A brain-inspired computational framework for image-based risk assessment

Abstract

Skin cancer risk prediction plays a critical role in early diagnosis and personalized healthcare. However, existing approaches often struggle to achieve both rich feature representation and computational efficiency. To address these issues, this study proposes Bicom, a brain-inspired framework for skin cancer risk prediction that integrates efficient attention mechanisms, multi-scale feature fusion, and confidence-aware refinement. Specifically, we develop F-ResNeSt, a multi-scale feature-extraction architecture that enhances the ResNeSt backbone by incorporating a Feature Pyramid Network (FPN) and Linformer-based linear-complexity attention. In addition, we propose L-CoAtNet, an optimized classification network that replaces conventional relative attention with Linformer-based attention to achieve scalable global contextual modeling. Furthermore, a brain-inspired Spiking Neural Network (SNN) module is introduced as a confidence-aware refinement mechanism to enhance prediction reliability for ambiguous samples. Comprehensive experiments on public and subject datasets demonstrate that the proposed framework achieves consistently competitive performance across multiple evaluation metrics, indicating its effectiveness, robustness, and scalability for assisted skin cancer risk prediction.

Introduction

Skin cancer is one of the most common malignancies worldwide, and its incidence has continued to increase in recent decades. Early detection and timely intervention are critical for improving patient survival rates and reducing treatment costs. However, skin cancer often exhibits subtle or atypical visual characteristics in its early stages, making accurate diagnosis highly dependent on clinical experience. As a result, intelligent risk prediction models have become an important auxiliary tool for supporting dermatologists in early screening and diagnosis. In recent years, deep learning–based methods have demonstrated promising performance in skin lesion analysis and cancer risk prediction. Most existing approaches rely on convolutional neural networks or attention-based architectures to extract discriminative features from dermoscopic images.

Despite their success, several challenges remain. First, many models suffer from high computational complexity when processing high-resolution medical images, limiting their practical deployment. Second, existing methods often lack sufficient multiscale representation, resulting in suboptimal recognition of lesions with diverse sizes and morphologies. Third, most current frameworks treat classification as a deterministic single-stage decision process and lack effective mechanisms for handling prediction uncertainty in ambiguous cases.

To address these limitations, this paper proposes a novel skin cancer risk prediction framework, termed Bicom (A Skin Cancer Risk Prediction Model Based on Brain-Inspired Computing Technology), designed for assisted diagnostic applications. As illustrated in Fig. 1, the proposed model can be deployed in both cloud-based and local clinical environments, providing real-time decision support for physicians while ensuring data security and privacy.

Fig. 1.

Bicom model application scenarios.

From a methodological perspective, the proposed framework integrates multiple complementary components to achieve efficient and reliable diagnosis. Specifically, a Linformer-enhanced multi-scale feature extraction architecture, F-ResNeSt, is developed to jointly improve global dependency modeling and scale adaptability. An improved classification network, L-CoAtNet, is further introduced to reduce memory consumption and computational complexity while preserving contextual modeling capability. More importantly, a brain-inspired spiking neural network (SNN) module is incorporated as a confidence-aware refinement mechanism, which is selectively activated for low-confidence predictions to re-evaluate ambiguous samples and enhance robustness.

Through this hierarchical, uncertainty-aware design, the proposed Bicom framework explicitly models prediction reliability and introduces a dynamic refinement process, thereby distinguishing it from conventional single-stage classification systems. Extensive experiments on public and subject datasets demonstrate that Bicom achieves superior accuracy, robustness, and generalization performance in skin cancer risk prediction tasks.

The main contributions of this work are summarized as follows:

• We propose a novel feature extraction architecture, termed F-ResNeSt, which integrates Linformer-based linear attention and Feature Pyramid Network into the ResNeSt backbone, enabling efficient multi-scale and global-aware representation learning for high-resolution skin lesion images.

• We develop an improved image classification framework, L-CoAtNet, by replacing the relative attention mechanism in CoAtNet with Linformer-based attention, significantly reducing computational and memory complexity while preserving effective global contextual modeling.

• We introduce an end-to-end skin cancer risk prediction framework, Bicom, which uniquely combines F-ResNeSt, L-CoAtNet, and a brain-inspired spiking neural network module to perform confidence-aware prediction refinement, resulting in enhanced robustness and predictive accuracy in assisted diagnostic scenarios.

The remainder of this paper is organized as follows. Section 2 reviews related work on skin cancer diagnosis and risk prediction. Section 3 introduces the preliminary techniques employed in this study. Section 4 presents the proposed methodology in detail. Section 5 reports experimental results and analysis. Section 6 discusses the advantages, limitations, and future directions. Finally, Sect. 7 concludes the paper.

Related work

With the rapid advancement of artificial intelligence and medical imaging technologies, a large body of research has focused on automated skin cancer diagnosis and risk prediction using machine learning and deep learning techniques. These studies aim to improve diagnostic accuracy, robustness, and clinical efficiency through advanced feature representation and decision modeling. Existing research in this field can be systematically categorized into three main directions.

Integrated deep learning frameworks for segmentation and classification of skin cancer

Integrated deep learning frameworks that jointly perform lesion segmentation and classification have emerged as a prominent research direction in automated cancer diagnosis. Deep learning–based frameworks have been widely investigated for skin cancer diagnosis, particularly through integrating lesion segmentation and classification to improve diagnostic robustness and accuracy. In this research direction, hybrid architectures combining deep neural networks with traditional classifiers have been commonly adopted. Representative studies include the work of Bassel et al.¹, who integrated deep feature extraction using ResNet50, Xception, and VGG16 with support vector machines and random forests, demonstrating that Xception-based features achieved the highest classification accuracy. Building upon this paradigm, Attique et al.² introduced a two-stream deep neural network evaluated on HAM10000, ISBI2018, and ISIC2019 datasets, incorporating contrast enhancement and parallel feature correlation mechanisms to enhance classification performance across multiple datasets.

