 17 Sep 2025). This combination places R-Net within a line of medical-imaging architectures that seek a compromise between predictive performance, resource efficiency, and auditability.

Within the EBHI task formulation, R-Net performs six-class classification. The output categories are Adenocarcinoma, High-Grade IN, Low-Grade IN, Normal, Polyp, and Serrated Adenoma (Ayon et al., 17 Sep 2025). The model operates on RGB inputs of size $64 \times 64 \times 3$, and the training pipeline uses 5-fold Cross Validation for robust evaluation (Ayon et al., 17 Sep 2025).

2. Architectural specification

R-Net follows a staged convolutional architecture with alternating convolutional blocks and max pooling, followed by dense classification layers. The architectural blueprint reported for the model comprises two consecutive $\text{Conv2D}$ layers with 64 filters each, followed by $\text{MaxPooling2D}$; two $\text{Conv2D}$ layers with 128 filters each, followed by $\text{MaxPooling2D}$; three $\text{Conv2D}$ layers with 256 filters each, followed by $\text{MaxPooling2D}$; and three $\text{Conv2D}$ layers with 512 filters each, followed by $\text{MaxPooling2D}$ (Ayon et al., 17 Sep 2025).

After the convolutional and pooling stages, the representation is flattened and passed through a dense layer with 256 neurons, then a dense layer with 64 neurons, and finally an output dense layer with 6 neurons corresponding to the six diagnostic classes (Ayon et al., 17 Sep 2025). The reported trainable parameter count is approximately 15.9 million, given explicitly as $15{,}911{,}430$ (Ayon et al., 17 Sep 2025).

Layer type	Number of filters/neurons
Conv2D (Block 1)	64, 64
Conv2D (Block 2)	128, 128
Conv2D (Block 3)	256, 256, 256
Conv2D (Block 4)	512, 512, 512
Dense	256
Dense	64
Dense (Output)	6

The paper characterizes this design as avoiding “complex multi-path/width/depth expansions” of architectures such as DenseNet, Xception, or Inception, and instead focusing on essential feature extraction (Ayon et al., 17 Sep 2025). This suggests that the “reasonable” designation refers not to a formally defined optimization criterion, but to an engineering stance: a CNN that is shallow relative to heavier competitors yet wide enough to capture discriminative histomorphological structure.

3. Training objective and optimization protocol

The loss function used for R-Net is Sparse Categorical Cross-Entropy, chosen as suitable for integer class labels. The reported form is

$\text{Conv2D}$0

where $\text{Conv2D}$1 denotes the predicted probability of the true class $\text{Conv2D}$2 for sample $\text{Conv2D}$3 (Ayon et al., 17 Sep 2025). The optimization algorithm is Adam with learning rate $\text{Conv2D}$4 (Ayon et al., 17 Sep 2025).

A learning-rate schedule is also specified:

$\text{Conv2D}$5

with early stopping at patience $\text{Conv2D}$6 based on validation accuracy (Ayon et al., 17 Sep 2025). The batch size is 32, and the input image size is $\text{Conv2D}$7 (Ayon et al., 17 Sep 2025).

Evaluation uses standard multiclass classification metrics: accuracy, precision, recall, F1-score, macro/micro/weighted averages, confusion matrix, and ROC/AUC per class (Ayon et al., 17 Sep 2025). The reported description further states that all classes achieved $\text{Conv2D}$8 in cross-validation (Ayon et al., 17 Sep 2025). In context, this evaluation protocol emphasizes both aggregate discrimination and class-wise behavior, which is particularly relevant in multiclass diagnostic settings where different lesion types can have different visual confounders.

4. Dataset, task formulation, and comparative setting

The benchmark dataset is EBHI, identified as the Enteroscope Biopsy Histopathological H&E Image Dataset (Ayon et al., 17 Sep 2025). The task is six-class image classification, and the data were “manually augmented with GANs to balance class distributions” (Ayon et al., 17 Sep 2025). After augmentation, the total number of images is 4800, with a class split of 70% train, 10% validation, and 20% test (Ayon et al., 17 Sep 2025).

The comparative study includes several state-of-the-art CNNs tested on the same dataset: DenseNet121, ResNet50, InceptionV3, Xception, MobileNetV2, and VGG16, along with transfer learning variants and two ensemble configurations (Ayon et al., 17 Sep 2025). The ensemble models are described as a multipath-depth-width combination, DenseNet121-InceptionV3-Xception, and a multipath-depth-spatial combination, ResNet50-InceptionV3-VGG16, with Soft, Hard, and Rank-based voting (Ayon et al., 17 Sep 2025).

The paper’s positioning of R-Net is therefore not against a single baseline, but against three families of alternatives: standalone CNNs, transfer learning pipelines, and ensemble models. This is important for interpretation. The claim is not merely that a lightweight custom CNN can beat one comparator; rather, the study situates R-Net inside a relatively broad comparison set on one dataset and one preprocessing regime.

5. Performance, efficiency, and interpretability

On the EBHI test set, R-Net is reported to achieve 99.37% accuracy (Ayon et al., 17 Sep 2025). The same source reports precision, recall, and F1-score all approximately 0.99, and “only 3/4800 misclassifications” (Ayon et al., 17 Sep 2025). Comparative test accuracies listed in the same study include DenseNet121 at 99.93, ResNet50 at 96.94, InceptionV3 at 99.44, Xception at 100.0, MobileNetV2 at 95.83, VGG16 at 96.74 (Adamax), and the top ensemble soft-voting model at 98.23 (Ayon et al., 17 Sep 2025).

