Morphology-Enhanced Transformer Attention Network for Accurate Pancreatic Cancer Detection from Computed Tomography Images

The diagnosis of pancreatic cancer has been one of the most demanding activities in medical imaging because of its very different morphologies, low lesion contrasts [1], and irregular lines that cannot be easily distinguished between lesions and normal surrounding soft tissues [2]. A common method of diagnosis has always been computed tomography (CT), as it allows visualizing local anatomical features clearly, but the traditional method of segmentation and classification often does not support finer anatomical features [3], resulting in the irregular delineation of tumors. Previous related research indicates multi-stage segmentation schemes and supervised networks to enhance the area of pancreatic identification [4], i.e., hyper-pairing networks, DeepOrgan-based pancreatic segmentation, and the holistically nested aggregation schemes [5]. Nonetheless, such classical Convolutional Neural Network (CNN) based systems lack the capacity to sustain structural continuity, particularly when morphological correlates differ greatly among tumor stages or where morphological information is misrepresented by noise and artifacts [6].

Recent progress in deep medical image segmentation has supported the significance of multi-scale feature extraction and enhanced skip-connection processes [7]. UNet++ illustrates the importance of dense multi-resolution pathways to learn heterogeneous lesion textures, and dense V-networks have proven to be quite competitive in the task of pulmonary segmentation and abdominal segmentation since they can learn progressively [8]. Much larger ones like DeepLesion further demonstrate the necessity of an effective, automated approach to feature extraction in the case of high-variability CT datasets [9]. However, even considering these efforts, pancreatic tumors are particularly hard to identify because of their perceptive morphological markers and extremely localized deviations [10], which explains the necessity to transform imaging in a more profound way that would empower structure-specific attributes [11].

Simultaneously, unsupervised and semi-supervised learning methods have become much more popular in medical imaging with the existence of a few large-scale datasets with annotations [12]. The multi-planar co-training, clustered instance learning, weakly supervised multi-resolution localization [13], and deeply supervised 3D architectures have all demonstrated that the ability to learn on heterogeneous sources of data can contribute a great deal to segmentation in limited annotation scenarios [14]. The original research on weak supervision, graph-based semi-supervised learning, and inductive methods served as inspiration for the current process in medical imaging by allowing better extrapolation of thin labels [15]. The techniques are especially applicable to pancreatic imaging, whereby annotation of the tumor delimitations is time-consuming, biased, and susceptible to inter-observer variation [16].

The recent developments of transformers and attention-based networks have helped medical image analysis to be pushed to a new level in that they permit long-range dependency modelling and consideration of global context [17], which traditional CNNs are necessarily deprived of. Guided feature attention closely supervised and partially unsupervised segmentation models [18], and semi-supervised consistency-based models are shown to greatly improve feature alignment and noise sensitivity [19]. Based on these developments, the current paper presents the Morphology-Enhanced Imaging and Transformer Attention (MEITA) framework to perform morphology-enhanced image preprocessing and transformer-based inductive attention to extract the most useful features in pancreatic cancer diagnostics. The proposed model exploits the use of morphological enhancement to highlight structural boundaries and transformer attention to infer global contextual relations, which overcome the shortcomings of previous CNN and semi-supervised models [20]. The combined framework aims to generate very discriminative representations for precise pancreatic tumor detection and segmentation, which will improve the reliability and clinical applicability of CT-based diagnostic systems. Figure 1 shows the general process of pancreatic tumor detection.

Recent progress in semi-supervised and weakly-supervised segmentation algorithms has offered substantial gains in cases where there are scarce annotated medical data [21]. The use of unlabeled data using robust augmentation-based learning has been shown to be effective in techniques that impose feature invariance to spatial transformations [22]. Prior-aware neural networks have also enhanced the reliability of segmentation in partially supervised settings by incorporating organ-specific priors into the learning process, and weakly-supervised thoracic disease localization and abdominal multi-organ segmentation models have indicated the increasing popularity of minimal-annotation learning models [23].