To further improve diagnostic precision, several studies have focused on jointly optimizing segmentation and classification within unified deep learning pipelines. Balaha et al.³ proposed a framework that combines deep transfer learning with a sparrow search algorithm to optimize multiple U-Net variants for skin cancer diagnosis. Similarly, Sethanan et al.⁴ developed an integrated segmentation and CNN-based classification model using HAM10000 and malignant/benign datasets, achieving notably high diagnostic accuracy. Other works emphasized architectural innovations in classification networks. Tahir et al. proposed DSCC_Net⁵, which was evaluated across ISIC 2020, HAM10000, and DermIS datasets⁶, demonstrating competitive performance over multiple CNN backbones.

More recent research has explored feature fusion and multimodal learning strategies to address the complexity of skin cancer diagnosis. He et al.⁷ proposed a shared-attention fusion network that integrates dermoscopic and clinical images via dual branches and a super-branch, enabling multi-stage feature fusion. Chen et al.⁸ further extended this idea by introducing a multimodal data fusion diagnosis network that combines skin images with clinical data, achieving approximately 9% higher accuracy compared with single-modality approaches. Collectively, these studies indicate that deep learning–based diagnostic systems, particularly those integrating segmentation, attention mechanisms, and multimodal information, constitute a dominant and effective research direction in automated skin cancer diagnosis.

Single-setting diagnostic systems for skin cancer

Single-setting diagnostic systems aim to accomplish cancer classification within a unified end-to-end architecture or training pipelines. In addition to deep learning, machine learning has been widely applied to disease analysis and clinical decision support in healthcare systems. Early investigations systematically evaluated traditional machine learning and deep learning models for skin cancer diagnosis. Bechelli et al.⁹ conducted a comparative analysis across multiple dermoscopic datasets, demonstrating that deep learning models generally outperform classical machine learning methods. Subsequently, Imran et al.¹⁰ explored CNN-based architectures, including VGG, CapsNet, and ResNet, and reported varying diagnostic accuracies across the different model designs.

Single-setting diagnostic systems that rely on unified pipelines or optimized learning strategies have also attracted significant attention. Malibari et al.¹¹ proposed a single-key CAD system that integrates Wiener filtering, SqueezeNet feature extraction, and an improved whale optimization algorithm, achieving high classification accuracy. Alam et al.¹² evaluated several CNN architectures on the HAM10000 dataset and identified RegNetY-320 as the most effective model. Similarly, Lee et al.¹³ developed a ResNet-50-based architecture tailored for dermoscopic images, achieving a notable improvement in sensitivity. Transformer-based and dense connectivity models have also been explored, as demonstrated by Aladhadh et al.¹⁴ and Kousis et al.¹⁵, who reported strong performance on HAM10000¹⁶. Further architectural refinements include capsule network variants with attention mechanisms¹⁷, ResNet-based feature separation strategies¹⁸, and CNN-based lesion detection models achieving competitive accuracy¹⁹.

Machine learning–based disease analysis in healthcare systems

Beyond image-level diagnosis, machine learning has been widely applied to disease analysis and clinical decision support in healthcare systems. Beyond image-based diagnosis, machine learning methods have been employed for broader disease analysis and healthcare decision-making. Hu et al.²⁰ utilized Bayesian models and time-series analysis to predict long-term skin cancer burden. Spectral and signal-based diagnostic approaches have been investigated using Raman spectroscopy combined with CNN classifiers²¹ and hybrid optical–AI frameworks based on Stokes–Mueller matrix decomposition²². Classical machine learning pipelines integrating handcrafted features with SVMs and random forests remain relevant, as demonstrated in bag–of–features–based CAD systems²³ and hybrid equilibrium models²⁴. CNN-based decision support systems trained on large-scale datasets such as HAM10000 have also been reported²⁵. Comprehensive reviews further confirm that CNN-driven diagnostic models generally achieve accuracies exceeding 90%^26,27. More advanced learning paradigms, including reinforcement learning²⁸ and hybrid intelligent computing approaches applied to TCGA datasets²⁹, highlight the expanding role of machine learning in predictive diagnosis and clinical decision support. Recent studies have also explored integrated diagnostic frameworks and machine learning–driven approaches for cancer detection in healthcare systems. For example, integrated deep learning frameworks have been proposed to jointly segment and classify cancer lesions, enabling more accurate lesion localization and diagnosis³⁰. In addition, machine learning–based disease analysis methods have been widely applied in healthcare systems to support clinical decision making and improve diagnostic reliability³¹. Furthermore, unified diagnostic systems based on a single-setting framework have been investigated to simplify diagnostic pipelines and improve efficiency in medical image analysis tasks³². These studies highlight the growing interest in integrated, intelligent healthcare diagnostic frameworks for medical diagnosis.

In summary, existing studies on skin cancer diagnosis can be broadly categorized into three major directions: (i) integrated deep learning frameworks for segmentation and classification, (ii) single-setting diagnostic systems with unified pipelines, and (iii) machine learning–based disease analysis in healthcare systems. These approaches have achieved promising results in improving diagnostic accuracy and robustness. However, most existing methods rely on conventional convolutional architectures or isolated network designs, which may limit their ability to comprehensively capture complex lesion characteristics. Moreover, some models exhibit high computational complexity when processing high-resolution medical images, restricting their applicability in real-time clinical environments. In addition, uncertainty-aware refinement and adaptive decision-making mechanisms remain insufficiently explored in current studies.