These results support two distinct but easily conflated claims. First, R-Net outperforms MobileNetV2 and ResNet50 on the reported test set (Ayon et al., 17 Sep 2025). Second, R-Net is not the single highest-accuracy model in the comparison table, because DenseNet121 and Xception are reported at 99.93 and 100.0, respectively (Ayon et al., 17 Sep 2025). The paper’s stronger claim is therefore about the balance between accuracy and resource demands, not about absolute top-line accuracy alone.

The interpretability component is implemented through LIME, SHAP, and Grad-CAM. LIME provides instance-level explanations and highlights positive and negative image regions; SHAP quantifies feature importance per prediction, with red and blue regions indicating positive and negative influence; Grad-CAM produces heatmaps showing which spatial regions contributed most to the final decision (Ayon et al., 17 Sep 2025). The paper attributes three functions to these methods: transparency, by helping verify that the model attends to pathologically relevant tissue structures; error analysis, by enabling inspection of misclassified cases; and trust and adoption, by making model decisions visualizable to clinicians (Ayon et al., 17 Sep 2025).

Model	Test Accuracy (%)
R-Net	99.37
DenseNet121	99.93
ResNet50	96.94
InceptionV3	99.44
Xception	100.0
MobileNetV2	95.83
VGG16	96.74
Soft Voting (Top Ensemble)	98.23

From an encyclopedic standpoint, the central significance of these results lies in the combination of near-ceiling classification performance on EBHI with an explicit claim of lower computational and memory overhead than heavier models such as ResNet50 and DenseNet121 (Ayon et al., 17 Sep 2025). A plausible implication is that the authors view R-Net less as a universal replacement for larger CNNs than as a deployment-oriented compromise for settings where compute, memory, and interpretability constraints are salient.

6. Interpretation of the “reasonable” design philosophy

The paper explicitly describes the main novelty as “building a reliable, lightweight CNN R-Net that requires fewer computing resources yet maintains strong prediction results” (Ayon et al., 17 Sep 2025). It also states that the architecture demonstrates “simplicity with SOTA performance,” and that a “carefully optimized, shallow-yet-wide architecture” can match or exceed deeper, more complex networks (Ayon et al., 17 Sep 2025).

Several concrete elements of this design philosophy are stated directly. The network has a smaller parameter count than heavier state-of-the-art models, avoids multi-path and depth-expansion mechanisms, integrates explainability from the outset, and relies on balanced synthetic augmentation and cross-validation to remain robust with small datasets (Ayon et al., 17 Sep 2025). The source also characterizes R-Net as showing consistently high precision, recall, and F1 across all folds and classes, which it interprets as suggesting low risk of overfitting and low variance (Ayon et al., 17 Sep 2025).

At the same time, the comparison table complicates any simplistic reading of “reasonable” as strictly “smallest” or “most accurate.” MobileNetV2 is reported at roughly 3.5M parameters, whereas R-Net is reported at approximately 15.9M (Ayon et al., 17 Sep 2025). Thus, “reasonable” is not equivalent to minimum parameter count. Instead, it appears to denote a balance among architecture size, predictive quality, and interpretability. This suggests a more specific editorial shorthand: R-Net is a performance-efficiency-interpretability compromise model (Editor’s term), not merely a compact model.

A common misconception would be to equate R-Net’s reported performance with proof of broad clinical generalization. The source material supports strong results on EBHI with GAN-based balancing, 5-fold Cross Validation, and the stated train/validation/test split, but does not provide a broader multicenter or external-validation claim in the supplied data (Ayon et al., 17 Sep 2025). Another misconception would be to treat XAI overlays as a direct guarantee of model correctness; the paper presents them as visualization tools for understanding and validating decision-making, not as a formal correctness certificate (Ayon et al., 17 Sep 2025).

7. Nomenclature and relation to similarly named models

The name “R-Net” is potentially ambiguous in the arXiv literature. In the present context, it refers specifically to the “reasonable network” for CRC detection and classification (Ayon et al., 17 Sep 2025). This model is distinct from several similarly named systems in other fields.

A different R-NET was used for behavioral robot navigation from natural language directions, where a modified R-NET with gated attention, self-matching attention, and a pointer network translated instructions into a high-level plan over a behavioral navigational graph (Shrestha et al., 2020). That architecture was adapted from reading comprehension and targeted combinatorial path planning rather than medical image classification (Shrestha et al., 2020).

ReNet in datacenter networking denotes a self-adjusting, demand-aware network architecture based on ego-trees, helper nodes, and a coordinator, with static optimality guarantees for sparse communication demands (Avin et al., 2019). This is a network-topology design problem rather than a CNN for image diagnosis (Avin et al., 2019).

The R-mAtrIx Net concerns neural-network construction of $\text{Conv2D}$9-matrices and integrable Hamiltonians under Yang–Baxter, regularity, and hermiticity constraints in quantum spin-chain theory (Lal et al., 2023). Its use of “R-matrix” has no connection to the “reasonable network” terminology of CRC histopathology (Lal et al., 2023).

Finally, $\text{MaxPooling2D}$0Net addresses 3D radio map estimation by embedding height information into 2D images and using residual U-Net-like architectures tailored separately to indoor and outdoor wireless propagation (Rao et al., 18 May 2026). Despite the superficial naming overlap, it belongs to wireless channel estimation rather than computational pathology (Rao et al., 18 May 2026).

This terminological disambiguation matters because “R-Net” is not a standardized architecture class across domains. In arXiv usage, the label has been applied to unrelated systems in robot navigation, self-adjusting networks, quantum integrability, radio-map estimation, and CRC histopathology. In the CRC literature considered here, however, Reasonable Network (R-Net) specifically denotes the lightweight, explainable CNN introduced for EBHI-based colorectal cancer detection and classification (Ayon et al., 17 Sep 2025).