1.png

Figure 1. General process of pancreatic cancer detection

The classical concepts of adaptive pattern recognition are still used to guide the contemporary algorithmic behavior, with label uncertainty and multi-label weak-semi-supervised learning systems having demonstrated good ability in image-level classification and localization [24]. Various-instance learning also continues to play a central role in working with heterogeneous histopathology and cancer data because it can utilize pattern-level significance based on bag-level labels [25]. The representation power of deep CNN models like Visual Geometry Group (VGG) and U-Net-based derivatives has also established foundations in the history of inductive semi-supervised learning solutions and has influenced more recent developments in the medical imaging process [26]. The combination of these advancements creates a solid justification as to why morphological improvement is coupled with transformer-induced attention processes, and it is on this basis that the proposed MEITA framework will be able to exploit both local anatomical specifics and global contextual linkages to achieve better pancreatic cancer detection [27].

The proposed MEITA framework simultaneously incorporates the complementary representations by employing a unified diagnostic pipeline, as opposed to the conventional CNN-based methods that mainly extract local texture information and the Transformer-based models that mainly focus on global contextual relationships. The novelty of MEITA is manifold: (i) morphology-enhanced preprocessing to ensure preservation of the anatomical boundaries and suppression of imaging artifacts; (ii) a hybrid CNN-Transformer architecture that accounts for the joint learning of local structural patterns, and the long-range, contextual dependency; and (iii) Feature Dependency Check and Ascendancy Linking mechanism to remove redundant features while retaining the highly discriminative representations; and (iv) extensive experimental testing which showed that this model consistently outperformed state-of-the-art baseline models in terms of classification, segmentation, and attention localization. The innovations are all combined to enhance the robustness of the diagnosis, to be explainable to the clinician, and to be generalizable in different pancreatic CT images.

1.1 Hypothesis

The research hypothesis includes

1. Boundary in pancreas is greatly enhanced with morphology-enhanced pre-processing. Visibility, which allows more precise feature extraction of low-contrast CT images.
2. Pancreas Region of Interest (ROI) segmentation eliminates redundant background patterns, causing an increase in the concentration of discriminative features and improved CNN learning.
3. The fine-grained morphological texture that is necessary in distinguishing between early-stage Pancreatic Ductal Adenocarcinoma (PDAC) and normal pancreatic tissue is obtained through CNN-based local feature extraction.
4. Transformer attention refinement improves global contextual thinking that enables the model to detect long-range structural relationships that are impossible to detect using CNNs.
5. The integrated MEITA process gives a more reliable prediction on heterogeneous scans of CT and is more robust to changes in imaging quality, noise, and acquisition parameters.

1.2 Research contributions

The research objectives are

1. An improved pancreatic region visualization scheme is proposed, boosting the quality of morphology of the pancreas, leading to a better visualization of tumor-related features in the CT scan. The reliability of downstream feature extraction is improved exclusively by this improvement, as the fine-grained organ boundaries are preserved, and noise, low contrast, and structural distortion are addressed.
2. An attention-based hybrid CNN Transformer architecture is designed that will extract features best, in which CNN layers extract the local morphological textures and Transformer attention modules model the long-range contextual relationships. The result of this intertwined mechanism is a discriminative and clinically significant feature representation of PDAC.
3. A strategy to refine features guided by attention is suggested, which allows the automatic retraining of important diagnostic areas in the pancreas. This facilitates an achievable balance between sensitivity and specificity by decreasing irrelevant responses and reducing false-positive forecasts in problematic CT imaging.
4. MEITA is designed to offer strong classification with heterogeneous CT scans, exhibiting excellent variability to changes in imaging quality and imaging acquisition protocols. The model improves diagnostic accuracy, offers better interpretability by using attention heatmaps, and detects early-stage pancreatic cancer with high reliability.

The suggested approach is called MEITA, and it involves Morphology Enhanced preprocessing along with a Hybrid CNN-Transformer architecture that can help improve the performance of the identification of pancreatic cancer on CT images. Morphology Enhancement ensures that there is no loss of anatomical borders and that the imaging artifacts have been reduced. In this case, CNN layers represent local textural features while the Transformer identifies the context relations. The Feature Dependency Check is responsible for removing any redundancy from the feature representation, while Ascendancy Linking gives preference to the most discriminating features. CNN layers are specifically efficient in localized spatial patterns, and hence they can be utilized to identify small or early-stage tumors. The feature extractor consists of several convolutional blocks with batch normalization, followed by ReLU activation, which increases the stability of features and guarantees good gradient flow. This component is the first stage of representation for the further attention mechanism.

