Rani Saritha, R.

et al: Enhanced transformer-based…… Archives for Technical Sciences 2024, 31(2), 175-191

ISSN 1840-4855
e-ISSN 2233-0046

Original scientific article

[Link]

ENHANCED TRANSFORMER-BASED DEEP KERNEL

FUSED SELF ATTENTION MODEL FOR LUNG NODULE
SEGMENTATION AND CLASSIFICATION
R. Rani Saritha1*, Dr.R. Gunasundari2
1*
Research Scholar, Department of Computer Science, Karpagam Academy of Higher
Education, Coimbatore, Tamil Nadu, India. e-mail: ranisaritha3090@[Link],
orcid: [Link]
2
Associate Professor, Department of Computer Science, Karpagam Academy of Higher
Education, Coimbatore, Tamil Nadu, India. e-mail: gunasundari.r@[Link],
orcid: [Link]

SUMMARY
Lung cancer remains a leading cause of mortality globally, with outcomes heavily dependent on the
timeliness and accuracy of diagnosis. Traditional medical imaging techniques, while foundational in
detecting lung nodules, often falter in distinguishing malignant from benign lesions with high precision,
largely due to their inability to contextualize the complex spatial relationships within the images. Precise
segmentation and classification of lung nodules is crucial for the early detection of lung cancer. This
paper presents a novel deep learning model that incorporates a transformer block to improve the
performance of lung cancer detection and classification. From performance evaluation, it is evident that
our proposed model has an average accuracy of 93%. 05%, which is superior to the existing D3DR _
MKCA model with a mean accuracy of 91.53%. These findings are especially important for the
identification of Adenocarcinoma and Small Cell Carcinoma, as improvements in the precision and recall
factors have been achieved in these cases.

Index Terms: lung cancer detection, deep learning models, transformer block, lung nodule
segmentation, adenocarcinoma classification, small cell carcinoma.

Received: July 27, 2024; Revised: September 01, 2024; Accepted: September 20, 2024; Published: October 30, 2024

INTRODUCTION

Lung cancer is on the list of top cancer-related deaths globally, and its high mortality rate is associated
with the advanced stage at diagnosis [1]. It is crucial to detect nodules early and classify them accurately,
as this increases the survival chance by allowing timely intervention and proper treatment planning [3].
lung nodules are usually small masses of tissue in the lung that are sometimes benign but at times
malignant. Lung tumors remain diagnostic challenges due to their complexity and variability and
therefore, there is need for improved medical imaging techniques to differentiate between malignant and
benign lesions [2, 8].

The conventional image segmentation techniques comprising of thresholding, region segmentation, and
edge segmentation have been the key fundamental elements in medical image analysis that mainly
addressed simple attributes of the images such as grayscale and texture [28]. However, these methods
may face issues such as low contrast, noise and the shapes of tumors which are difficult to define in lung

Technical Institute Bijeljina, Archives for Technical Sciences. Year XVI – N0 31 175
Rani Saritha, R. et al: Enhanced transformer-based…… Archives for Technical Sciences 2024, 31(2), 175-191
imaging [5]. Although some of these techniques such as standard U-Nets have enhanced results utilizing
deep learning, they are not effective in exploiting context information comprehensively and usually have
a high possibility of false positive results that might result in biopsy and increased patient stress [6, 7].
To overcome these limitations various organizations and recent works in the field of medical imaging
and CAD systems have started exploring the concepts and algorithms included in machine learning and
deep learning. These methods have the possibility to provide ideas for the automation of lung nodule
detection and classification to minimize the radiologists’ burden and increase accuracy. Therefore, deep
learning models, especially, have revealed substantial potentials which are even higher than enabling
traditional CAD systems, as they can identify intricate patterns from large datasets.

Motivation and Problem Statement

Medical imaging has experienced significant progress in the last few decades, which has led to the ability
to diagnose such large diseases like lung cancer with the need for additional high accuracy of the source
of imaging data [4, 7]. However, the current technological advancement calls for development of better
diagnostic tools that can easily incorporate detailed image analysis with the clinical decisions made
regarding the same [29]. The main rationally for this research lies in the presumed possibilities of
enhancing the diagnostic accuracy and delivering the best treatment to the patient when the disease is in
its early stages [15]. Improving image segmentation technologies with the most advanced machine
learning methods may be one of the possibilities of implementing these clinical goals and develop new
approaches to cancer diagnosis [20].

The primary difficulty in lung cancer detection is the separation and identification of the lung nodules
in association with radiographic imaging which is compounded by non-homogeneous appearance of
tumors, overlapped tissues, and when it comes to distinguishing malignant and benign nodules [7, 9].
Current segmentation methods have its utility but they are not intelligent enough to incorporate the
spatial and textural information required for the accurate segmentation of these complexities [10]. This
limitation can result in either very high false positive and negative values, meaning the patient might
undergo surgeries that they do not need, or they might not be diagnosed at all. This is further
compounded by the inconsistency of the imaging techniques and qualities for which the developed
methods need to be robust and flexible enough that they are able to generalize the findings across
different imaging conditions and yet, maintain a high level of diagnostic precision.

Contribution of the Research Work

The aim of this paper is to overcome these difficulties through proposals with the usage of deep learning
model, where transformer blocks are adaptive to contextual processing due to their structure while self-
attention mechanisms are highly accurate in signal-localizing tasks. This unique approach is expected to
notably improve accuracy of separating and grouping lung nodules as well as equip radiologists with a
dependable tool to facilitate right diagnosis.

This paper proposes a new architecture called deep convolutional self-attention model with a transformer
block, a relatively new and unique idea in imaging analysis. The deep convolutional and transformer
blocks are capable of learning large-scale spatial dependencies between different regions in an image
for better contextual relationship learning in medical images. This will be useful in ascertaining how
distant patches interact in order to detect finest, less conspicuous trends that can point to malignancy.
The transformer block takes the input data layer by layer: It decomposes the input data into multiple
layers that enable the capture of high-level information such as the detail, edges, texture of a nodule or
more coarser cues like the shape or size of a nodule. The transformer’s mechanism of self-attention lets
it pay attention to only relevant parts of the image that contribute to the final decision. When extended
with attention mechanism, which aims at improving the segmentation accuracy by narrowing down the
area of interest, the proposed method will significantly enhance the sensitivity and specificity of the lung
cancer detection.

Organization of the Paper

This research article is organized as follows: Section 2 presents a literature review of related work,
setting the theoretical and empirical context of the technologies under consideration. Section 3 describes
Technical Institute Bijeljina, Archives for Technical Sciences. Year XVI – N0 31 176
Rani Saritha, R. et al: Enhanced transformer-based…… Archives for Technical Sciences 2024, 31(2), 175-191
the data sources, model design, and training strategies in general terms. Section 4 also measures the
effectiveness of the proposed model in terms of the existing metrics. Section V presents the conclusions
of the analysis, the advantages and drawbacks of the proposed methodology, and potential research
avenues. Lastly, section six synthesizes the results through a conclusion and illustrates the effect of the
findings on the area of medical diagnosis.

