Cross-Shuffled Fusion Based Multi-Modal Hybrid Deep Learning Network for Breast Cancer Detection

Globally, over 2.3 million women will be diagnosed with breast cancer (BrC) by 2020, representing more than a quarter of all cancers among women. In its diagnosis, treatment, and prognosis, BrC is a highly heterogeneous disease [1]. Tumor histology techniques are enabling the detection of prognostic and phenotypic information about tumors. It is difficult to determine histology objectively, as different observers can come to differing conclusions [2]. Patients' quality of life may be negatively affected by inappropriate treatment and complications that arise from conventional pathologic analysis. BrC recurrences are common in the early stages of the disease, which necessitates proactive and appropriate treatments [3]. In spite of this, predicting whether a patient will recur remains one of the biggest challenges patients facing early-stage BrC face. For diagnosis and prognosis of BrC patients, molecular tests are not readily available because of cost and other barriers [4]. A cost-effective method for stratifying patients for surgery may be available through digital pathology analysis of their histopathology images [5].

The heterogeneity of BrC means that there is a need for timely detection of any anomaly in the tissue in order to make correct clinical decisions. Mammography can provide information about the structural characteristics of breast abnormalities while histopathological analysis provides information regarding cellularity of such abnormalities. However, analyzing these complementary images independently may restrict the potential of an automated system to learn all characteristics of the breast lesion. Integration of data obtained from mammography and histopathology can provide a more comprehensive understanding of the breast abnormalities and can help in enhancing the reliability of computer-aided detection of BrC. Driven by such clinical requirements, this work introduces Cross-Shuffled Fusion based Multi-Modal Hybrid Deep Learning Network (CSFHD-Net), a novel multimodal deep learning approach to fuse complementary information from mammogram and histopathology images.

BrC diagnosis, treatment, and prognosis can be greatly improved by deep learning (DL). Diagnostic accuracy and efficiency are greatly enhanced by the use of DL models through enhancement of image quality and rapid segmentation of lesions [6]. Automated and accurate tumor detection and segmentation can be accomplished using these models utilizing Convolutional Neural Networks (CNNs) [7]. Treatments based on DL technology include staging and subtyping BrC based on image-based techniques [8]. By identifying different tumor features and classifying them into different cancer subtypes, these models provide a foundation for developing personalized treatment plans [9]. In addition, multi-omics data can also be incorporated into DL models to help doctors choose the most suitable treatment regimen for their patients. DL technology was used to build robust and generalizable prognostic models based on large-scale databases [10].

BrC remains one of the most complex and challenging diseases to diagnose and manage. Although DL has advanced diagnostic, prognostic, and treatment planning strategies, several key limitations persist [11]. Conventional single-modality approaches often suffer from noise, class imbalance, and poor generalizability, reducing their robustness in real-world clinical settings. Moreover, most existing methods are confined to binary classification (normal vs. abnormal) and fail to provide consistent sub-classification into benign and malignant categories, which is crucial for clinical decision-making. Histopathological evaluation, while essential for assessing tumor heterogeneity is subjective and prone to inter-observer variability, whereas molecular tests for precise stratification remain expensive and inaccessible in many healthcare systems. These challenges highlight a cost-effective and reliable diagnostic framework of conventional single-modality approaches. The proposed CSFHD-Net integrate complementary information from mammography and histopathology to enable accurate and clinically meaningful BrC detection and classification. The main contribution of the research are as follows:

(1) An Advanced 3D Gabor filter is employed in the preprocessing phase to simultaneously reduce noise and enhance texture quality ensuring that critical cancer features are preserved for reliable analysis.

(2) A Shuffled with MobileNet is introduced for separate feature extraction from mammography and histopathology images by enabling lightweight yet highly discriminative representation of modality-specific patterns.

(3) A novel Cross-Shuffled Fusion strategy is proposed to effectively integrate multimodal feature sets while preserving inter-modality correlations, reducing redundancy, and generating a unified and robust feature representation.

(4) The fused features are further optimized using a combination of Bidirectional Gated Recurrent Unit (Bi-GRU) and Stacked Autoencoder (SAE), which capture sequential dependencies, and enable accurate classification of normal, benign and malignant cases.