Local characteristics alone, however, do not provide enough PDAC diagnostics, as the anatomical environment is complicated, and the tumor has the propensity to mix with the tissue around it. To solve this, the proposed structure incorporates a Transformer Attention Module, which is used to model global contextual relationships between the features extracted. Transformer attention reweights tokens that are reweighted according to their global importance, enabling the model to capture long-range dependencies, which CNNs lack by nature. This will help to make sure that, in addition to local abnormalities, global structural cues are also involved in the diagnostic decision, thus enhancing interpretability and reducing false positives. The morphology-enhanced pre-processing, CNN-local extraction, and transformer global refinement have complementary advantages that yield a very informative feature vector.

The resulting aggregated feature vector is then passed through a fully connected classification head, which does the final PDAC prediction. Depending on the setup of classification, a softmax or a sigmoid output layer is adopted, and the cross-entropy loss is used during training to optimize the weights in the model. The model creates attention activation maps to make high-importance regions visibly prominent for better clinical interpretability. The radiologists can use these maps to confirm that the predictive decision is using a clinically relevant anatomical location, which makes the system more reliable in the diagnostic process. Adaptive learning strategies such as learning-rate warm-up, weight regularization, and data augmentation are used to optimize the MEITA framework. These mechanisms will keep the network stable and eliminate overfitting, as well as be robust when applied to heterogeneous CT datasets with varying acquisition parameters. The proposed model combines morphology amplification, sophisticated attention processing, and feature extraction (local and global) to create a strong pipeline for early and accurate pancreatic cancer detection to deal with the fundamental shortcomings of the current diagnostic tools. The proposed model architecture is shown in Figure 2.

2.png

Figure 2. Proposed model architecture

3.1 Dataset description

The proposed research makes use of a publicly available set of CT images, which is obtained at Kaggle available at https://www.kaggle.com/datasets/jayaprakashpondy/pancreatic-ct-images, and is comprised of 1411 pancreatic scan slices divided into the tumor and non-tumor groups. The dataset used in the proposed study is publicly available images of the pancreas from CT scans of the abdomen, repurposed from Kaggle, consisting of 1411 slices of the axial plane of the pancreas from CT scans of the abdomen obtained from contrast-enhanced CT imaging. Images were saved in PNG format with an original spatial resolution of 512 × 512 pixels, and all slices were resized to 224 × 224 pixels prior to network training. The Kaggle repository contains images from several clinical institutions, so detailed information about the scanner manufacturer is not available, but the data in the repository are heterogeneous scans extracted from multidetector CT scanners on the basis of imaging protocols for the portal venous phase. There are around 54% of tumor slices (762 images) and 46% of normal pancreas slices (649 images) in the dataset. All images were rescaled in the range [0,1] in the pre-processing stage and then processed by adaptive contrast enhancement and ROI extraction using morphology operations to reduce scanner-dependent intensity variations. The dataset was randomly split into three categories: an 80% training set, a 10% validation set, and a 10% testing set, with a fixed class balance across all three sets.

In addition to its balanced structure, the dataset is densely diversified with morphological structures, with subtle lesions, irregular tumor margins, and low-contrast pancreatic regions that may often be problematic to deep learning models. The proposed Morphology-Enhanced Transformer Attention framework is able to learn the fine-grained structural cues and promote boundary-preserving feature extraction due to this variability.

The proposed MEITA framework is a six-step method to extract discriminative features from pancreatic CT images effectively: The first step in the processing of CT images is pre-processing such as noise reduction, intensity normalization, adaptive contrast enhancement, and morphology-based extraction of regions-of-interest, which helps to enhance the visibility of the boundaries. The Transformer Attention Module is then used to model long-range spatial dependency and global contextual relationships, which are difficult to model using only conventional CNNs. An optimized feature vector is classified with fully connected layers with Softmax activation between tumor and non-tumor CT images. This streamlined pipeline boosts the representation of features, reduces computational redundancies, and increases the robustness of classification.