RELATED WORK

Zhi et al. [11] have also performed the latest and detailed survey on segmentation of pulmonary nodules
in CT imagery using deep neural network approach emphasizing the benefits of those medical models
against traditional segmentation. In the current study, heterogeneity of pulmonary nodules is analyzed
based on aspects like subtlety, sphericity, margins, texture, and size as regards the segmentation
performances. From the LIDC and Luna16 dataset they successfully tested the effectiveness of the open-
source 2D and 3D models and found that 2D models are more accurate. Instead, it found that features
such as ‘Subtlety’ and ‘Texture’ do have effects on the segmentation performance, and the design of
contextual information acquisition and attention mechanisms has positive effects on the results.

Shashikala et al. [12] proposed a novel approach for lung nodule detection named Cross- Spectral Vision
Transformer with an Improved Moth Flame Optimization (CSViT- IMFO). This approach helped
improve the classification accuracy on the LIDC dataset to 99.5%. This paper focuses on three-step
process that includes preprocessing, segmentation, and classification, with ALPM & RC Methods.
MATLAB was used for testing the proposed method and confirmed higher performance as compared to
the traditional learning-based methods.

Li et al. [13] offered a more detailed review of deep learning attention mechanisms for medical image
analysis. The study expounded on how the transformers replicate human cognitive systems in Attention
Mechanisms which boosts the performance of medical image tasks through selective attention on the
data. Some of the applications like classification, segmentation, detection, and enhancement were also
reviewed with reference to their technical definitions and possible difficulties. The paper stated that
attention mechanisms have significantly enhanced model performance in the medical imaging domain
but introduced new issues that need to be studied further to achieve excellent performance.

Shamshad et al. [14] presented a comprehensive review of the literature on transformer models in
medical imaging, where the authors discussed the recent developments and limitations. The paper also
focuses on the effectiveness of the transformers in capturing the global context, something that has been
noted as vital for handling medical images and not found in the traditional CNNs. It provides an
overview of several medical imaging tasks like segmentation, classification, detection that use
Transformers. The present shortcoming and future development of this area are also involved, it has
been pointed out that, although Transformers provide a better improvement, there are still many
unknown problems regarding Transformers and its application in clinical practices.

Chen et al. [30] the authors presented a comparison of several DL models in the context of lung nodule
segmentation in CT images with a focus on data preprocessing methods. The study also noted that the
preprocessing techniques such as adding of a lung mask and cropping the area of interest to enhance
segmentation impact on the results. TransUNet proved to be the most accurate out of all the models
analyzed with high Dice coefficients in various datasets. Using preprocessing the models greatly
improved the results and segmentation accuracy was also highlighted as an important factor by the
research conducted.

Annavarapu et al. [16] proposed a new deep learning-based segmentation technique for lung nodule
using Bi-FPN between encoder and decoder improving the speed of segmentation. As such, a model
adopting the Mish activation function, and class weight, was employed to analyze the LUNA 16 and
QIN Lung CT Dataset. It performed better than U-Net with Dice Similarity Coefficients of being 82.
82% and 81. 66%. Compared with previous models, this one successfully takes a step forward in dealing
with various and complicated lung nodule features.

Technical Institute Bijeljina, Archives for Technical Sciences. Year XVI – N0 31 177
Rani Saritha, R. et al: Enhanced transformer-based…… Archives for Technical Sciences 2024, 31(2), 175-191
Liu et al. [17] proposed an attention-based 3D CNN model for categorizing lung nodules as benign or
malignant. Thus, the model combined CT imaging features with the clinical data and yielded the average
of diagnostic sensitivity as 96.2% for malignant nodules and an accuracy of 81%. This performance
exceeded typical ResNet and VGG networks, which verified the effectiveness of the attention
mechanism in improving diagnostic accuracy in lung cancer evaluations.

Wu et al. [18] developed an enhanced version of the U-Net architecture, termed RAD-UNet, tailored for
lung nodule semantic segmentation. This model integrates a ResNet encoder and an atrous spatial
pyramid pooling module, with a cross-fusion feature modulethat incorporates channel and spatial
attention mechanisms. Tested on the LIDC and a local dataset, RAD-UNet outperformed traditional
SegNet and U-Net models, achieving mean intersection-over-union (mIoU) scores of 87.76% and
88.13%, and F1-scores of 93.56% and 93.72%, respectively, on these datasets. These results underscore
its superior ability to delineate lung nodules accurately.

Saihood [19] proposed another deep learning model based on Multi-Orientation Local Texture Features
(MOGAM) that classifies lung nodule. The model incorporates guided attention mechanisms to integrate
the textural features computed across different orientations and enhance the emerging classification
results on the LIDC-IDRI dataset. The proposed MOGAM is ideal for the spatial variation of lung
nodules and improves the distinction between benign and malignant cases. This approach improved not
only the accuracy and AUC but also offered an overall framework for analyzing the spatial distribution
of textural features of the lung nodules.

In a work by Zheng et al. [31], a multichannel attention-guided deep neural network model was proposed
for lung nodule classification from CT scans. To achieve accurate feature representation and fusion, the
model incorporates multi-view and multi-scale mechanisms. As a result, this particular study adopted
the LIDC-IDRI dataset and obtained a classification accuracy of 90.11% and an AUC of 95.66% By
using the attention mechanism, the process of localization of nodules made it possible to achieve high
classification results while decreasing the time needed for analysis of images. It outperformed other deep
learning models and revealed the enhanced performance and utility compared to the previous models.
of the local texture features or Multi-Oriented Gaussian-derivate based Angle Measurements (MOGAM)
to classify lung nodules. These guided attention mechanisms are employed to integrate the textural
features computed in various orientations and achieve enhanced classification performance on the LIDC-
IDRI dataset. As the result, the proposed MOGAM deals well with the spatial variation of the lung
nodules improving discrimination of benign and malignant cases. In addition to increasing the accuracy
and AUC, this approach offered a method for identifying the spatial distribution of textural features in
lung nodules.

Bhattacharyya et al. [21] proposed a bi-directional DL model named as DB-NET for lung nodule
segmentation. This model also introduced the bidirectional feature network and Mish activation function
to improve the segmentation accuracy. On the LUNA-16 dataset, DB-NET showed improvements over
the traditional U-NET model, with the Dice coefficient equal to 88.89%. In this account, several
incidences were highlighted that demonstrated how DB-NET can easily handle segmentations with
many segments and complex spatial details while incurring minimal computational costs.

Canayaz et al. [22] utilized deep learning models on pulmonary nodules and segmented and classified
them on the CT images. Introducing EffxNet for classification, the work proposed impressive
performance figure including accuracy of 97.98% as well as kappa value of 0.9690. As for the
segmentation, some of the used models include InceptionV3, DenseNet121 and SeResNet101; the best
Dice coefficient was 0.8877 sets attained following the implementation of FPN model with
DenseNet121. The paper pays a lot of attention to the authors showing how the deep learning models
can be implemented to enhance medical diagnosis of lung nodules.