The rest of the work was pre-systematized as follows: Section 2 examines the recent works on BrC classification. Section 3 contains the clear description of the proposed CSFHD-Net for different BrC cases. The outcomes of the proposed CSFHD-Net along with a contrast to the proposed and existing methods are determined in Section 4. The conclusion encloses with Section 5.

In this research, a novel CSFHD-Net has been proposed for leverages mammogram and histopathology images for accurate diagnosis. In the pre-processing phase, noise is reduced using a 3D Gabor filter by enhancing texture quality and preserving critical tumor features. For training, modality-specific feature extraction is performed by Shuffled-MobileNet separately for histopathology and mammograms.

These features are integrated using a Cross-Shuffled Fusion mechanism, designed to preserve inter-modality correlations and generate a unified fused vector representation. During testing, images are processed through Bi-GRU networks with SAE, which refine feature representations and reduce dimensionality. These SAE branches are concatenated and classified normal and abnormal tissue, followed by sub-classification of abnormal cases is performed to differentiate benign or malignant. The overall workflow of the proposed CSFHD-Net shown in Figure 1.

tu_pian_14.png

Figure 1. Overall workflow of the proposed Cross-Shuffled fusion based Multi-Modal Hybrid Deep Learning Network (CSFHD-Net)

3.1 Dataset description

In this research Curated breast imaging subset of digital database for screening mammography (CBIS-DDSM) is used for BrC classification. CBIS-DDSM is better suited for DDSM mammography datasets 33 Once decompressed, images can be viewed on a computer in DICOM format. DICOM data was converted to PNG files and processed according to CBIS DDSM guidelines. This research aimed to develop a classification scheme to separate benign from malignant images.

The BACH dataset contains of microscopy images meticulously categorized by medical experts. The annotation process included two medical experts assessing each image, and any disagreements led to the image being excluded from the dataset. Microscopy images consisted of 400 images in total, evenly distributed between four classes, each containing 100 images.

.tiff images contained specific characteristics: red–green–blue (RGB) color model, 2048 × 1536-pixel dimensions, and 0.42 μm × 0.42 μm pixel scale. Furthermore, each image required between 10 and 20 MB of memory. The dataset was labelled image-wise by providing detailed information regarding each cancer type. An image collection of BrC histopathology images was utilized for training and evaluating DL models.

BACH Dataset Class Distribution: The BACH dataset consists of 400 histopathological images equally distributed among four initial diagnostic categories: Normal, Benign, In situ carcinoma, and Invasive carcinoma, with 100 images in each category. In this study, the final classification task was formulated as a binary classification of Benign and Malignant. The Benign category was retained as the benign class, while in situ carcinoma and Invasive carcinoma were grouped into the malignant class. The Normal category was excluded because the objective of the study is to distinguish benign from malignant BrC tissue. Thus, the resulting BACH classification comprises 100 benign and 200 malignant images. No additional class-balancing technique was applied, and the resulting class distribution was retained during the experiments.

Dataset Selection: The Curated Breast Imaging Subset of the Digital Database for Screening Mammography (CBIS-DDSM) and Breast Cancer Histology images (BACH) datasets were chosen to complement each other in terms of mammographic and histopathological image modalities. While CBIS-DDSM offers mammographic images of breast lesions, BACH offers microscopic tissue-level images from histopathological images. The two datasets, thus, offer modality-specific image representations, which can be used for extracting features to be fused using the proposed CSFHD-Net framework. It is important to note that CBIS-DDSM and BACH are independent public datasets and do not offer paired mammographic and histopathological images for the same patients. Thus, the proposed framework performs feature-level fusion of modality-specific features instead of patient-level multimodality fusion. While these datasets offer useful benchmarks for evaluating the effectiveness of the BrC classification algorithms, they might not fully represent the real-world conditions, where there could be differences in patient demographic characteristics, imaging instruments, acquisition procedures, institutional practices, and histopathological preparation methods. Thus, evaluation using larger and multiple institutional datasets is required.