3.2 Pre-processing

The pre-processing phase is used to make sure that CT images are uniform, clear, and structurally enhanced, prior to the application of morphological operations and a Transformer-based feature extractor.

3.2.1 Noise reduction using Gaussian smoothing

Gaussian filtering reduces high-frequency noise while preserving coarse anatomical boundaries. The filter kernel $G(u, v)$ ensures smooth transitions, improving the clarity of pancreatic tissue patterns.

$I_G(x, y)=\sum_{u=-k}^k \sum_{v=-k}^k I(x-u, y-v) * G(u, v)$                (1)

3.2.2 Intensity normalization

$I_{{norm }}(x, y)=\frac{I_G(x, y)-\min \left(I_G\right)}{\max \left(I_G\right)-\min \left(I_G\right)}$               (2)

This min-max scaling makes the pixel intensities vary within a uniform range [0,1]. It minimizes variations in the difference that may arise due to different CT acquisition parameters and maintains a constant gradient flow of the deep model.

3.2.3 Adaptive Histogram Equalization

$I_{{enh }}(x, y)=I_{ {norm }}(x, y)+\eta \cdot\left(H_{ {local }}(x, y)-H_{ {global }}\right)$               (3)

The enhancement of local contrast is achieved through AHE which enhances subtle intensity differences in the pancreas. η regulates enhancement to avoid over-amplification in homogeneous regions.

3.2.4 Adaptive threshold-based Region of Interest segmentation

$S(x, y)=\left\{\begin{array}{lc} 1, & I_{\text {enh }}(x, y) \geq T_{\text {adaptive }} \\ 0, & \text { otherwise } \end{array}\right.$                  (4)

The given step of segmentation identifies the pancreatic region with the help of the dynamic threshold calculated on the basis of the local intensity statistics.

3.2.5 Morphological refinement (closing operation)

$S_{\text {morph }}=(S \oplus B) \ominus B$              (5)

where, B denotes the structuring element. The closing operation fills small gaps and smooths region borders, ensuring cleaner pancreatic masks for subsequent feature extraction.

3.2.6 Gradient-based edge stabilization

$E(x, y)=\sqrt{\left(\frac{\partial I_{e n h}}{\partial x}\right)^2+\left(\frac{\partial I_{e n h}}{\partial y}\right)^2}$              (6)

The improved representation of edges helps in more precise morphology-aware Transformer attention.

3.3 Input cloud service datasets and pre-processed datasets

Prior to feature extraction, all the images are subjected to an integrated cloud-based pre-processing pipeline that aims at standardizing the format, dealing with inconsistent resolutions, and eliminating unnecessary metadata that is usually added during cloud uploads. To ensure uniformity, images are first converted to a standardized resolution and intensity scale. Let an input CT image from the cloud dataset be denoted as:

$I_c=\left\{I_1, I_2, \ldots, I_N\right\}$             (7)

where, each $I_i$ represents a cloud-retrieved slice. A normalization map $\phi$ is applied to resolve heterogeneous acquisition settings, expressed as:

$I_{s t d}(x, y)=\phi\left(I_i(x, y)\right)=\frac{I_i(x, y)-\mu_c}{\sigma_c}$              (8)

where, $\mu_c$ and $\sigma_c$ denote dataset-wide mean and standard deviation. This measure is necessary to assure statistical consistency of scans of various clinical sources.

Every picture is then scaled with an interpolation operation distributed on the cloud:

$I_{{res }}(x, y)=I_{ {std}}(a x, b y)$               (9)

where, scaling factors a and b control height–width adjustments. This standardization makes batch processing possible in the Transformer encoder, which has fixed spatial dimensions.

Cloud preprocessing also eliminates artifacts that are created during compression or cloud transfer of images. A structural stabilizing denoising kernel D(x,y) is used:

$I_{\text {clean }}(x, y)=I_{\text {res }}(x, y) * D(x, y)$                 (10)

where, * denotes convolution. This improves the pancreas-tumor boundaries, setting the dataset to be used in morphology-sensitive attention extraction.