Gai et al. [23] specifically focused on the comparison between CNN-based and transformer-based
models for diagnosing lung cancer from CT images. They employed sophisticated approaches such as
transfer learning, self-supervised learning, and sharpness-aware minimization for improving the model.
The findings revealed that CNNs especially the one trained utilizing self-supervised learning yielded

Technical Institute Bijeljina, Archives for Technical Sciences. Year XVI – N0 31 178
Rani Saritha, R. et al: Enhanced transformer-based…… Archives for Technical Sciences 2024, 31(2), 175-191
higher accuracy with a recall of 93.4% indicating good predictive accuracy and a higher AUC of 98.1%
while a ViT achieved lesser accuracy when trained only with less number of training samples. This work
highlights the superiority of CNNs against ViTs in medical image analysis when data is limited.

Ali et al. [24] performed a scoping review to assess recent advances in vision transformer-based AI
approaches for lung cancer imaging. In this review, I included results of 34 published studies and focused
on using vision transformers for classifying lung cancer and guiding the disease’s progression. Another
interesting model mentioned was the vision transformers which were said to improve diagnostics and
prognoses, especially in classification problems, and in terms of survival prediction. The review raises
the bar for vision transformers to promote future progress of lung cancer diagnosis and prognosis and it
also acknowledges that exploration of more effective and efficient approaches to training is required, as
well as discussing the significance of the subject’s translation to clinical practice.

Cao et al. [25] proposed a multi-scale nodule detection network for pulmonary nodules with attention
mechanism for improving its detection capabilities. The authors have incorporated a method of pseudo-
color processing to the data pre-processing in the proposed method, incorporating contextual data and
enhancing the textural characteristics of lung nodules. It employs the ResSCBlock for feature extraction
and the feature pyramid network for feature fusion across multiple scales. They have also demonstrated
an enhancement in the detection of pulmonary nodules on the LUNA16 dataset with the mAP value of
83% thus establishing a strong capacity to deal with different sizes and characteristics of nodules.

Yang et al. [26] trained a Natural Language Processing (NLP) system based on transformer models to
identify pulmonary nodules and related features from radiology reports. They evaluated the different
pretrained transformer models of BERT, RoBERTa, and ALBERT for clinical concept extraction,
relation identification, and negation detection. RoBERTa-mimic gave the best accuracy in concept
extraction test with F1–score of 0. 9279 while ALBERT-base performed best in relation identification
test with F1-score of 0. 9737. The system also successfully focused on the expected negation detection,
with multiple transformers scoring a perfect F1-score. This shows how the rich transformer models can
successfully capture clinical details by analyzing textual data.

DEHA-Net was proposed by Usman and Shin [27] for lung nodule segmentation, and it is a dual-
encoder-based hard attention network with A-ROI. Approach developed with this method, tested on the
LIDC/IDRI dataset, on average achieved Dice score of 87.84%, and positive predictive value of 89.56%.
DEHA-Net takes advantage of the full axial slice and feeds an ROI mask through a dual-encoder
structure, as a result focusing on the region of interest for improved segmentation performance. This
approach proved superior to traditional models in its ability to operate regardless of nodule size and
shape changes, without having to resize the ROI.

Research Gaps

In recent years, deep learning techniques have greatly enhanced the medical image analysis, specifically
comprising lung nodule detection and segmentation, but there is a significant lack of incorporating these
two technologies into a single, coherent system that will involve global context information and specific
image specifics. Extensive state-of-the-art works can perform better on either the thorough contextual
associations by utilizing the architectures like CNNs or on the fine-grained feature elaboration by
segmentation networks like U- Nets. Nonetheless, combining these strengths in a single architecture that
can interpret highly professional medical images while considering both macroaspect and microaspect
is not well understood. Existing approaches mainly pay attention to increasing the goodness of fit of
segmentation or of classification techniques but not of both at the same time. Furthermore, the ability of
transformer models, which were previously demonstrated in self-attention mechanism for discovering
long-range dependencies in data, has not been fully investigated in the medical image analysis especially
in the case of lung cancer where the spatial organization of tissues can greatly affect the diagnostic
results.

This research gap is in composing a model that integrates the global context understanding of the
transformer models with the fine-grained localization offered by Attention Gates. This integration is
aimed at handling some of the difficulties associated with medical imaging including differences in
Technical Institute Bijeljina, Archives for Technical Sciences. Year XVI – N0 31 179
Rani Saritha, R. et al: Enhanced transformer-based…… Archives for Technical Sciences 2024, 31(2), 175-191
nodule presentation and malign symptoms, to provide a significant improvement in the imaging of lung
cancer.

PROPOSED MODEL

Lung cancer continues to be one of the most prevalent types of cancer with high mortality rate. Lung
nodules that suggest malignancy need to be detected at an early stage and classified properly to facilitate
patient care. Among the various data deep learning approaches, Convolutional neural networks or CNNs
have been quite effective in medical image analysis. The model brings in CNN layers, self-attention,
and transformer blocks to enhance the efficacy and resilience of the proposed lung nodule classifier.
This framework should be capable of analyzing the internal structure of CT images and properly
differentiate between the four categories of malignant nodules.

Dataset Descriptions

The dataset for lung nodule cancer type classification, provided by Drs. Huiping Han, Funing Yang, and
Rui Wang, includes CT images and PET-CT DICOM images, accompanied by XML annotation files.
These images were retrospectively collected from patients with suspected lung cancer who had
undergone standard lung biopsy and PET/CT scans. Patients were classified based on their
histopathological diagnoses. The radiologists categorized the cases into four distinct groups based on
the tissue type. Group 'A' represents Adenocarcinoma, noted for mucus-like substances within the cells.
Group 'G' corresponds to Squamous Cell Carcinoma, typically located in the lung airways. Group 'B'
includes Small Cell Carcinoma, which is known for its rapid progression relative to other types. Finally,
group 'E' consists of Large Cell Carcinoma, which can appear in any area of the lungs.

According to the radiologists' markings, a total of 29,121 slices have been identified as cancerous. Group
A contains 20,894 of these slices, Group B has 3,116 slices, Group E includes 801 slices, and Group G
has 4,310 slices. From these, 7,894 slices from Group A, 3,116 slices from Group B, 1,005 slices from
Group E, and 4,310 slices from Group G were selected for further analysis. Since Group E is the smallest,
data augmentation techniques were applied to increase its size. Ultimately, a total of 16,325 slices were
chosen for further processing, with 80% allocated for training and 20% set aside for testing shows in
figure 1.

Figure 1. LIDC-IDRI dataset sample with original CT slices (left), ground truth (center) and U-Net predicted
segmentation mask (right)

Architectural Setup

Data pre-processing is a crucial stage in the construction of a lung nodule segmentation and classification
model as it prepares the data for the next phase. Some of the typical pre-processing steps that are
employed before analysis are

Technical Institute Bijeljina, Archives for Technical Sciences. Year XVI – N0 31 180
Rani Saritha, R. et al: Enhanced transformer-based…… Archives for Technical Sciences 2024, 31(2), 175-191
• Image Normalization: This step is employed in normalizing or scaling the intensity values of CT
images to a range of 0-1 or normalizing the intensity values based on the method of mean and
standard deviation. Following this it assists in reducing fluctuations of intensity due to varying
scanning conditions and hence provide a standardization in the data set.