Based on these observations, current research still faces challenges in balancing representation capability, computational efficiency, and diagnostic reliability. To address these limitations, the proposed Bicom framework integrates neuromorphic computing, Linformer-based attention, Feature Pyramid Network, ResNeSt, and CoAtNet within a unified architecture. Unlike conventional single-stage classifiers, Bicom incorporates a confidence-aware refinement mechanism to enable dynamic re-evaluation of ambiguous samples. This design supports efficient multi-scale feature learning, global contextual modeling, and adaptive decision refinement for skin cancer risk prediction. The effectiveness of the proposed framework is evaluated through the comparative and ablation experiments presented in the following sections.

Preliminary

This section mainly describes the Linformer, Feature Pyramid Network, ResNeSt, and CoAtNet that we used to propose the Bicom (A Skin Cancer Risk Prediction Model Based on Brain-Inspired Computing Technology) model.

Linformer

Linformer is an efficient improvement method for the self-attention mechanism of the Transformer model, aiming to solve the problems of high computational complexity and memory usage of traditional Transformer when processing long sequence data³³. The complexity of the traditional self-attention mechanism in time and space is the square of the sequence length, that is, O(n²), which means that when the sequence length increases, the amount of calculation and memory requirements will increase exponentially, limiting the Transformer Applications in tasks such as long text understanding, high-resolution image processing. The innovation of Linformer is introducing a low-rank linear projection method to reduce the dimensionality of the key and value matrices and map them from high-dimensional space to lower dimensions, thereby reducing the complexity of self-attention. The degree is reduced from O(n²) to O(n). This low-rank approximation method takes advantage of the redundancy of sequence data and retains the most crucial feature information while significantly reducing the amount of calculation and memory usage. Linformer introduces learnable projection matrices, which are optimized during the training process to adapt to different data distribution and task requirements, enhancing the flexibility and expressiveness of the model. In addition, the design of the Linformer does not require significant modifications to the overall architecture of the Transformer. Low-rank projection only needs to be applied in the self-attention module. This allows Linformer to integrate existing training techniques and pre-trained models easily. Linformer’s contribution is to significantly improve the efficiency of the Transformer model in long sequence processing, expand its application scope, and enable it to play a role in long sequence tasks in natural language processing, computer vision, and other fields while maintaining the performance of the model. and accuracy³³. This innovation provides new ideas for researching efficient self-attention mechanisms and promotes the continued development of deep learning regarding computational efficiency and model optimization.

Feature pyramid network

Feature Pyramid Network (FPN) is a deep learning architecture for improving the performance of object detection and segmentation tasks, which Tsung-Yi Lin et al. proposed in 2016. FPN was born to address the limitations of traditional convolutional neural networks (CNNs) in dealing with multi-scale objects because objects may appear in images of different scales, shapes, and postures. Traditional CNNs have rich semantic information in high-level features but low resolution, while low-level features have high resolution but weak semantic information³⁴. This characteristic causes the model to perform poorly when detecting small objects or dealing with objects with large-scale differences. The innovation of FPN is that it introduces a top-down feature fusion path and lateral connections to construct a feature pyramid structure with semantic solid information. Specifically, FPN uses the hierarchical features of CNN to upsample high-level feature maps layer by layer through a top-down path and fuse them with low-level feature maps of the corresponding scale. This fusion is achieved through lateral connections, usually element-wise addition, combining high-resolution spatial and high-level semantic information. As a result, each layer of the pyramid has rich semantic and detailed information and can effectively detect objects of different scales. The contributions of FPN are mainly reflected in the following aspects:

• Improved multi-scale object detection performance: By predicting objects at each level of the feature pyramid, FPN can simultaneously process objects of different scales in a single forward propagation, especially in small object detection.

• Efficient computational cost: FPN utilizes the existing CNN hierarchy and adds minimal overhead. It avoids the tedious process of multiple scaling of the input image required by traditional multi-scale methods, thereby improving computational efficiency.

• Strong versatility and scalability: FPN can seamlessly combine with various detection frameworks (such as Faster R-CNN, Mask R-CNN, and RetinaNet), proving its applicability in different models and tasks.

• Promoted subsequent research: The success of FPN has inspired many subsequent works, such as PANet, NAS-FPN, BiFPN, etc. These studies further improved the method of multi-scale feature fusion based on FPN and promoted the development of the object detection field.

• Wide range of application scenarios: Besides object detection, FPN is also widely used in tasks such as instance segmentation, semantic segmentation, key point detection, and pose estimation, reflecting its profound influence on computer vision.

In summary, Feature Pyramid Network successfully solves the shortcomings of traditional CNN in multi-scale object detection through clever architecture design. Its top-down feature fusion strategy and lateral connection mechanism enable the model to significantly improve the detection ability of objects of different scales while maintaining high efficiency. FPN has become a standard component in modern object detection and segmentation models, has had an essential impact on the research and application of computer vision, and has promoted the continuous development of multi-scale feature fusion methods³⁴.

ResNeSt

ResNeSt (Residual Networks with Split-Attention) is a convolutional neural network architecture with significant influence on computer vision. Its research background stems from the need for deep learning models’ deeper and broader feature expression capabilities when processing complex visual tasks³⁵. Traditional networks such as ResNet, ResNeXt, and SENet have improved model performance to a certain extent by increasing depth and width or introducing attention mechanisms. However, these methods still have room for improvement when processing more complex visual tasks, especially in improving the richness and robustness of feature representation while maintaining computational efficiency. The innovation of ResNeSt lies in the introduction of the Split-Attention module, which combines the grouped convolution idea of ​​ResNeXt and the channel attention mechanism of SENet to realize the dual attention mechanism for feature channels and feature groups. Specifically, ResNeSt first divides the feature channel into multiple subgroups (splits) and performs convolution operations in each subgroup. This process is similar to the grouped convolution of ResNeXt, which aims to increase the network’s width and improve the model’s expressiveness. Then, ResNeSt applies the attention mechanism within each subgroup and adaptively adjusts the influence of each subgroup feature through the learned weights. The contributions of ResNeSt are mainly reflected in the following aspects. First, ResNeSt has achieved excellent performance improvements in multiple computer vision tasks. In the ImageNet image classification task, ResNeSt achieved higher accuracy than the models of the same period (such as EfficientNet and ResNeXt). In downstream tasks such as object detection, instance segmentation, and semantic segmentation, using ResNeSt as the backbone also achieved significant performance improvements. Secondly, the Split-Attention module proposed by ResNeSt provides new ideas for the design of the attention mechanism and inspires subsequent research on multi-dimensional attention and efficient feature fusion. In addition, while maintaining high performance, ResNeSt has high computational efficiency and is suitable for various resource-constrained environments, and excessive computational costs will not limit its usage scenarios³⁵.