Finally, an intensity clipping operation eliminates extreme outlier values:

$I_{\text {clip }}(x, y)=\min \left(\max \left(I_{\text {clean }}(x, y), L\right), U\right)$               (11)

where, L and U represent lower and upper HU thresholds. This step suppresses cloud-induced artifacts and improves contrast stability across all CT slices. Algorithm 1 briefly explains the pre-processing procedure.

Access and download each CT slice $I_i$ from the Kaggle cloud service.

Store all retrieved images in a temporary buffer.

•Standardize intensity distribution: For each image $I_i$, apply global normalization:

$I_{s t d}(x, y)=\frac{I_i(x, y)-\mu_c}{\sigma_c}$              (12)

•Resize image to standard dimensions: Apply interpolation-based resizing:

$I_{r e s}(x, y)=I_{s t d}(a x, b y)$               (13)

•Apply noise reduction filter: Perform convolution with a denoising kernel:

$I_{\text {clean }}(x, y)=I_{\text {res }}(x, y) * D(x, y)$                   (14)

•Intensity clipping for artifact reduction: Remove extreme HU values using Eq. (11).

•Store pre-processed image: Add each processed slice $I_{ {clip }}$ to the final dataset:

$D_{ {prep }}=D_{ {prep }} \cup\left\{I_{ {clip}}\right\}$                (15)

Output the complete pre-processed CT dataset $D_{{prep}}$ for subsequent morphology-based and Transformer-based feature extraction.

3.4 Feature Dependency Check and Ascendancy Linking

Once morphology-enhanced and Transformer-based feature maps have been generated, the model uses a two-step refinement process, which includes: Feature Dependency Check and Ascendancy Linking to only keep the most discriminative, non-redundant, and class-separating features. These unnecessary aspects retard calculation and reduce generalization. Thus, the dependency analysis procedure makes sure that the statistically independent and clinically meaningful features are left.

3.4.1 Feature dependency computation

The correlation analysis evaluates the similarity between features using a normalized linear dependence measure:

$\rho_{i j}=\frac{{cov}\left(f_i, f_j\right)}{\sigma_{f_i} \sigma_{f_j}}$                (16)

where, covariance is computed as:

${cov}\left(f_i, f_j\right)=\frac{1}{N} \sum_{k=1}^N\left(f_i^{(k)}-\mu_{f_i}\right)\left(f_j^{(k)}-\mu_{f_j}\right)$              (17)

A dependency score is also computed via mutual information, enabling detection of nonlinear redundancy:

$M I\left(f_i, f_j\right)=\sum_u \sum_v p(u, v) \log \left(\frac{p(u, v)}{p(u) p(v)}\right)$                 (18)

To combine linear and nonlinear dependency:

$D_{i j}=\beta\left|\rho_{i j}\right|+(1-\beta) \cdot \frac{{MI}\left(f_i, f_j\right)}{M I_{max }}$               (19)

A feature pair is considered redundant if:

$D_{i j} \geq \tau_{d e p}$               (20)

where, $\tau_{{dep}}$ is the dependency threshold.

3.4.2 Discriminative ascendancy score

For each feature remaining after dependency filtering, the model computes discriminative capability. The Fisher-based ascendancy score is computed as:

$\alpha_i=\frac{\left(\mu_{i, 1}-\mu_{i, 2}\right)^2}{\sigma_{i, 1}^2+\sigma_{i, 2}^2}$             (21)

To incorporate morphological importance, a morphology relevance factor (MRF) is added:

$M R F_i=\frac{\left|\nabla f_i\right|}{1+\text {texture_smoothness}\left(f_i\right)}$              (22)

The final ascendancy score becomes:

$A_i=\gamma \alpha_i+(1-\gamma) M R F_i$                (23)

A feature is retained if:

$A_i \geq \tau_{ {asc }}$               (24)

3.4.3 Entropy-based stability evaluation

To ensure selected features are consistent across patient samples, an entropy score is computed:

$H\left(f_i\right)=-\sum_{k=1}^K p_k \log \left(p_k\right)$               (25)

The normalized stability score:

$S_i=1-\frac{H\left(f_i\right)}{\log K}$             (26)

Final weighted discriminative measure:

${ Score }_i=A_i \cdot S_i$               (27)

Feature selection criterion:

$Score_i \geq \theta$             (28)

where, θ is a stability-aware threshold.