Diversity of the Datasets and Their Drawbacks: Despite the availability and popularity of the CBIS-DDSM and BACH datasets, they have limitations in representing the full real-world clinical diversity of BrC patients. These datasets do not provide sufficiently detailed demographic information to assess their representativeness across different age groups and racial or ethnic populations. In addition, variations in imaging equipment, acquisition protocols, institutional practices, and data-labelling procedures may not be fully represented. CBIS-DDSM consists of mammographic examinations, whereas BACH contains expertly labelled histopathological images, enabling the investigation of complementary structural and cellular information. Nevertheless, the use of these datasets does not encompass the full variability observed across different hospitals, patient populations, scanners, staining procedures, and image acquisition conditions. Other publicly available datasets, such as BreaKHis for histopathological images and mini-DDSM for mammographic images, can also be considered above in Table 1; however, these datasets may have similar limitations regarding demographic diversity and acquisition variability. Therefore, the high performance achieved by CSFHD-Net should be interpreted within the context of the datasets used in this study. Validation using larger, more diverse, and multi-institutional datasets is necessary to further establish the generalizability, robustness, and clinical applicability of the proposed framework.

Table 1. Description for the Curated Breast Imaging Subset of the Digital Database for Screening Mammography (CBIS-DDSM) dataset

3.2 Three-dimensional Gabor filter formulation for 2D multimodal images

In the preprocessing stage, a 3D Gabor filter is employed to suppress noise while enhancing texture information and preserving critical tumor features in mammography and histopathology images. While conventional 2D Gabor filters effectively extract spatial frequency features, they process multi-channel information independently, thereby failing to capture inter-channel cross-correlations and multi-scale depth relationships. To overcome this limitation, a 3D Gabor filter is utilized to capture joint spatial, chromatic, and multi-scale textural patterns.

For 2D multimodal medical images, the 3D tensor representation $\text{I(x,y,z)}$ is constructed as follows:

Histopathology Images: The 2D RGB biopsy images (H × W × 3) are mapped such that (x, y) represent spatial coordinates and z ∈ {R, G, B} represents the color channel axis. The 3D Gabor filter performs volumetric spatial-spectral convolution, extracting coherent structural boundaries while preserving colorimetric stain distributions essential for distinguishing hematoxylin-stained nuclei from eosin-stained stroma.

Mammogram Images: The 2D grayscale mammograms (H × W) are transformed into a multi-scale tensor (H × W × D), where the z-axis (z ∈ {1, 2 …, D} corresponds to multi-resolution scale decomposition layers. This enables simultaneous multi-frequency directional texture enhancement and quantum noise suppression.

The mathematical formulation of the three-dimensional Gabor filter $G\left( x,y,z \right)$ is computed as follows:

$G\left( x,y,z \right)=\frac{1}{{{(2\pi )}^{3/2}}{{\sigma }_{x}}{{\sigma }_{y}}{{\sigma }_{z}}}exp\left( -\frac{1}{2}\left[ {{\left( \frac{{{x}'}}{{{\sigma }_{x}}} \right)}^{2}}+{{\left( \frac{{{y}'}}{{{\sigma }_{y}}} \right)}^{2}}+{{\left( \frac{{{z}'}}{{{\sigma }_{z}}} \right)}^{2}} \right] \right)exp\left( j2\pi \left( {{f}_{x}}x+{{f}_{y}}y+{{f}_{z}}z \right)+\phi  \right)$      (1)

Here, (σ_x,σ_y,σ_z) denote the Gaussian envelope bandwidths along the spatial and channel dimensions, (f_x,f_y,f_z) represent the spatial-spectral frequencies, and ϕ is the phase offset. The enhanced filtered representation $I_{\text texture}(x, y, z)$ is obtained by convolving the 3D image tensor with the 3D Gabor filter:

$I_{\text texture}(x, y, z)=I(x, y, z) \otimes G(x, y, z)$         (2)

This 3D Gabor formulation ensures that both macro-scale mammographic spicules/micro calcifications and micro-scale cellular textures retain their critical structural and textural characteristics during preprocessing. This directly strengthens the feature quality for subsequent multimodal fusion in CSFHD-Net, thereby improving classification accuracy.