CoAtNet

CoAtNet (Convolutional and Attention Network) is a hybrid model that combines convolutional neural networks (CNN) with self-attention mechanisms (Transformers)³⁶. It aims to simultaneously utilize the locality of convolution and the globality of attention mechanisms to improve performance and efficiency in computer vision tasks. Its research background stems from the advantages and limitations of deep learning models in processing visual tasks: CNN is good at extracting local features and has beneficial inductive biases such as translation invariance and local connections, but it is insufficient in capturing global dependencies; while the Transformer model has global solid modeling capabilities and can effectively capture long-distance dependencies, but lacks the inductive bias of CNN, which may lead to poor performance on small data sets. The innovation of CoAtNet lies in the design of a unified architecture that organically combines convolution operations and self-attention mechanisms. Specifically, the model is divided into multiple stages. The first few stages mainly use convolutional layers to capture low-level local features, and the subsequent stages introduce self-attention mechanisms to focus on high-level global feature extraction. In addition, CoAtNet introduces technologies such as relative position encoding and depthwise separable convolution. Relative position encoding enables the model to capture spatial position information better and enhances the model’s robustness to position changes; deep separable convolution effectively reduces the number of parameters and computational costs and improves the model’s computational efficiency³⁶.

The contributions of CoAtNet are mainly reflected in the following aspects. First, it has achieved excellent performance in large-scale image classification tasks such as ImageNet, significantly surpassing traditional CNN and pure Transformer models and setting new performance records. Second, CoAtNet performs well in data efficiency and can train high-performance models on relatively small data sets, which is significant for practical application scenarios with limited resources³⁶. In addition, while maintaining high performance, the model has optimized computational efficiency and parameter scale, making it suitable for environments with limited computing resources. Finally, the architectural design of CoAtNet provides new ideas for integrating different deep learning mechanisms, demonstrating the great potential of hybrid models in computer vision tasks. Overall, CoAtNet overcomes both limitations by cleverly combining the local feature extraction capabilities of convolutional neural networks and the global modeling capabilities of self-attention mechanisms, achieving a dual improvement in performance and efficiency. Its success has provided a new direction for the design and optimization of deep learning models, has broad application prospects and significant reference value, and has had a positive impact on the development of the field of computer vision.

This section synthesizes the core methodological components that constitute the theoretical and architectural foundation of the proposed Bicom framework. Linformer enables scalable self-attention modeling of long-range dependencies with reduced computational complexity, while the Feature Pyramid Network facilitates hierarchical multi-scale feature representation essential for capturing lesion structures of varying sizes. ResNeSt enhances feature discrimination via split-attention mechanisms, thereby improving robustness to complex visual patterns. CoAtNet further integrates convolutional inductive biases with global self-attention, achieving a balanced combination of local detail extraction and contextual modeling.

More importantly, these components are not simply combined but are strategically coordinated within a unified architecture. The interaction between linear-complexity attention, hierarchical feature fusion, and hybrid convolution–attention modeling enables efficient and discriminative representation learning across diverse lesion characteristics. This synergistic design serves as the methodological basis for the proposed diagnostic model and supports reliable skin cancer risk prediction in auxiliary clinical applications.

Methodology

This section presents the methodological design of the proposed Bicom framework, including feature extraction, classification, confidence-aware refinement, and end-to-end optimization. Rather than naive model stacking, Bicom is designed as a hierarchical and confidence-driven decision architecture that integrates deep learning and neuromorphic computing for robust skin cancer risk prediction.

Multi-scale feature extraction based on improved ResNeSt

To achieve efficient and discriminative representation learning for skin lesion images, an improved ResNeSt-based feature extraction module, termed F-ResNeSt, is developed. The proposed module integrates Linformer-based linear-complexity attention with Feature Pyramid Network (FPN)–based multi-scale feature fusion into the original ResNeSt backbone, thereby enhancing both global dependency modeling and scale-aware lesion representation.

In conventional ResNeSt blocks, Split-Attention is employed to enhance channel-wise feature interaction. In contrast, the proposed L-ResNeStBlock replaces Split-Attention with Linformer attention to improve global contextual modeling with lower computational cost. The structure of L-ResNeStBlock is illustrated in Fig. 2. By introducing learnable low-rank projection matrices, Linformer approximates the self-attention operation with linear complexity in sequence length, significantly reducing computational and memory overhead when processing high-resolution dermoscopic images.

Fig. 2.

Architecture of the L-ResNeStBlock feature pyramid.

To further enhance scale adaptability, the Feature Pyramid Network (FPN) is incorporated to construct a hierarchical feature representation, as illustrated in Fig. 3. Each feature-extraction stage consists of stacked L-ResNeStBlocks, with lateral connections and a top-down pathway used to fuse high-level semantic information with low-level spatial features. This hierarchical fusion enables the model to capture lesion structures at multiple spatial resolutions. In addition, task-specific prediction heads are attached to multiple pyramid levels, allowing the network to perform classification using features extracted at different scales.

Fig. 3.

Architecture of the F-ResNeSt feature pyramid.

The combination of Linformer-based linear attention and FPN-based hierarchical fusion enables the proposed F-ResNeSt module to capture long-range contextual dependencies while preserving fine-grained spatial information. As a result, the feature extraction process achieves efficient global modeling, improved multi-scale representation capability, and better adaptability to lesions with diverse morphological characteristics.

