ResLink: CNN for CT Brain Tumor Classification

• ResLink is a CNN architecture that combines area attention with residual connections for robust binary brain tumor classification from CT images.
• It employs a multi-stage convolutional pipeline including a convolutional stem, intermediate attention blocks, and an attention-based classification head for localized feature extraction.
• The model achieves a high test accuracy of 95% on a balanced Kaggle dataset, demonstrating rapid convergence and efficient computation.

ResLink is a custom convolutional neural network architecture designed specifically for binary brain tumor classification from CT scan images (“Healthy” vs “Tumor”). It combines Area Attention with residual connections in a multi-stage convolutional pipeline intended to improve feature learning and spatial understanding while avoiding heavy pre-trained backbones and keeping computational complexity modest. In the reported study, CT images are resized to $224 \times 224$, training is performed on a balanced Kaggle brain tumor CT dataset, and the model reaches an overall test accuracy of 95% (Arya et al., 24 Aug 2025).

1. Problem formulation and imaging setting

ResLink addresses a binary image-classification problem in which the input is a 2D brain CT scan and the output is a probability $\hat{y} \in [0,1]$ indicating tumor presence. The label space is binary, with “Healthy” encoded as $0$ and “Tumor” encoded as $1$. The architecture is therefore framed as a binary classifier rather than a tumor-subtype recognizer or a volumetric detection system (Arya et al., 24 Aug 2025).

The imaging context is central to the model design. CT scans are described as lower-contrast than MRI for soft tissue, but widely available and fast. Because tumors typically occupy localized regions, the task is presented as one of extracting robust spatial features while emphasizing suspicious areas against complex anatomical background. This motivates the use of Area Attention as a locality-preserving attention mechanism over spatial regions rather than a global transformer-style attention layer.

A plausible implication is that ResLink is intended for settings in which CT accessibility and inference efficiency matter as much as raw classification performance. The model is explicitly positioned as an alternative to heavier transfer-learning pipelines, with the design emphasis placed on localized feature refinement and stable deep optimization.

2. Network organization and computational pipeline

At the architectural level, ResLink is organized into three functional parts: a stem, a sequence of intermediate attention and residual stages, and a final attention-based classification head (Arya et al., 24 Aug 2025).

The stem applies an initial $7 \times 7$ convolution, batch normalization, ReLU, and max pooling. If the input image is denoted

$X \in \mathbb{R}^{H \times W \times C},$

the stem computes

$$F_0 = \text{MaxPool}\big(\text{ReLU}(\text{BN}(W_s * X))\big).$$

The convolution uses stride $=2$, and the stage performs both low-level feature extraction and early downsampling.

The intermediate pipeline contains AreaAttentionLayer modules, residual CNN blocks, and downsampling layers. Residual blocks use two $3 \times 3$ convolutions with batch normalization and skip connections, while deeper-stage downsampling is performed by strided convolution rather than pooling. The downsampling operation is written as

$$F_{\text{down}} = \text{ReLU}\big(\text{BN}(W_d * H)\big),$$

with a $\hat{y} \in [0,1]$0 convolution of stride $\hat{y} \in [0,1]$1.

The final stage applies another Area Attention layer to the last convolutional feature map, followed by Global Average Pooling, dropout, and a dense layer with sigmoid activation. For channel $\hat{y} \in [0,1]$2, the pooled representation is

$\hat{y} \in [0,1]$3

After dropout,

$\hat{y} \in [0,1]$4

is used for binary prediction.

The exact number of residual stages and filters per stage is not enumerated. What is specified is the functional composition and kernel sizes. This suggests that ResLink is defined more by its architectural grammar—stem, local attention, residual refinement, strided downsampling, and attention-based head—than by a single rigid depth configuration.

3. Area Attention and residual connectivity

The core novelty of ResLink is the Area Attention Layer, which operates over non-overlapping spatial areas of an intermediate feature map rather than over the entire map at once (Arya et al., 24 Aug 2025).

Let

$\hat{y} \in [0,1]$5

be a feature map. ResLink partitions $\hat{y} \in [0,1]$6 into areas

$\hat{y} \in [0,1]$7

where $\hat{y} \in [0,1]$8 is the area size. For each area, the attention map is computed as

$\hat{y} \in [0,1]$9

where $0$0 is a $0$1 convolution for channel mixing and $0$2 is a $0$3 convolution for local spatial refinement. The area-level attention maps are reassembled into a global attention tensor $0$4, and the attended feature map is obtained by elementwise multiplication: $0$5

This mechanism is explicitly distinguished from standard query–key–value attention. It is a local convolutional attention method in which each area learns a multiplicative importance mask. The locality of the computation makes it computationally lighter than global self-attention while remaining aligned with the observation that tumors are spatially localized.