• Resizing and Resampling: For the consistency in the data, the CT images are then normalised
and scaled to the depth of 128 x128 dimension. Standardization of spatial resolution by resampling
is important with regard to segmentation where the images are to be segmented.

• Noise Reduction: There is a lot of noise present in the images which can be minimized after
applying the noise reduction techniques such as Gaussian smoothing or median filtering. This
enhancement helps identify lung nodules more clearly and also decreases the probabilities of false
positive during the classification step.

The A-plan goes through several subsequent stages starting at first by segmenting lung nodules from the
CT images through an efficient Attention U-Gate semantic model for both, precise localization and
delineation of the nodules. This process helps to separate the nodule regions from the other thoracic
structures, which helps to limit further analysis to these zones. The segmented nodule is then subjected
to the proposed DKFSA-ET framework for the classification of the segmented nodule into four different
types of cancer tissue with the class label as A, B, E, G so as to enhance the overall lung cancer detection
accuracy and reliability of the proposed model through the utilization of the comprehensive
characteristics obtained from the different scale fused attention deep network model. This will make it
easier to accurately categorize and manage the disease based on the phases of the disease that have been
observed. Figure 2 shows the block diagram of the lung nodule classification.

Figure 2. Block diagram of lung nodule classification

The proposed architecture of DKFSA-ET known as the Enhanced Transformer based Deep Kernel Fused
Self Attention Model is depicted in the Figure 3. Its purpose is to sort and categorize into four groups
(A, B, E and G) the lung cancer. To improve classification task, this design combines transformer blocks,
self-attention techniques, and convolutional layers. Three primary convolution phases, each with unique
configurations, compose the process. A number of processes are then performed to guarantee accurate
categorization. The different stages of the model is explained below.

Technical Institute Bijeljina, Archives for Technical Sciences. Year XVI – N0 31 181
Rani Saritha, R. et al: Enhanced transformer-based…… Archives for Technical Sciences 2024, 31(2), 175-191

Figure 3. DKFSA-ET architecture for lung nodule classification

First Stage of Convolution (Conv_Stage-1)

The first stage of the model contains three sub-stages of different kernel and filter sizes.

• Sub-stage A: Six convolutional layers, among this first two (process the input) and 4th and 5th layers
each with a kernel size of 3 × 1 and a filter size of 16. Similarly, two convolutional layers with
filter size 32 and kernel size 3 × 1 are used after every two convolutional layers of 16 filter size.
That is illustrated in the following equations 1 to 2.

𝐂𝐨𝐧𝐯𝟏𝐀−𝐒𝐭𝐚𝐠𝐞𝟏 = 𝐂𝐨𝐧𝐯𝟑×𝟏,𝟑𝟐 (𝐂𝐨𝐧𝐯𝟑×𝟏,𝟏𝟔 (𝐂𝐨𝐧𝐯𝟑×𝟏,𝟏𝟔 (𝒊𝒏𝒑𝒖𝒕))) (𝟏)

𝐂𝐨𝐧𝐯𝟏𝐀𝐒𝐀−𝐒𝐭𝐚𝐠𝐞𝟏 = 𝐒𝐀(𝐂𝐨𝐧𝐯𝟏𝐀−𝐒𝐭𝐚𝐠𝐞𝟏 )
𝐂𝐨𝐧𝐯𝟐𝐀−𝐒𝐭𝐚𝐠𝐞𝟏 = 𝐂𝐨𝐧𝐯𝟑×𝟏,𝟑𝟐 (𝐂𝐨𝐧𝐯𝟑×𝟏,𝟏𝟔 (𝐂𝐨𝐧𝐯𝟑×𝟏,𝟏𝟔 (𝐂𝐨𝐧𝐯𝟏𝐀−𝐒𝐭𝐚𝐠𝐞𝟏 ))) (𝟐)
• Sub-stage B: This contains six convolutional layers and it is identical to sub-stage A, but with a 3
× 3 kernel size. That is illustrated in the following equations 3 to 4.

𝐂𝐨𝐧𝐯𝟏𝐁−𝐒𝐭𝐚𝐠𝐞𝟏 = 𝐂𝐨𝐧𝐯𝟑×𝟑,𝟑𝟐 (𝐂𝐨𝐧𝐯𝟑×𝟑,𝟏𝟔 (𝐂𝐨𝐧𝐯𝟑×𝟑,𝟏𝟔 (𝒊𝒏𝒑𝒖𝒕))) (𝟑)

𝐂𝐨𝐧𝐯𝟏𝐁𝐒𝐀 = 𝐒𝐀(𝐂𝐨𝐧𝐯𝟏𝐁−𝐒𝐭𝐚𝐠𝐞𝟏 )
𝐂𝐨𝐧𝐯𝟐𝐁−𝐒𝐭𝐚𝐠𝐞𝟏 = 𝐂𝐨𝐧𝐯𝟑×𝟑,𝟑𝟐 (𝐂𝐨𝐧𝐯𝟑×𝟑,𝟏𝟔 (𝐂𝐨𝐧𝐯𝟑×𝟑,𝟏𝟔 (𝐂𝐨𝐧𝐯𝟏𝐁−𝐒𝐭𝐚𝐠𝐞𝟏 ))) (𝟒)
• Sub-stage C: This stage also contains six convolutional layers with 1 × 3 kernel size. That is
illustrated in the following equations 5 to 6.

𝐂𝐨𝐧𝐯𝟏𝐂−𝐒𝐭𝐚𝐠𝐞𝟏 = 𝐂𝐨𝐧𝐯𝟏×𝟑,𝟑𝟐 (𝐂𝐨𝐧𝐯𝟏×𝟑,𝟏𝟔 (𝐂𝐨𝐧𝐯𝟏×𝟑,𝟏𝟔 (𝒊𝒏𝒑𝒖𝒕))) (𝟓)

𝐂𝐨𝐧𝐯𝟏𝐂𝐒𝐀−𝐒𝐭𝐚𝐠𝐞𝟏 = 𝐒𝐀(𝐂𝐨𝐧𝐯𝟏𝐂−𝐒𝐭𝐚𝐠𝐞𝟏 )
𝐂𝐨𝐧𝐯𝟐𝐂−𝐒𝐭𝐚𝐠𝐞𝟏 = 𝐂𝐨𝐧𝐯𝟏×𝟑,𝟑𝟐 (𝐂𝐨𝐧𝐯𝟏×𝟑,𝟏𝟔 (𝐂𝐨𝐧𝐯𝟏×𝟑,𝟏𝟔 (𝐂𝐨𝐧𝐯𝟏𝐂−𝐒𝐭𝐚𝐠𝐞𝟏 ))) (𝟔)
A Self-Attention (SA) mechanism is incorporated into each sub-stage subsequent to the first set of kernel
layers. One kernel flow's intermediate output from the SA layer is then added to another kernel flow's
final output. This guarantees that the data's local and global dependencies are captured by the model.
That is illustrated in the following equations 7 to 9.

