Page 275 - 2024-Vol20-Issue2

271 | Qadir, Abdalla & Abd

drastically lowers the number of required parameters. TABLE II.
The XGBoost classifier was added to replace the top layer ANALYZING PERFORMANCE MEASURES IN COMPARISON
of the previously mentioned network, namely VGG16. This
study uses a CT scan image dataset that is labeled. Therefore, TO OTHER TECHNIQUES
inductive transfer learning is the type of learning that is used
in this context since the ImageNet weights are used by the Study Technique Dataset Accuracy%
VGG16 model. The VGG16 model’s weight, which was used
to classify 1000 pictures out of a dataset of 14 million, is trans- The proposed Hybrid- IQ- 98.54
ferred and utilized to conduct feature extraction on the CT method LCSCDM OTHNCCD 94.38
image dataset. The VGG16 model feature extraction findings AL-Huseiny GoogLeNet IQ- 89.88
are sent to the XGBoost algorithm [26, 27]. et al. [29] OTHNCCD 93.54
Kareem et al. SVM IQ-
E. Classification using XGBoost classifier [30] OTHNCCD
Extreme Gradient boosting, or XGboost, uses a decision tree Al-Yasriy et al. AlexNet IQ-
technique with gradient boosting. XGboost is built for speed [31] OTHNCCD
and efficiency. Its engineering objective is to push the com-
putational resource limitations considering enhanced tree al- whereas the XGboost model was applied for the calcification
gorithms. Python, C++, Java, R, and other programming phase in both the training and testing stages.
languages are just a few of the interfaces available for XG-
boost. In this study, the Python interface was used. The IV. RESULTS AND DISCUSSION
approach optimizes computing speed and memory use, while
the term “Boosting” is a strategy that uses an ensemble tech- In this part, the performance of the suggested (Hybrid-LCSCDM)
nique in which new models are introduced to older models to is evaluated, and experimental findings are shown. We em-
rectify their flaws. Models are successively introduced until ployed a transfer learning method using VGG16 as the foun-
no more improvements are possible. A well-known exam- dation model to extract features from CT scan data, which
ple of boosting is the AdaBoost algorithm, which weights were then used to train an XGBoost classifier.
difficult-to-predict data points, while Gradient boosting is a For the testing and evaluation phases of the proposed hybrid
technique that works on creating new models to forecast mis- model, the publicly available (IQ-OTH/NCCD) dataset was
takes or the residuals of previous models and merging them used. For efficient testing of the model, the Our research
to produce the final prediction. While adding new models experiments were conducted on a computer with an Intel Core
to minimize loss, a gradient descent approach is employed, i7-13700F CPU, NVIDIA RTX 4060 GPU, and 64 GB of
which the name Gradient boosting comes from. With each RAM, providing the necessary computational power and effi-
additional model, the accuracy of the prediction increases. ciency for our data-intensive tasks.
The aim of the model of gradient descent for a boosted tree In Table II, in terms of Recall and accuracy, we contrasted our
is derived from (5) using Taylor expansion. Equation (6) in- performance measures with the outcomes of existing models.
cludes the regularization, whereas (7) accounts for the tree’s In Table II, the evaluation metrics are shown; the Hybrid-
generalization. The regularization aim of XGboost is to LCSCDM (VGG-16-XGboost) model has achieved an accu-
choose a model with simple prediction functions [28]. racy of 98.5%, which is higher than the compared models
related to lung cancer detection. In Fig. 4, we have provided

? * = - B Aj ? (5) TABLE III.
j j+ EVALUATION METRICS FOR THE HYBRID-LCSCDM

MODEL

?f (” ”) = -1 T A2j +?T (6) Fold Accuracy Precision Recall F1-Score
obj 2 j=1 Bj +? 1 97.73% 0.9818 0.9306 0.9519
2 99.10% 0.9821 0.9821 0.9821
1 A2L + A2R + (AL + AR)2 - ? (7) 3 98.17% 0.9853 0.9543 0.9684
”Gain” = BL + ? BR + ? BL + BR + ? 4 99.09% 0.9930 0.9821 0.9874
5 98.64% 0.9891 0.9682 0.9780
2 98.54% 0.9863 0.9635 0.9736
Average

Fig. 3 demonstrates the suggested model’s design. Mainly, a visual representation of the results obtained by the Hybrid-
The VGG-16 model was employed for feature extraction, LCSCDM model. This visualization aims to offer a clear and