3.3 Pre-processing module Cross-Shuffled Fusion based Multi-Modal Hybrid Deep Learning Network

The proposed CSFHD-Net Employs Shuffled-MobileNet as a lightweight backbone to extract discriminative features separately from mammogram and histopathology images. It ensures efficiency while reducing computational complexity through depth-wise separable convolutions (DSC) and channel shuffle attention. A CSFN module is then used to integrate modality-specific features and enhance inter-modality correlations. It also reduces redundancy and enhances cross-attention. Lastly, the classification network receives the optimized feature representation, enabling high accuracy differentiation between benign and malignant BrC. Architecture of CSFHD-Net shown in Figure 2.

tu_pian_15.png

Figure 2. Architecture of Cross-Shuffled Fusion based Multi-Modal Hybrid Deep Learning Network (CSFHD-Net)

3.3.1 Shuffled-Mobilenet for feature extraction

Shuffled-MobileNet extracts features for histopathology and mammograms separately, resulting in feature sets Fhis and Fmam. A pre-trained feature extraction network with numerous characteristics that drastically reduce the model size without sacrificing effectiveness is the MobileNet model. Convolutions that form the foundation of a Mobile-Net are depth-wise separable and point-wise separable. It reduces the computational requirements to separate the spatial and channel dimensions. In order to balance model size and performance, the MobileNet model relies heavily on the usage of width multipliers to regulate the network's width. Using an additional option, a resolution multiplier regulates the input resolution of MobileNet.

MobileNet was pre-trained as a transfer learning model using an ImageNet and other large-scale datasets. The MobileNet model is a factorized convolution that converts a conventional convolution into a depth-wise, one- dimensional convolution called a pointwise convolution by using DSC. The depthwise convolution trains each input channel separately using a single filter. Utilize dimension pointwise convolution to merge these distinct filters. A traditional convolution creates a new set of outputs by combining and filtering inputs in a single step. These outputs are then split into two layers by the DSC. This process produces a large reduction in both processing and model size. Two square feature maps are produced by a conventional convolution layer: one for input F with size $D F \times D F \times N$ and one for output $G$ with size $D G \times D G \times M$. The $D F$ and $D G$ stand for the input and output feature map spatial dimensions (width and height). The convolution kernel K of size $D k \times D k \times N \times M$ describes the CNN layer parameters, where $D k$ stands for the kernel's spatial breadth and output channels' height, respectively. The output feature map for standard convolution is calculated using the formula below:

${{G}_{k,l,m}}=\underset{i,j,n}{\mathop \sum }\,{{K}_{i,j,m,n}}.{{F}_{k+i-1,i+j-1,n}}$         (3)

The input and output channels N and M, the feature map size (D_F), the kernel size (D_k), and the computational cost are all multiplicative. All of these elements and their interconnections are covered by MobileNet models. First, DSC were used to isolate the relationship between kernel size and output channel numbers. Depth-wise and point wise convolutions are the two tiers of DSC. Depth-wise convolutions are used to obtain the filter for each input channel. Pointwise convolution, a straightforward 1 × 1 convolution, is then used to linearly aggregate the output from the depth wise layer. For every layer, MobileNet uses either the batch norm or the ReLU nonlinearity.

$DS{{C}_{\text{cost}}}={{D}_{k}}\times {{D}_{k}}\times N\times {{D}_{F}}\times {{D}_{F}}+N\times M\times {{D}_{F}}\times {{D}_{F}}$          (4)

Convolution is a two-phase process that involves filtering and combining, which lowers the computing cost. Channel-wise dependencies were fully captured by using the SE block. In terms of balancing speed and accuracy, it will introduce an excessive number of factors, which is detrimental to the creation of a more lightweight attention module. Furthermore, faster 1-D convolutions of size k, ECA are not suitable for producing channel weights since k is typically greater. To improve the global information, channel-wise statistics like $s \in \mathbb{R} \mathrm{C} / 2 \mathrm{G} \times 1 \times 1$ was obtained by employing global averaging pooling (GAP), which was calculated by reducing ${{X}_{k1}}$ through spatial dimension $H \times W$.

$s={{F}_{gp}}\left( {{X}_{k1}} \right)=\frac{1}{H\times W}\overset{H}{\mathop \sum }\,\underset{k1}{\mathop \sum }\,\left( i,j \right)i={{1}_{j=1}}$          (5)

Additionally, to facilitate assistance for accurate and flexible choosing, a compact feature is developed. This is accomplished by sigmoid activation and a straightforward gating mechanism. The final channel attention output can then be obtained by:

$X_{\text{k}1}^{'}=\sigma \left( {{F}_{c}}\left( s \right) \right)\cdot {{X}_{k1}}=\sigma \left( {{W}_{1}}s+{{b}_{1}} \right)\cdot {{X}_{k1}}$           (6)

Efficiently extract lightweight yet highly discriminative features from both mammogram and histopathology images. The proposed feature sets Fhis and Fmam are designed to enhance the accuracy and robustness of BrC classification through a Cross-Shuffled Fusion process.