The CNN layers extract local features, and the Transformer Attention Module creates a global context representation by calculating the relationships between spatial feature tokens to improve discriminative representation. The Feature Dependency Check module, which is then applied to the feature maps built in the previous step, extracts statistically redundant features from the feature maps, using the basis of correlation and mutual-information analysis. The rest are ranked according to the Ascendancy Linking strategy, which is prior to classification, based on discriminative power, morphological relevance, and feature stability. Sequential optimization increases computational efficiency, while maintaining clinically relevant anatomy data. The pseudocode for performing the pre-processing is also included. Algorithm 2 presents the Feature Dependency Check and ascendancy linking process.

The Feature Dependency Check step is used to ensure that only non-redundant and statistically independent features are forwarded to classification. Because of high-density and high-dimensional feature maps created by morphology-enhanced filters and Transformer attention blocks, a significant portion of extracted attributes have linear or nonlinear redundancy. To remedy this, a hybrid measure of pairwise dependency is calculated, which is a combination of Pearson correlation and mutual information, which enables the model to model both simple and complex dependencies between features. The characteristics that are greater than the dependency level are filtered out to eliminate overfitting and speed up calculation. In addition to making the feature representation compact, this refinement procedure permits the ensuing classification network to act on structurally different as well as clinically significant attributes.

The Feature Dependency Check mainly filters redundant feature representations produced by the CNN and Transformer modules based on the analysis of linear correlation and nonlinear mutual information. Only statistically independent features are chosen to be further processed. The Ascendancy Linking module then ranks the other features by their discriminating power and morphological importance prior to classification. This refinement process has two stages to eliminate the computational redundancy while retaining the clinically relevant anatomical information, which brings higher accuracy of classification and better generalization for the model.

After the eliminated features are gone, the Ascendancy Linking stage measures the discriminative power of the remaining features by approximating their separability between healthy and cancerous pancreatic areas. An ascendancy score using Fisher is calculated for each feature, and it is combined with a morphology relevance score based on the magnitude of the gradients and amount of smoothness in the textures, which makes sure that shape-, boundary-, and structure-preserving features are given more weight. In order to ensure consistency between samples of patients, an entropy-based score on stability is used, and the features that satisfy the ascendancy-stability score are kept.

Although there are indications of good diagnostic performance, some limitations need to be noted. First, the images were taken from a publicly available repository of images, but the data set does not contain detailed scanner-specific metadata, which would allow for detailed analysis of scanner-specific variability. Second, this proposed framework has never been tested on a large-scale multi-center clinical study, so its validity in other hospitals, patient populations, and imaging protocols needs to be explored. Third, the model performance may be affected by domain shifts stemming from various acquisition parameters, such as slice thickness, reconstruction kernels, contrast administration protocol, and scanner manufacturers. Moreover, the current framework simply treats individual two-dimensional CT slices independently and does not explicitly make use of the three-dimensional volumetric information. Last but not least, prospective external clinical validation, with multiple institutions and radiologists, is required before clinical routine use.

Future studies will be conducted on developing lighter Transformer architectures to further decrease computational complexity and enable deployment on resource-limited clinical systems. Large-scale multi-center datasets collected from various hospitals, scanner manufacturers, and imaging protocols will be used to validate the proposed framework and assess its generalizability. In addition, volumetric learning will be explored, as it will enable using the spatial continuity between adjacent CT slices to enhance the localization accuracy of the tumor. In addition, future versions of MEITA will incorporate multimodal clinical details such as MRI, Positron Emission Tomography (PET), genomic biomarkers, and eHealth records data to boost diagnostic confidence and customized threat determination. Additionally, explainable artificial intelligence techniques like Grad-CAM and attention visualization will be integrated to improve model transparency and aid clinical acceptance by supporting the radiologist in interpretation.