𝐂𝐨𝐧𝐯𝟑𝐀−𝐒𝐭𝐚𝐠𝐞𝟏 = 𝐂𝐨𝐧𝐯𝟐𝐀−𝐒𝐭𝐚𝐠𝐞𝟏 + 𝐂𝐨𝐧𝐯𝟏𝐁𝐒𝐀−𝐒𝐭𝐚𝐠𝐞𝟏 (𝟕)

𝐂𝐨𝐧𝐯𝟑𝐁−𝐒𝐭𝐚𝐠𝐞𝟏 = 𝐂𝐨𝐧𝐯𝟐𝐁−𝐒𝐭𝐚𝐠𝐞𝟏 + 𝐂𝐨𝐧𝐯𝟏𝐂𝐒𝐀−𝐒𝐭𝐚𝐠𝐞𝟏 (𝟖)
𝐂𝐨𝐧𝐯𝟑𝐂−𝐒𝐭𝐚𝐠𝐞𝟏 = 𝐂𝐨𝐧𝐯𝟐𝐂−𝐒𝐭𝐚𝐠𝐞𝟏 + 𝐂𝐨𝐧𝐯𝟏𝐀𝐒𝐀−𝐒𝐭𝐚𝐠𝐞𝟏 (𝟗)

Technical Institute Bijeljina, Archives for Technical Sciences. Year XVI – N0 31 182
Rani Saritha, R. et al: Enhanced transformer-based…… Archives for Technical Sciences 2024, 31(2), 175-191
Second Stage Convolution (Conv_Stage-2)

• Sub-stage A: Like Stage 1, the model utilizes six convolutional layers in this stage to handle the
input using different filter and kernel sizes. 32 filters featuring a kernel size of 5 × 1, as well as 16
filters with the same kernel size. The process is similar to the satege1.

• Sub-stage B: With a kernel size of 5 × 5, this arrangement is similar to Substage A.

• Sub-stage C: Comparable to sub-stage A, but with a 1 × 5 kernel size.

Similar to Conv_Stage-1, SA methods are included following the initial set of kernels. To improve the
feature representation, their intermediate outputs are combined with the final outputs of the parallel kernel
flows. The intermediate convolution and self-attention modules are generated as in the stage1 mode. Then
the outcomes are further fused to provide the output of satge-2 in each substage A, B and C which is
illustrated in the following equations (10 to 12).
𝐂𝐨𝐧𝐯𝟑𝐀−𝐒𝐭𝐚𝐠𝐞𝟐 = 𝐂𝐨𝐧𝐯𝟐𝐀−𝐒𝐭𝐚𝐠𝐞𝟐 + 𝐂𝐨𝐧𝐯𝟏𝐁𝐒𝐀−𝐒𝐭𝐚𝐠𝐞𝟐 (𝟏𝟎)
𝐂𝐨𝐧𝐯𝟑𝐁−𝐒𝐭𝐚𝐠𝐞𝟐 = 𝐂𝐨𝐧𝐯𝟐𝐁−𝐒𝐭𝐚𝐠𝐞𝟐 + 𝐂𝐨𝐧𝐯𝟏𝐂𝐒𝐀−𝐒𝐭𝐚𝐠𝐞𝟐 (𝟏𝟏)
𝐂𝐨𝐧𝐯𝟑𝐂−𝐒𝐭𝐚𝐠𝐞𝟐 = 𝐂𝐨𝐧𝐯𝟐𝐂−𝐒𝐭𝐚𝐠𝐞𝟐 + 𝐂𝐨𝐧𝐯𝟏𝐀𝐒𝐀−𝐒𝐭𝐚𝐠𝐞𝟐 (𝟏𝟐)
Third Stage Convolution (Conv_Stage-3)

• Substage A: In this stage, the input is processed via three convolutional layers with different filter
and kernel sizes: 32 filters with a 3×1 kernel size and 16 filters with a 3 × 1 kernel size.

• Substage B: Identical setup, but with a 3 × 3 kernel.

• Substage C: Identical to Substage A and B, but with a 1 × 3 kernel size.

Similar to the above two convolution stages the outcome from the third convolution stage is detected
and shown its process in the following equations 13 to 15.

𝐂𝐨𝐧𝐯𝟑𝐀−𝐒𝐭𝐚𝐠𝐞𝟑 = 𝐂𝐨𝐧𝐯𝟐𝐀−𝐒𝐭𝐚𝐠𝐞𝟑 + 𝐂𝐨𝐧𝐯𝟏𝐁𝐒𝐀−𝐒𝐭𝐚𝐠𝐞𝟑 (𝟏𝟑)

𝐂𝐨𝐧𝐯𝟑𝐁−𝐒𝐭𝐚𝐠𝐞𝟑 = 𝐂𝐨𝐧𝐯𝟐𝐁−𝐒𝐭𝐚𝐠𝐞𝟑 + 𝐂𝐨𝐧𝐯𝟏𝐂𝐒𝐀−𝐒𝐭𝐚𝐠𝐞𝟑 (𝟏𝟒)
𝐂𝐨𝐧𝐯𝟑𝐂−𝐒𝐭𝐚𝐠𝐞𝟑 = 𝐂𝐨𝐧𝐯𝟐𝐂−𝐒𝐭𝐚𝐠𝐞𝟑 + 𝐂𝐨𝐧𝐯𝟏𝐀𝐒𝐀−𝐒𝐭𝐚𝐠𝐞𝟑 (𝟏𝟓)
Following these convolutional stages, the multi-dimensional feature maps are transformed into 1D
feature vectors that are ready for further processing by applying a Flatten Layer independently to each
substage of third stage-Conv output. That is illustrated in the following equations 16.

𝑭𝑨 = 𝑭𝒍𝒂𝒕𝒕𝒆𝒏(𝐂𝐨𝐧𝐯𝟑𝐀−𝐒𝐭𝐚𝐠𝐞𝟑 + 𝐂𝐨𝐧𝐯𝟑𝐁−𝐒𝐭𝐚𝐠𝐞𝟑 + 𝐂𝐨𝐧𝐯𝟑𝐂−𝐒𝐭𝐚𝐠𝐞𝟑 ) (𝟏𝟔)

Proposed Enhanced Transformer Block Integration

Next, an Enhanced Transformer Block (ETB) receives each flattened output. Substantially ETB
improves the model's capacity to categorize the lung nodule cancer types utilizing self-attention to
capture long-range correlations within the feature vectors. The flattened output 𝑭𝑨 is processed by the
Transformer block shown in figure 2. To create a full feature vector, the features from these ETB are
concatenated and combined together which is represented as ETBout

Final Classification Layers

The concatenated feature vector undergoes several processes to ensure accurate classification. First,
dropout is applied to help regularize the model by avoiding overfitting. This is followed by a fully
connected layer that aggregates the features and reduces their dimensionality for the final categorization.
Finally, the last layer employs a sigmoid activation function to provide a probability for each of the four

Technical Institute Bijeljina, Archives for Technical Sciences. Year XVI – N0 31 183
Rani Saritha, R. et al: Enhanced transformer-based…… Archives for Technical Sciences 2024, 31(2), 175-191
stages of lung cancer (A, B, E, and G). That is illustrated in the following equations 17 to 18.