Image classification based on improved CoAtNet

Although CoAtNet achieves strong performance in visual recognition by combining convolutional operations with self-attention mechanisms, its relative position attention often introduces considerable computational and memory overhead. In addition, maintaining an effective balance between global contextual modeling and local feature preservation remains challenging when processing high-resolution dermoscopic images in resource-constrained clinical settings.

To address these limitations, an improved classification network, termed L-CoAtNet, is developed by replacing CoAtNet’s relative attention mechanism with Linformer-based attention, as illustrated in Fig. 4. This modification reduces the computational complexity of attention from quadratic to linear in sequence length, thereby improving scalability while preserving effective contextual representation.

Fig. 4.

Architecture of the L-CoAtNet classification module.

Through linearized attention approximation, L-CoAtNet maintains stable contextual modeling ability across varying input resolutions and lesion patterns. At the same time, the simplified attention structure reduces memory usage and facilitates more efficient optimization during training. As a result, the proposed classification module provides scalable global modeling and computationally efficient feature discrimination for skin lesion classification.

Bicom framework

To achieve accurate and reliable skin cancer risk prediction, the proposed Bicom framework integrates F-ResNeSt, L-CoAtNet, and a brain-inspired Spiking Neural Network (SNN) module within a unified architecture, as shown in Fig. 5. Unlike conventional single-stage classifiers, Bicom explicitly introduces a confidence-aware refinement mechanism to improve the reliability of predictions for ambiguous samples.

Fig. 5.

Architecture of the Bicom framework.

In the proposed framework, input dermoscopic images are first augmented and preprocessed, and then divided into training, validation, and test subsets. F-ResNeSt extracts multi-scale feature representations, which are subsequently classified by L-CoAtNet. Let denote the extracted feature representation and the corresponding primary prediction.

A confidence score is computed from the classifier’s output to quantify the reliability of the prediction. When low-confidence samples are identified, the SNN module is selectively activated to refine the prediction via sparse, adaptive neural responses. This refinement mechanism enables dynamic re-evaluation of uncertain cases and helps reduce misclassification near decision boundaries.

Rather than naive model stacking, the proposed framework adopts a hierarchical decision structure in which feature extraction, classification, and refinement are functionally decoupled yet dynamically coupled through confidence-driven activation. This design improves robustness to noise, blur, and illumination variation, which are commonly observed in clinical dermoscopic images.

Bicom model algorithm

The end-to-end training procedure of the proposed Bicom framework is summarized in Algorithm 1. The algorithm explicitly defines data preprocessing, feature extraction, classification, confidence-aware refinement, loss computation, and parameter optimization steps.

Algorithm 1.

End-to-End Training of the Bicom Framework.

Algorithm 1 summarizes the end-to-end training procedure of the proposed Bicom framework. The algorithm begins with data preprocessing and dataset partitioning, followed by multi-scale feature extraction using the F-ResNeSt module. The extracted features are then processed by the L-CoAtNet classifier to produce the primary prediction . A confidence score is computed to evaluate the reliability of the prediction. For samples with confidence below a predefined threshold, a brain-inspired Spiking Neural Network module is selectively activated to perform refinement, enabling adaptive re-evaluation of ambiguous cases. Model parameters are optimized by minimizing the classification loss through gradient-based learning. Validation-guided model selection and early stopping are adopted to improve generalization and prevent overfitting. Finally, the optimized model is evaluated on the independent test set.

Mathematical formulation of the Bicom framework

Let denote the skin lesion dataset, represents an input dermoscopic image, and denotes the corresponding label indicating benign or malignant skin lesions. The proposed F-ResNeSt module maps the input image to a multi-scale feature representation, as shown in Eq. 1.

1

In Formula 1, represents the feature-extraction network parameters. By integrating a Linformer-based attention mechanism, self-attention operations can be approximated as shown in Eq. 2.

2

As shown in Eq. 2, W and H are learnable low-rank projection matrices, reducing the computational complexity from to . represents the unified representation obtained, i.e., features at multiple spatial scales further aggregated by the feature pyramid network. The extracted features are fed into the L-CoAtNet classifier to generate initial predictions, as shown in Eq. 3.

3

In Formula 3, represents the classifier parameters, and corresponds to the predicted probability. To enhance robustness, the confidence score is computed from the classifier’s output, as shown in Eq. 4.

4

In this paper, we use a predefined confidence threshold . If , we activate a brain-inspired spiking neural network (SNN) to improve the prediction, as shown in Eq. 5.

5

In Eq. 5, represents the SNN module’s parameters. This mechanism allows the model to capture subtle temporal and structural changes in blurred samples. The overall training objective of the Bicom model is defined as shown in Eq. 6.

6

In the classification loss function of Eq. 6, represents the complete set of model parameters. The Bicom model parameters are optimized through iterative gradient-based learning until convergence.

This section presents the methodological framework of the proposed Bicom model, including hierarchical feature extraction, efficient classification, confidence-aware refinement, mathematical formulation, and end-to-end optimization. Rather than a simple aggregation of existing modules, Bicom is constructed as a coordinated architecture that combines deep learning and neuromorphic computing to improve representation learning and adaptive decision refinement for skin cancer risk prediction.

Specifically, Sect. 4.1 introduces F-ResNeSt, which incorporates Linformer-based linear-complexity attention into ResNeSt blocks and integrates an FPN-based multi-scale fusion pathway to strengthen lesion feature representation across different spatial scales. Section 4.2 describes L-CoAtNet, in which the relative attention mechanism in CoAtNet is replaced with Linformer-based attention to reduce computational and memory overhead while preserving global contextual modeling capability for skin lesion classification. Section 4.3 further integrates F-ResNeSt and L-CoAtNet into the unified Bicom architecture and introduces a brain-inspired Spiking Neural Network (SNN) module that is selectively activated under low-confidence predictions to refine ambiguous cases and improve robustness against noise, blur, and illumination variation.