Residual processing then refines the attended representation. For a feature map $0$6,

$0$7

$0$8

The shortcut path is

$0$9

The residual output is

$1$0

When attention is enabled at a block, $1$1 is replaced by $1$2 before convolution. The stated rationale is that attention modulates the content of suspicious regions, while residual connections preserve spatial integrity and gradient stability. This suggests a division of labor in which Area Attention performs localized emphasis and skip connections constrain that emphasis so that important structure is not catastrophically suppressed.

4. Training protocol, preprocessing, and regularization

The dataset source is the Kaggle brain tumor CT dataset, used as a benchmark for binary “Healthy” versus “Tumor” classification. The preprocessing pipeline consists of label encoding, class balancing with RandomOverSampler, stratified train–validation–test splitting at 80% / 10% / 10%, image resizing to $1$3, pixel normalization to $1$4, and data loading through ImageDataGenerator (Arya et al., 24 Aug 2025).

The learning objective is binary cross-entropy: $1$5 and for a batch of $1$6 examples,

$1$7

Optimization uses Adam, and the reported training schedule consists of 5 epochs. The text emphasizes rapid convergence. Batch size, learning rate, dropout rate, and weight decay are not specified as fixed experimental values. Dropout is applied after Global Average Pooling and before the final dense layer; batch normalization appears in the stem, inside Area Attention, within residual blocks, and in downsampling layers.

Regularization is distributed across several mechanisms. Dropout reduces feature co-adaptation at the classification head. Batch normalization improves training stability and functions as a regularizer. RandomOverSampler addresses class imbalance by duplicating minority-class examples in the training set. Heavy augmentation is not described; only resizing and rescaling are explicitly confirmed, while richer augmentation is reserved for future work. This combination indicates a training strategy that prioritizes architectural inductive bias and balanced supervision over extensive augmentation or long optimization schedules.

5. Reported performance and empirical interpretation

The reported training dynamics show rapid improvement during early epochs. At Epoch 1, training accuracy is 77.68%, training loss is 0.7382, and validation accuracy is 92.67%. At Epoch 2, training accuracy reaches 95.25% and validation accuracy reaches 94.83%. The final reported test-set overall accuracy is 95% (Arya et al., 24 Aug 2025).

The paper also reports class-sensitive metrics. Precision for class 1 (Healthy) is 0.97, recall for class 1 (Healthy) is 0.92, and the macro-average F1-score is 0.95. The standard definitions of accuracy, precision, recall, and macro-average F1 are given, but specificity, AUC, and sensitivity for the tumor class are not reported.

The empirical interpretation is partly quantitative and partly qualitative. The model is presented as achieving high accuracy on a balanced CT dataset without using pre-trained models. Formal ablation studies are not included, and no direct numerical comparison against alternative CT-based systems is reported. The qualitative explanation for the reported performance attributes improved generalization and F1-score to two factors: Area Attention focuses on informative regions, and residual connections preserve spatial integrity and gradient stability.

Confusion-matrix analysis is described as showing strong performance on both classes with limited misclassifications. At the same time, the absence of cross-dataset validation, k-fold cross-validation, tumor-class sensitivity, and formal ablations limits the strength of causal claims about which architectural component contributes most strongly.

6. Generalizability, limitations, and prospective extensions

ResLink is described as demonstrating strong generalizability, and the use of stratified splitting together with RandomOverSampler is presented as supporting generalization across data splits. However, no external validation dataset is used, and no k-fold cross-validation metrics are provided (Arya et al., 24 Aug 2025).

Several limitations are explicit. The evaluation is CT-only, so performance may differ on MRI, which has different intensity distributions and richer contrast. The task is binary classification only, with no tumor subcategories such as glioma, meningioma, or pituitary tumor. Training lasts only 5 epochs, which supports quick convergence but may constrain maximal performance. The dataset size is not quantified, and this raises questions about robustness across scanners, institutions, and patient populations.

Future work is framed along several axes. One direction is dataset augmentation and diversity, including rotations, flips, intensity shifts, and additional CT datasets from different institutions. Another is hyperparameter tuning, including learning-rate scheduling and adjustment of dropout rate and area size. A further extension is adaptation to MRI and multi-modal data, with the possibility of integrating 3D convolutions for volumetric inputs. The classifier head can also be generalized from sigmoid to softmax for multi-class tumor classification.

The broader significance of ResLink lies in its role as an architectural template for localized, convolutional attention integrated with residual connections in medical image classification. This suggests relevance beyond the reported benchmark: the design can be implemented, adapted, and scaled in settings where pathology is spatially localized and where lightweight, task-specific CNNs are preferred over heavier pre-trained backbones.