𝑭𝑪 = 𝑭𝒖𝒍𝒍𝒚𝑪𝒐𝒏𝒏𝒆𝒄𝒕𝒆𝒅(𝑫𝒓𝒐𝒑𝒐𝒖𝒕(𝑬𝑻𝑩𝒐𝒖𝒕 )) (𝟏𝟕)
𝑶𝒖𝒕𝒑𝒖𝒕 = 𝑺𝒊𝒈𝒎𝒐𝒊𝒅(𝑭𝑪) (𝟏𝟖)
Enhanced Transformer Block

In the DKFSA-ET frame work the Transformer block, as depicted in Figure 4, the process begins by
flattening the final stage Conv-features (𝐹𝐴 ). This flattened representation undergoes Layer
Normalization to standardize the feature values, ensuring consistent scaling across different inputs. The
normalized features are then fed into the Multi-Head Attention (MHA) layer, where attention
mechanisms capture intricate relationships within (𝐹𝐴 ). The output of MHA, combined with features of
(𝐹𝐴 ) is added (𝐴1 ), undergoes another round of Layer Normalization for enhanced stability.
Subsequently, an MLP layer processes (𝐴1 ) to introduce nonlinear transformations and further refine
feature representations. Finally, (𝐴2 ) is formed by adding (𝐴1 ) to the output of the MLP, incorporating
enriched features for robust classification of lung cancer stages. This makes the model utilize the
normalization layer, attention layer and the MLP layer in sequence to achieve feature extraction and
model the entire dependency of the data well.

Figure 4. Transformer block architecture

𝑳𝑵𝑭𝑨 = 𝑳𝒂𝒚𝒆𝒓𝑵𝒐𝒓𝒎(𝑭𝑨 ) (𝟏𝟗)

𝑴𝑯𝑨𝑭𝑨 = 𝑴𝑯𝑨(𝑳𝑵𝑭𝑨 ) (𝟐𝟎)
𝑨𝟏 = 𝑳𝑵(𝑴𝑯𝑨𝑭𝑨 + 𝑭𝑨 ) (𝟐𝟏)
𝑴𝑳𝑷𝑨𝟏 = 𝑴𝑳𝑷(𝑳𝑵(𝑨𝟏)) (𝟐𝟐)
𝑨𝟐 = 𝑴𝑳𝑷𝑨𝟏 + 𝑨𝟏 (𝟐𝟑)
The following formulas demonstrate how multi-head attention, multi-layer perceptron operations, and
normalization are employed by the transformer block to enhance the feature set in equation 19 to 23. In
the Enhanced Transformer Block, the basic MHA module has been enhanced with the activation based
co-efficient generation for attention co-efficient. Which is illustrated in the following figure 5.

Figure 5. Workflow of multi-head attention in transformer block

Technical Institute Bijeljina, Archives for Technical Sciences. Year XVI – N0 31 184
Rani Saritha, R. et al: Enhanced transformer-based…… Archives for Technical Sciences 2024, 31(2), 175-191
Figure 5 outlines the workflow of Enhanced Multi-Head Attention (EMHA) in the Transformer Block,
illustrating the process of Query (Q), Key (K), and Value (V) interactions. Initially, intermediate
matrices M1 and M2 are computed by multiplying 𝑄 × 𝐾 𝑇 and 𝑉 × 𝐾 𝑇 , respectively. Outcomes of
SoftMax functions SM1 and SM2 are formed independently to create the attention co-efficient of them
which is then applied on V for forming M3 and M4. At last, there is the SoftMax SM3 that fuses M3
and M4 for the output of attention-weighted values. Lastly, these weighted values are used in conjunction
with the original input feature map by performing a matrix multiplication in order to produce the output
of the Enhanced Multi-Head Attention mechanism. This layered approach guarantees that the model
considers dependencies between Q, K, and V – which is mandatory for a proper feature extraction and
context merging in the Transformer.

RESULTS

The comparative analysis offers a comprehensive evaluation of the performance of the suggested model
in comparison with the baseline D3DR_MKCA [28] model. This assessment is important to identify the
gains of the proposed new model in relation to the diagnostic results, accuracy, precision, recall, and F1
scores for the classification of lung cancer.

Overall Performance Comparison

Table 1 provides the test and training image counts. Table 2 shows the performance of a D3DR_MKCA
[28] and table 3 gives the performance of the proposed model.

Table 1. Training and testing description of dataset

Cancer Type Total Training (80%) Testing (20%)

A 8010 6408 1602
B 4116 3292 824
E 1407 1125 282
G 4310 3448 862
Table 2. Performance Analysis of D3DR_MKCA [28] model

Class Accuracy Precision Recall F1 Score

A 0.9401 0.9660 0.9401 0.9529
B 0.9041 0.8922 0.9041 0.8981
E 0.8865 0.8475 0.8865 0.8666
G 0.9304 0.9103 0.9304 0.9203
Average 0.9153 0.9040 0.9153 0.9094
Table 3. Performance Analysis of Proposed model

Class Accuracy Precision Recall F1 Score

A 0.9544 0.9683 0.9544 0.9613
B 0.9102 0.9259 0.9102 0.9180
E 0.8582 0.8551 0.8582 0.8566
G 0.9292 0.8920 0.9292 0.9102
Average 0.9305 0.9103 0.9130 0.9117
Accuracy: It is found out that the model proposed has a mean accuracy of 93.05%, which is a new level
to the huge improvement than the previous 91.55% was obtained by the D3DR_MKCA model. This
increase will point out to the increased capacity of the proposed model in accurately identifying the cases
of lung cancers in all the classes shows in figure 6.

Technical Institute Bijeljina, Archives for Technical Sciences. Year XVI – N0 31 185
Rani Saritha, R. et al: Enhanced transformer-based…… Archives for Technical Sciences 2024, 31(2), 175-191

Figure 6. Comparison of Accuracy

Precision: Hence it has an average precision of 9.03% and above in all the experiments conducted on
it. Compared to 90.15% for the D3DR_MKCA model, the proposed model has better capability to
reduce False Positives. This is especially the case in clinical diagnosis whereby the accuracy of cancer
identification may affect treatment procedures and prognosis in figure 7.

Figure 7. Comparison of precision

Recall: The proposed model proves to have a slightly enhanced recall averaging 91.30% compared to
91.15 % for the baseline model. This shows that the proposed model has a better capability in identifying
true positive cases across the dataset a significant factor when it comes to efficient and accurate detection
of cancer at its early stage shows in figure 8.