The CSFHD-Net framework processes mammogram and histopathology images through separate Shuffled-MobileNet branches to extract modality-specific features. The CSFN module integrates these features using cross-attention, feature interaction, channel shuffling, and concatenation for effective multimodal fusion in Figure 3. The fused features are refined using Bi-GRU, compressed through SAE, and classified into Normal/Abnormal and Benign/Malignant categories.

tu_pian_16.png

Figure 3. Annotated architecture of the Cross-Shuffled Feature Fusion (CSFN)

3.3.2 Cross-Shuffle Feature Fusion

The extracted features are integrated using a Cross-Shuffled Fusion Network (CSFN) designed to preserve inter-modality correlations and generate a unified fused vector representation. The CSFN enhances reciprocal learning between histopathology and mammography feature representations by combining cross-attention with channel shuffle operations.

Given a histopathology feature vector Fhis and a mammography feature vector Fmam, their element-wise product is first computed as mth=fhis⊗fmam.

${{s}_{th}}=\text{SeLU}\left( {{W}_{th}}\left( {{a}_{th}}\otimes {{m}_{th}} \right)+{{b}_{th}} \right)$           (7)

The cross-modal correlation score is then obtained through a dense transformation with SeLU activation.

$S'=softmax\left( S \right)$             (8)

To achieve diverse feature interactions, the histopathology and mammography feature representations are updated using the normalized cross-modal affinity matrix along with channel shuffle options.

$\begin{gathered}\text { Fhis } *=\text { Shuffle }(\text { Fhis } \cdot \text { S }) \\ \text { Fmam } *=\text { Shuffle }\left(\text { Fmam } \cdot S^{\mathrm{T}}\right)\end{gathered}$         (9)

Table 2 ablation results demonstrate the effectiveness of the proposed Cross-Shuffled Feature Fusion (CSFN) module. Early fusion achieved an accuracy of 96.84%, while late fusion improved the accuracy to 97.46%. Attention-based fusion further increased the accuracy to 98.31%. In comparison, the proposed CSFN achieved the highest accuracy of 99.69%, together with a specificity of 97.95%, precision of 98.67%, recall of 96.20%, and F1-score of 97.93%. These results indicate that the Cross-Shuffled Fusion strategy effectively preserves complementary information between mammographic and histopathological modalities while reducing redundant representations and improving inter-modality feature interaction. Bi-GRU with SAE for Classification.

Table 2. Ablation study of different fusion strategies in the proposed Cross-Shuffled Fusion based Multi-Modal Hybrid Deep Learning Network (CSFHD-Net)

During testing, images are processed through Bi-GRU networks with SAE, which refine feature representations and reduce dimensionality. These SAE branches are concatenated and classified normal and abnormal tissue followed by sub-classification of abnormal cases is performed to differentiate benign or malignant. The Bi-GRU is employed to effectively model sequential dependencies in the extracted multimodal features. Bi-GRU processes information both forward and backward, allowing it to record dependencies from past and future sequences. This bidirectional learning is essential for identifying complex spatial-temporal patterns in histopathological and mammography images.

The number of hidden units in the GRU layers and scaled dot-product attention are two other attention techniques that have been modified in addition to the number of GRU layers. To balance computational cost and recognition accuracy, a single Bi-GRU layer with a scaled dot-product attention mechanism was employed. These decisions on the Bi-GRU and attention components are crucial for the approach to properly identify BrC classification. It highlights significant frames or moments, eliminates superfluous information, and highlights vital aspects of the BrC by delicately assigning differential weights.

Formulas for the backward-pass Bi-GRU:

$m_{r}^{'}t=\sigma \left( W_{m}^{'}*\left[ h_{\left( t-1 \right),}^{'}{x}'\left( t \right) \right] \right)$            (10)

$n_{z}^{'}t=\sigma \left( W_{z}^{'}*\left[ h_{\left( t-1 \right),}^{'}{x}'\left( t \right) \right] \right)$           (11)

$\hbar _{t}^{'}=ReLU\left( W_{h}^{'}*\left[ m_{r}^{'}t*h_{\left( t-1 \right),}^{'}{x}'\left( t \right) \right] \right)$         (12)