Sect. 4.4 summarizes the end-to-end algorithmic workflow, including data augmentation, preprocessing, dataset partitioning, iterative optimization, validation-guided model selection, and final evaluation. Section 4.5 then presents the mathematical formulation of the proposed framework by explicitly defining the feature extraction, classification, confidence estimation, refinement, and optimization processes, thereby establishing the theoretical basis for the design and training of the Bicom model.

Experimental results and analysis

This section presents a systematic experimental evaluation of the proposed Bicom framework, including dataset description, evaluation metrics, comparative analysis with representative baselines, ablation studies, and parameter optimization experiments.

Datasets

This subsection describes the dataset and experimental settings used to evaluate the proposed Bicom framework.

All experiments were conducted on dermoscopic images obtained from the International Skin Imaging Collaboration (ISIC) archive³⁷. The dataset contains two categories, benign and malignant skin lesions, with 1,800 benign images and 1,497 malignant images. All images were resized to a unified resolution of 224 × 224 pixels for model training and evaluation.

The ISIC archive is a widely used public benchmark for skin lesion analysis and provides standardized dermoscopic images for research on automated skin cancer diagnosis. In this study, the dataset serves as the primary benchmark for evaluating the proposed framework under controlled experimental settings.

Prior to training, all images underwent standardized preprocessing and data augmentation, including normalization, resizing, and geometric transformations. The dataset was randomly partitioned into training, validation, and test subsets at a ratio of 6:2:2. To improve the reliability and reproducibility of the reported results, all experiments were repeated 5 times with different random seeds, and the final performance values are reported as mean ± standard deviation.

Baseline models for comparison

To provide a comprehensive evaluation of the proposed framework, several representative models from the related literature were selected as baseline methods. These models include widely used convolutional neural networks and hybrid architectures that have been applied to skin cancer or medical image classification tasks.

Table 1 summarizes representative baseline models, their architectural characteristics, and the performance reported in the original studies. These models cover various deep learning approaches, including conventional convolutional neural networks, dense connectivity networks, capsule networks, and multi-feature fusion architectures. The reported accuracy values correspond to the results presented in the original studies and are included here to summarize representative baseline methods in the literature.

Table 1.

Representative baseline models used for experimental comparison.

Model	Reference	Key Architecture	Reported Accuracy
Xception	Bassel3	CNN-based architecture	90.9%
Vgg-19	Tahir, M5	Deep convolutional network	92.51%
ConvNet	Sharma, A. K6	Convolutional neural network	98.3%
Fusion Network	He, X7	Multi-feature fusion network	76.8%
VGG	Imran10	Deep CNN architecture	79%
DenseNet169	Kousis, I15	Dense convolutional network	92.25%
Capsule Network	Lan, Z17	Capsule-based deep model	96.49%
CNN	Siddique19	Convolutional neural network	96.25%

In our experiments, several representative and widely used models, including EfficientNetV2, ResNeSt, and CoAtNet, were used as baselines under identical experimental settings to ensure a fair comparison with the proposed Bicom framework.

Evaluation metrics

We used the following indicators and AUC to evaluate the skin cancer risk prediction model Bicom³⁸.

7

8

9

10

11

The above five formulas are accuracy, precision, recall³⁹, F1 score, and specificity. TP stands represents true positive, TN stands for true negative, FP stands for false positive, and FN stands for false negative⁴⁰. In the experiment, AUC was used to evaluate the practical effect of the skin cancer risk prediction model in an all-around way^47–49.

Effect comparison experiment

To comprehensively evaluate the effectiveness of the proposed Bicom framework, comparative experiments were conducted against several representative baseline models. EfficientNetV2, ResNeSt, and CoAtNet were selected as widely adopted convolutional and hybrid-attention architectures for image classification tasks. In addition, three previously proposed disease risk prediction models—Diplin, Linge, and Motico—were included to further evaluate the general applicability of the proposed framework across different model designs.

All baseline models were implemented and trained under identical experimental settings to ensure a fair comparison. Model performance was evaluated using multiple metrics, including Accuracy, Precision, Recall, F1-score, Specificity, and the Area Under the Receiver Operating Characteristic Curve (AUC). To reduce the influence of random initialization, each experiment was repeated five times with different random seeds, and the final results are reported as mean ± standard deviation. The comparative results on the public dataset are presented in Fig. 6.

Fig. 6.

Comparative experimental results of the Bicom model on public datasets.

As illustrated in Fig. 6, all baseline models—including EfficientNetV2, ResNeSt, CoAtNet, Diplin, Linge, and Motico—achieve competitive performance on the skin cancer risk prediction task. Among these methods, the proposed Bicom framework consistently achieves the best overall performance across multiple evaluation metrics. In particular, Bicom demonstrates measurable improvements over the strongest baseline model, CoAtNet, in Accuracy, Precision, Recall, Specificity, and AUC, indicating its enhanced ability to discriminate between malignant and benign skin lesions.

To further evaluate the robustness and generalization capability of the proposed framework, additional comparative experiments were conducted on another dataset. The detailed quantitative results are reported in Table 2.

Table 2.

Comparison of experimental results of Bicom model on subject dataset.