Figure 8. Comparison of Recall

Technical Institute Bijeljina, Archives for Technical Sciences. Year XVI – N0 31 186
Rani Saritha, R. et al: Enhanced transformer-based…… Archives for Technical Sciences 2024, 31(2), 175-191
F1 Score: The mean of F1 score of the proposed model stands 91.17%, compared to 90.93% for the
D3DR_MKCA model. This is way better even though there is a minimal improvement the scale shows
that there was a consistent improvement across both the precision and recall boundaries as a sign of an
improved model when it comes to dealing with the imbalanced data set which is common with medical
datasets shows in figure 9.

Figure 9. Comparison of F1-Score

Class-specific Performance

Class A: Demonstrated the most significant improvement in all the metrics, which is important since
Adenocarcinoma is one of the most prevalent forms of lung cancer. The F1 score was enhanced from
95.29% to 96.13% that elicited in the proposed model of this class.

Class B: Other measurements also increased in all categories, and a notable increase in F1 scores from
89.81% to 91.80%, which indicates that the proposed model has improved performance in identifying
Small Cell Carcinoma; a highly malignant form of cancer.

Class E: However, it was the least improved regarding performance indicators. The lower values of the
recall and precision coefficients could mean that the model still might not meet the level of specificity
that is necessary for Large Cell Carcinoma, so the further tuning of the model and investigation of this
class is needed.

Class G: Showed better accuracy and a constant F1 score, which reestablish that the model was
appropriate in distinguishing squamous cell carcinoma.

Confusion Matrix Insights

Comparative analysis of control matrices for both models (figure 10 and 11) shows that the proposed
model detects less Cases A and B with lower false negatives, which is important to ensure no many
cancer cases are omitted. Thus, we see an increase of false positives for Class E – while the model seems
to be very sensitive, it is less accurate for Class E – thus, the balance must be further optimized. Using
the data presented in the document table 4, which is the set of results from the proposed lung cancer
diagnosis model, one can identify the ability of the model to accurately segment and classify lung
nodules from the medical imaging scans. The dataset includes cases where nodules are segmented and
assigned to particular classes that correspond to various types of lung cancer, including Adenocarcinoma
(A), Small Cell Carcinoma (B) and other.

Technical Institute Bijeljina, Archives for Technical Sciences. Year XVI – N0 31 187
Rani Saritha, R. et al: Enhanced transformer-based…… Archives for Technical Sciences 2024, 31(2), 175-191

Figure 10. Confusion matrix of D3DR_MKCA model [15]

Figure 11. Confusion matrix of proposed model

Specifically, the images demonstrate cases of proper classification, for example the classification of
class ‘A’ nodules, yet there are also issues of misclassification, where nodules belonging to one class
are classified into another class. This mixed outcome further speaks to the effectiveness of the model in
managing complicated imaging information while paying attention to the potential limitations in
discriminating between the two forms of nodules that are quite similar. The visualization of segmented
nodules along with their detected classes gives an accurate representation of the model’s results that can
be useful in the future to fine tune this methodology for even better diagnostic results.

Technical Institute Bijeljina, Archives for Technical Sciences. Year XVI – N0 31 188
Rani Saritha, R. et al: Enhanced transformer-based…… Archives for Technical Sciences 2024, 31(2), 175-191
Table 4. Classification Results for Segmented Lung Nodules

Test Sample Segmented nodule Detected Class

A

DISCUSSION

The overall significance of improvement in most of the metrics is a clear sign of how much the proposed
model enhances the idea of applying Transformer blocks in deep convolutional blocks. This has
obviously increased the merits of both architectures enabling the model to better make sense of complex
image features at better and precise levels. The improvements in accuracy and the F1 score for Classes
A and B indicate the model’s improved ability to detect and classify common types of lung cancer with
high reliability. As mentioned earlier, false negatives can be reduced significantly by the proposed
model, especially in the important classes of cancers. This reduction is important when the costs of
incorrectly labeling a given condition as negative compare to the actual positive status are very high as
in diagnosing certain medical conditions. The proposed model architecture ensures a better
understanding of medical images than before and can detect slight nodular patterns that were not visible
to other models earlier the proposed model architecture also has its drawbacks particularly in the
handling of Class E, Large Cell Carcinoma. The performance that has been achieved in this grouped
category could be even lower because there is more variability and less clear image characteristics of
Large Cell Carcinoma compared to other types of tumor. This would indicate the need for the model to
be fine-tuned to be more sensitive and specific especially for cancer types that may not have well defined
or small dataset.

CONCLUSION

This paper has proposed a very novel technique for the lung cancer diagnosis with the help of deep
learning techniques incorporating Transformer blocks as well as Self attention based deep convolutional
blocks. When compared to the baseline D3DR_MKCA model, the proposed model shows better
accuracy, precision, recall, and F1 scores on the detection and classification of Adenocarcinoma and/or
Technical Institute Bijeljina, Archives for Technical Sciences. Year XVI – N0 31 189
Rani Saritha, R. et al: Enhanced transformer-based…… Archives for Technical Sciences 2024, 31(2), 175-191
Small Cell Carcinoma. These modifications further confirm the model’s ability to fully utilize global
and local image features to enhance diagnosis accuracy and possibly benefit the patients through early
detection. Technological application of the Transformer enables a significant improvement in processing
spatial and contextual dependencies in medical image data. They have demonstrated special efficiency
in improving the model’s ability to identify diseases and differentiate between various types that are
essential in a healthcare setting where the classification of diseases determines the subsequent treatment
and the prognosis of patients’ conditions. Future research works for this study are to test other
combinations of deep learning architectures that may potentially increase the sensitivity and specificity
levels. Augmenting the dataset, especially by providing more samples of the underrepresented classes,
may aid in alleviating the current shortfalls in performance. Moreover, introducing approaches for the
comprehensible interpretation of the model and methods for showing their effectiveness in clinical
practice would be important intermediate steps for the enhancement of existing research progress into
efficient diagnostic tools that can be used in actual health care facilities.