$h_{t}^{'}=\left( 1-n_{z}^{'}t \right)*h_{\left( t-1 \right)}^{'}+n_{zt}^{'}t*\hbar _{t}^{'}$         (13)

Here, m_rt and $m_{r}^{'}t$ represent the reset gate. n_zt and $n_{z}^{'}t$ denote the update gate where both update gates. $\hbar _{t}^{'}$ denotes the candidate’s hidden state. h_t and $h_{t}^{'}$ denote hidden state. The model can learn complex behaviors from images that usually contain spatial and temporal information because to this mix of characteristics. Furthermore, adding a nonlinearity aspect to the model by substituting Tanh activation for ReLU activation makes it easier for deep networks like Bi-GRU to converge. Finally, the models are improved by the attention mechanisms.

3.3.3 Stacked Autoencoder

The Stacked Autoencoder (SAE) is integrated into the framework to provide robust feature compression and hierarchical representation learning. By successively encoding and decoding the feature space, SAE is able to extract increasingly abstract representations of the input data. Classification with this technique is particularly beneficial for intrinsic datasets because it can learn hierarchical representations of data. By stacking layers of auto encoders, the model can capture increasingly features and reduce dimensionality. Additionally, SAE were effectively high-dimensional data and are noise-resistant, which makes them appropriate for use in medical imaging applications. As a result, classification tasks are more generalized and accurate. Figure 4 shows a Bi-GRU with SAE for Classification.

tu_pian_17.jpg

Figure 4. Architecture of Bidirectional Gated Recurrent Unit (Bi-GRU) with Stacked Autoencoder (SAE)

The encoder and decoder layers are proportioned and have equivalent hidden neuron counts. The input images are transformed into intricate coding by the encoding layers, which capture multiple high-level representations of multimodal medical images. The decoding layers then reconstruct output signals that closely resemble the original inputs by utilizing these detailed coding. Experimental feature vectors are encoded and decoded using non-linear functions as follows,

${{h}_{k}}={{f}_{k}}\left( x \right)={{\phi }_{k,e}}\left( {{W}_{k,e}}{{x}_{k,i}}+{{\beta }_{k,e}} \right)$     (14)

${{\overset{}{\mathop{h}}\,}_{l}}={{g}_{l}}\left( x \right)={{\phi }_{1,d}}\left( {{W}_{l,df}}\left( {{x}_{l,i}} \right)+{{\beta }_{l,d}} \right)$          (15)

where, the functions of encoding and decoding in the k-th hidden layer (HL) ($h k$), and l-th HL (${{\overset{}{\mathop{h}}\,}_{k}}$) are indicated by ${{f}_{k}}\left( x \right)$ and ${{g}_{l}}\left( x \right)$, respectively. Weighted matrices in ${{h}_{k}}$ and ${{\overset{}{\mathop{h}}\,}_{l}}$ are indicated by the ${{W}_{k,e}}$ and ${{W}_{l,df}}$, and modified by tasks such as a hyperbolical tangent, linear, Maxout unit, sigmoid, step, ramp, and ReLU. In order to achieve an experienced network of SAE using gradient background approaches based on the error results during fine-tuning. The Bi-GRU with SAE modules combine to contextualize multimodal features sequentially and compress them into robust hierarchical embeddings. Combining these two networks provides highly discriminative feature vectors to reliably classify BrC as normal, benign, and malignant.

This research proposed a novel CSFHD-Net for early classification BrC stages using mammogram and histopathology images. Shuffled-MobileNet provides a lightweight and efficient method for training by learning discriminative features from both histopathology and mammography images. The Cross-Shuffled Fusion mechanism integrates these features to preserve inter-modality correlations and generate a unified fused vector representation. A GRU combined with SAE refines feature representation and reduces dimensionality during testing. The SAE branches are concatenated and classified as normal or abnormal tissue, followed by sub-classification of abnormal cases for differentiation of benign from malignant. From the experimental analysis, the experimental results provided in Section 4 show that the proposed CSFHD-Net has better classification performance than the current methods, whereas the comparison study proves its effectiveness in multimodal BrC classification. Future work includes extending CSFHD-Net to integrate additional imaging modalities, such as ultrasound and MRI, to enhance early BrC detection. Incorporating explainable AI techniques to provide interpretable and transparent decision-making, improving clinical applicability across diverse patient populations.