Basic algorithm	Accuracy and standard deviation	Precision and standard deviation	Recall and standard deviation	Specificity	AUC
EfficientNetV241	0.8875 ± 0.00398	0.8509 ± 0.00698	0.8703 ± 0.00613	0.8752	0.8792
ResNeSt35	0.8906 ± 0.00567	0.8593 ± 0.00631	0.8759 ± 0.00552	0.8867	0.9036
Diplin42	0.9076 ± 0.00329	0.8692 ± 0.00592	0.8897 ± 0.00539	0.9003	0.9201
Linge43	0.9137 ± 0.00512	0.8896 ± 0.00397	0.9089 ± 0.00513	0.9197	0.9396
Motico44	0.9208 ± 0.00537	0.9098 ± 0.00326	0.9218 ± 0.00506	0.9369	0.9531
CoAtNet36	0.9326 ± 0.00397	0.9258 ± 0.00296	0.9397 ± 0.00308	0.9512	0.9608
Bicom	0.9457 ± 0.00362	0.9367 ± 0.00237	0.9532 ± 0.00307	0.9681	0.9766

A paired t-test was conducted between Bicom and the strongest baseline, CoAtNet, using results obtained from multiple independent runs. The improvements in Accuracy, Precision, Recall, and AUC were statistically significant (). As shown in Table 2, Bicom consistently outperforms all baseline models across all evaluation metrics on the subject dataset. Compared with CoAtNet, Bicom achieves relative improvements of 1.31% in Accuracy, 1.09% in Precision, 1.35% in Recall, 1.69% in Specificity, and 1.58% in AUC. These results suggest that the observed performance gains are stable and reproducible rather than due to random variation, thereby supporting the reliability of the proposed framework for skin cancer risk prediction.

Ablation comparison experiment

To quantify the contributions of individual components within the Bicom framework, a systematic ablation study was performed. All ablation variants were trained and evaluated under identical experimental settings to ensure fair comparison.

Specifically, five model variants were constructed by progressively incorporating the key architectural components of the proposed framework. The first variant employed the original ResNeSt backbone for feature extraction and CoAtNet for classification. The second variant introduced Linformer-based attention into both ResNeSt and CoAtNet. The third variant incorporated the Feature Pyramid Network (FPN) into the ResNeSt backbone while retaining the original CoAtNet classifier. The fourth variant combined Linformer-based attention with FPN-enhanced feature extraction and Linformer-based classification. The fifth variant further introduced a brain-inspired SNN module for confidence-aware refinement, completing the Bicom framework.

By progressively adding these components, the ablation design enables controlled evaluation of the effects of linear-complexity attention, hierarchical multi-scale fusion, and neuromorphic refinement on the final prediction performance. The ablation results obtained on the public dataset are presented in Fig. 7.

Fig. 7.

Ablation comparison experiment results of the Bicom model on public datasets.

The experimental results in Fig. 7 indicate that both Linformer-based attention and the Feature Pyramid Network improve the performance of the proposed framework. Compared with the baseline configuration, introducing Linformer improves global contextual modeling, leading to more discriminative feature representations. The incorporation of the FPN module further enhances hierarchical multi-scale feature fusion, which is particularly beneficial for recognizing lesions with varying sizes and irregular morphologies.

In addition, the SNN-based refinement module yields further performance improvements by selectively re-evaluating low-confidence samples, thereby reducing misclassification near decision boundaries. The progressive performance gains observed across the ablation variants provide empirical evidence that Linformer-based attention, multi-scale fusion, and confidence-aware neuromorphic refinement contribute complementary benefits to the robustness and predictive accuracy of the Bicom framework.

Parameter optimization experiment

During training, model parameters were optimized by minimizing the classification loss between predicted outputs and ground-truth labels. Model performance on the validation set was continuously monitored to guide hyperparameter adjustment and prevent overfitting. Training was terminated when the convergence criterion was satisfied.

To improve model stability and generalization, systematic parameter optimization was performed by tuning key hyperparameters, including the optimizer type, initial learning rate, learning rate scheduler, batch size, regularization strength, and loss-function configuration. These hyperparameters were selected through controlled experiments based on validation performance.

In addition, architectural parameters associated with Linformer, FPN, ResNeSt, CoAtNet, and the SNN module were jointly adjusted to balance representation learning capacity and classification efficiency. Figure 8 presents a quantitative comparison of model performance before and after parameter optimization, showing consistent improvements across multiple evaluation metrics.

Fig. 8.

Comparison of accuracy, precision, recall, specificity, F1 score, and AUC value before and after optimization of the Bicom model. (a) Comparison of accuracy, precision, and recall. (b) Comparison of specificity, F1 score, and AUC value.

Parameter optimization experiments indicate that appropriate configurations of learning rate⁴⁵, batch size, loss function, and regularization strategies play important roles in model convergence and generalization⁴⁶.As shown in Fig. 8, the optimized Bicom model achieves consistent improvements across all evaluation metrics. These results suggest that validation-guided hyperparameter tuning contributes to stable learning dynamics and effective overfitting control.

This section systematically evaluates the proposed Bicom framework through comprehensive experiments, including dataset description, evaluation metrics, comparative analysis with representative baselines, ablation studies, and parameter optimization. Comparative experiments on both public and subject datasets show that Bicom maintains consistently competitive performance across multiple evaluation metrics. The ablation studies further verify the contributions of individual architectural components, indicating that Linformer-based attention, multi-scale feature fusion via FPN, and the brain-inspired SNN module provide complementary benefits to model performance. In addition, parameter-optimization experiments show that appropriately tuned learning strategies and model hyperparameters improve stability and generalization.

Overall, these experimental results provide empirical support for the effectiveness of the proposed Bicom framework. The observed performance improvements are consistent with the integrated design of linear-complexity attention, hierarchical feature fusion, and confidence-aware neuromorphic refinement. Compared with conventional skin cancer diagnosis models based primarily on standard convolutional backbones or quadratic-complexity self-attention, the proposed framework combines F-ResNeSt, L-CoAtNet, and an adaptive SNN-based refinement mechanism within a unified end-to-end architecture, enabling efficient and robust skin cancer risk prediction.

Discussion

In this section, we discuss the proposed Bicom model from three aspects: its advantages, limitations, and potential directions for future work.