REFERENCES

[1] Leiter A, Veluswamy RR, Wisnivesky JP. The global burden of lung cancer: current status and future trends.
Nature reviews Clinical oncology. 2023 Sep;20(9):624-639.
[2] Liao J, Li X, Gan Y, Han S, Rong P, Wang W, Li W, Zhou L. Artificial intelligence assists precision medicine
in cancer treatment. Frontiers in oncology. 2023 Jan 4;12:998222. [Link]
[3] Bi WL, Hosny A, Schabath MB, Giger ML, Birkbak NJ, Mehrtash A, Allison T, Arnaout O, Abbosh C, Dunn
IF, Mak RH. Artificial intelligence in cancer imaging: clinical challenges and applications. CA: a cancer
journal for clinicians. 2019 Mar;69(2):127-157.
[4] Alamer L, Alqahtani IM, Shadadi E. Intelligent Health Risk and Disease Prediction Using Optimized Naive
Bayes Classifier. Journal of Internet Services and Information Security. 2023 Jan;13(1):1-10.
[5] Mridha MF, Prodeep AR, Hoque AM, Islam MR, Lima AA, Kabir MM, Hamid MA, Watanobe Y. A
comprehensive survey on the progress, process, and challenges of lung cancer detection and classification.
Journal of Healthcare Engineering. 2022;2022(1):5905230. [Link]
[6] Yin XX, Hadjiloucas S, Zhang Y, Tian Z. MRI radiogenomics for intelligent diagnosis of breast tumors and
accurate prediction of neoadjuvant chemotherapy responses-a review. Computer methods and programs in
biomedicine. 2022 Feb 1;214:106510. [Link]
[7] Rajapaksa SM. Weakly Supervised Perturbation Based Method for 3D Brain Tumour Segmentation (Master's
thesis, University of Toronto (Canada)). 2022.
[8] Ramakrishnan J, Ravi Sankar G, Thavamani K. Publication Growth and Research in India on Lung Cancer
Literature: A Bibliometric Study. Indian Journal of Information Sources and Services. 2019;9(1):44-47.
[9] Quanyang W, Yao H, Sicong W, Linlin Q, Zewei Z, Donghui H, Hongjia L, Shijun Z. Artificial intelligence
in lung cancer screening: Detection, classification, prediction, and prognosis. Cancer Medicine. 2024
Apr;13(7):e7140. [Link]
[10] Sun G, Pan Y, Kong W, Xu Z, Ma J, Racharak T, Nguyen LM, Xin J. DA-TransUNet: integrating spatial
and channel dual attention with transformer U-net for medical image segmentation. Frontiers in
Bioengineering and Biotechnology. 2024 May 16;12:1398237. [Link]
[11] Zhi L, Jiang W, Zhang S, Zhou T. Deep neural network pulmonary nodule segmentation methods for CT
images: Literature review and experimental comparisons. Computers in Biology and Medicine. 2023 Aug
9:107321. [Link]
[12] Shashikala D, Chandran CP, Rajathi S. Cross-spectral vision transformer for lung nodule detection with
improved moth flame algorithm using deep learning. e-Prime-Advances in Electrical Engineering,
Electronics and Energy. 2024 Jun 1;8:100556. [Link]
[13] Li X, Li M, Yan P, Li G, Jiang Y, Luo H, Yin S. Deep learning attention mechanism in medical image
analysis: Basics and beyonds. International Journal of Network Dynamics and Intelligence. 2023 Mar 27:93-
116.
[14] Shamshad F, Khan S, Zamir SW, Khan MH, Hayat M, Khan FS, Fu H. Transformers in medical imaging: A
survey. Medical Image Analysis. 2023 Aug 1;88:102802. [Link]
[15] Odilov BA, Madraimov A, Yusupov OY, Karimov NR, Alimova R, Yakhshieva ZZ, Akhunov SA. Utilizing
Deep Learning and the Internet of Things to Monitor the Health of Aquatic Ecosystems to Conserve
Biodiversity. Natural and Engineering Sciences. 2024 May 5;9(1):72-83.
[16] Annavarapu CS, Parisapogu SA, Keetha NV, Donta PK, Rajita G. A Bi-FPN-based encoder–decoder model
for lung nodule image segmentation. Diagnostics. 2023 Apr 13;13(8):1406.
[Link]

Technical Institute Bijeljina, Archives for Technical Sciences. Year XVI – N0 31 190
Rani Saritha, R. et al: Enhanced transformer-based…… Archives for Technical Sciences 2024, 31(2), 175-191
[17] Liu G, Liu F, Gu J, Mao X, Xie X, Sang J. An attention-based deep learning network for lung nodule
malignancy discrimination. Frontiers in Neuroscience. 2023 Jan 9;16:1106937.
[Link]
[18] Wu Z, Li X, Zuo J. RAD-UNet: Research on an improved lung nodule semantic segmentation algorithm
based on deep learning. Frontiers in Oncology. 2023 Mar 23;13:1084096.
[Link]
[19] Saihood A, Karshenas H, Naghsh-Nilchi AR. Multi-Orientation local texture features for guided attention-
based fusion in lung nodule classification. IEEE Access. 2023 Feb 16;11:17555-17568.
[20] Ramana RH, Ravisankar V. Precision in Prostate Cancer Diagnosis: A Comprehensive Study on Neural
Networks. Journal of Wireless Mobile Networks, Ubiquitous Computing, and Dependable Applications
(JoWUA). 2024;15(2):109-122. [Link]
[21] Bhattacharyya D, Thirupathi Rao N, Joshua ES, Hu YC. A bi-directional deep learning architecture for lung
nodule semantic segmentation. The Visual Computer. 2023 Nov;39(11):5245-5261.
[22] Canayaz M, Şehribanoğlu S, Özgökçe M, Akıncı MB. A comprehensive exploration of deep learning
approaches for pulmonary nodule classification and segmentation in chest CT images. Neural Computing
and Applications. 2024 May;36(13):7245-7264.
[23] Gai L, Xing M, Chen W, Zhang Y, Qiao X. Comparing CNN-based and transformer-based models for
identifying lung cancer: which is more effective?. Multimedia Tools and Applications. 2024
Jun;83(20):59253-59269.
[24] Ali H, Mohsen F, Shah Z. Improving diagnosis and prognosis of lung cancer using vision transformers: a
scoping review. BMC Medical Imaging. 2023 Sep 15;23(1):129. [Link]
z
[25] Cao Z, Li R, Yang X, Fang L, Li Z, Li J. Multi-scale detection of pulmonary nodules by integrating attention
mechanism. Scientific Reports. 2023 Apr 4;13(1):5517. [Link]
[26] Yang S, Yang X, Lyu T, Huang JL, Chen A, He X, Braithwaite D, Mehta HJ, Wu Y, Guo Y, Bian J.
Extracting Pulmonary Nodules and Nodule Characteristics from Radiology Reports of Lung Cancer
Screening Patients Using Transformer Models. Journal of Healthcare Informatics Research. 2024 May 17:1-
5.
[27] Usman M, Shin YG. DEHA-Net: a dual-encoder-based hard attention network with an adaptive ROI
mechanism for lung nodule segmentation. Sensors. 2023 Feb 10;23(4):1989.
[Link]
[28] Jasim WA, Mohammed RJ. A Survey on Segmentation Techniques for Image Processing. Iraqi Journal for
Electrical & Electronic Engineering. 2021 Dec 1;17(2):73-93.
[29] Sanchez-Martinez S, Camara O, Piella G, Cikes M, González-Ballester MÁ, Miron M, Vellido A, Gómez E,
Fraser AG, Bijnens B. Machine learning for clinical decision-making: challenges and opportunities in
cardiovascular imaging. Frontiers in cardiovascular medicine. 2022 Jan 4;8:765693.
[Link]
[30] Chen W, Wang Y, Tian D, Yao Y. Ct lung nodule segmentation: A comparative study of data preprocessing
and deep learning models. IEEE Access. 2023 Apr 6;11:34925-34931.
[31] Zheng R, Wen H, Zhu F, Lan W. Attention-guided deep neural network with a multichannel architecture for
lung nodule classification. Heliyon. 2024 Jan 15;10(1). [Link]

Technical Institute Bijeljina, Archives for Technical Sciences. Year XVI – N0 31 191