Advantages of the proposed Bicom model

The proposed Bicom framework exhibits several strengths in skin cancer risk prediction for auxiliary diagnostic applications. First, by integrating Feature Pyramid Network (FPN), ResNeSt, Linformer, and CoAtNet within a unified architecture, Bicom enables coordinated multi-scale feature extraction and global contextual modeling. The incorporation of FPN enhances hierarchical feature representation, which is particularly beneficial for handling skin lesions with diverse sizes and morphological variations. This integrated design facilitates the capture of both local structural patterns and broader contextual information from dermoscopic images.

Second, adopting the Linformer attention mechanism reduces the computational complexity of self-attention from quadratic to linear with respect to sequence length. This modification improves computational efficiency while preserving effective representation capability, contributing to more scalable model design for high-resolution medical images.

Third, incorporating a brain-inspired Spiking Neural Network (SNN) module introduces a confidence-aware refinement mechanism. By selectively activating the SNN under low-confidence conditions, the framework provides an adaptive strategy for handling ambiguous cases, which is particularly relevant in medical risk prediction scenarios where uncertainty management is critical.

Overall, the experimental findings indicate that the integrated design of multi-scale feature fusion, linear-complexity attention, and neuromorphic refinement contributes to consistent improvements in predictive performance and robustness. These characteristics suggest that the Bicom framework provides a reliable and extensible solution for assisted skin cancer risk prediction.

Limitations of the proposed model

Despite its promising performance, the proposed Bicom framework has several limitations that merit further investigation. First, the experimental evaluation in this study is primarily conducted on publicly available datasets, which may not fully capture the diversity and complexity of real-world clinical environments. Variations in imaging devices, acquisition protocols, and patient demographics may influence model generalization. Future work will focus on validating the proposed framework using large-scale multi-center clinical datasets to further assess its real-world applicability.

Second, although integrating multiple advanced architectures enhances predictive performance, it also increases model complexity and computational cost. This reflects a typical trade-off between accuracy and efficiency in deep learning–based medical systems. Future research will explore lightweight model compression, pruning, and hardware-aware optimization strategies to improve deployment feasibility in resource-constrained settings.

Third, the current study focuses on skin cancer risk prediction as a representative application scenario. While the proposed framework demonstrates strong performance in this context, its generalizability to other disease risk prediction tasks remains to be systematically evaluated. Extending the proposed architecture to multi-disease and multimodal diagnostic applications represents an important direction for future research.

Future work

Future research will focus on addressing the above limitations and further improving the proposed Bicom framework. First, additional validation will be conducted on larger and more diverse clinical datasets, including multi-center and real-world data, to more comprehensively evaluate model generalization and clinical applicability. Second, model simplification and optimization strategies will be investigated to reduce computational and memory overhead while preserving predictive accuracy. These efforts may include lightweight network design, model compression, and hardware-aware optimization to facilitate practical deployment.

Third, the proposed framework will be extended to support risk prediction for multiple diseases in auxiliary diagnostic scenarios. By adapting the feature extraction and classification modules, Bicom can be generalized to a broader range of medical imaging and healthcare applications. Finally, more advanced brain-inspired computing mechanisms and adaptive learning strategies will be explored to further enhance uncertainty modeling and decision reliability, thereby contributing to the development of more robust and trustworthy medical decision support systems.

Conclusion

In this study, we proposed Bicom, an integrated skin cancer risk prediction framework designed for auxiliary diagnostic applications. The framework combines linear-complexity attention, hierarchical multi-scale feature fusion, and a confidence-aware neuromorphic refinement mechanism within a unified end-to-end architecture. By incorporating Linformer-enhanced F-ResNeSt for efficient feature extraction, L-CoAtNet for scalable classification, and an adaptive SNN module for low-confidence refinement, the proposed approach achieves coordinated representation learning and robust decision enhancement.

Comprehensive experiments on public and subject datasets demonstrate that Bicom consistently achieves competitive performance across multiple evaluation metrics. Ablation studies further confirm that integrating linearized attention, multi-scale fusion, and brain-inspired refinement provides complementary benefits for predictive robustness and generalization capability.

Although the proposed framework demonstrates strong performance for skin cancer risk prediction, several limitations remain. Future research will focus on validating the model using large-scale multi-center clinical datasets, exploring lightweight optimization strategies for efficient deployment, and extending the framework to broader multi-disease diagnostic scenarios. Furthermore, advanced brain-inspired learning mechanisms and uncertainty modeling strategies will be investigated to further enhance decision reliability in medical risk prediction tasks.

Overall, the proposed Bicom framework provides an efficient, extensible, and robust solution for assisted skin cancer risk prediction, offering practical value for intelligent medical decision support systems.

Author contributions

F.Z., and S.H., wrote the main manuscript text, T.Z. provided the idea. X.D. prepared the data and figures. N.L. provided guidance on the writing of the manuscript. All authors reviewed the manuscript. The authors read and approved the final manuscript. All authors have read and agreed to the published version of the manuscript.

Data availability

The dataset is from https://www.kaggle.com/datasets/rm1000/skin-cancer-isic-images.

Declarations

Competing interests

The authors declare no competing interests.

Consent to participate

This study did not involve any human participants or personal data. Therefore, no informed consent or institutional review board/ethics committee approval was required.

Consent to publish

No personal or identifiable data are included in this manuscript; hence, no consent for publication was necessary.

Ethical statement

No ethical approval was needed because this is not a human study, but only online information was used.

Scientific responsibility statement

The authors declare that they are responsible for the article’s scientific content including study design, data collection, analysis and interpretation, writing, some of the main line, or all of the preparation and scientific review of the contents and approval of the final version of the article.

Animal and human rights statement

All procedures performed in this study were in accordance with the ethical standards of the institutional and/or national research committee and with the 1964 Helsinki declaration and its later amendments or comparable ethical standards. No animal or human studies were carried out by the authors for this article.