AN INVESTIGATION INTO DETECTING PNEUMONIA THROUGH IMAGE PROCESSING AND OBJECT DETECTION

International Journal of Advanced Information Technology (IJAIT) Vol.14, No.5, October 2024
DOI : 10.5121/ijait.2024.14501 1
AN INVESTIGATION INTO DETECTING PNEUMONIA
THROUGH IMAGE PROCESSING AND
OBJECT DETECTION
Prithvi Sairaj Krishnan
Department of Computer Science, Westwood High School, Austin, United States of
America
ABSTRACT
One of the most common respiratory infections that causes substantial morbidity and mortality worldwideis
pneumonia, particularly in poorer countries with poor medical infrastructure. Chest X-ray imaging is
essential for early diagnosis, although it can be difficult. In order to identify pneumonia from chest X-rays,
this study created an automated deep learning computer-aided diagnosis method. Three pre-trained
convolutional neural network models (ResNet-18, DenseNet-121), together with a newly developed
weighted average ensemble approach based on evaluation metric scores, were used in the ensemble. Tested
using five-fold cross-validation on two public X-ray datasets for pneumonia, the methodoutperformed state-
of-the-art techniques with high accuracy (98.2%, 86.7%) and sensitivity (98.19%, 86.62%). Over 2.5
million fatalities globally are attributed to pneumonia each year. This precise automated model can help
radiologists diagnose patients in a timely manner, particularly in situations with limited resources. How it
is included into clinical decision assistance systems has the potential to improve pneumonia management
and outcomes significantly.
KEYWORDS
Convolutional Neural Networks, Pneumonia, Infection, X-Rays, Model, Machine Learning
1. INTRODUCTION
Pneumonia is a serious lung illness caused by bacteria, viruses, or fungus. Pneumonia can lead to
pleural effusion, a disorder characterized by fluid collection and inflammation of the air sacs in
the lungs. Pneumonia is a major cause of mortality for children under five, particularly in
developing and growing countries where there is a high pollution rate, overcrowding, poor
hygiene, and limited access to healthcare. For pneumonia to be treated effectively and prevent
death, it must be detected early. Radiological tests like computed tomography (CT), magnetic
resonance imaging (MRI), or X-rays are commonly used to detect pneumonia. X-ray imaging isa
fairly cost, non-invasive method of assessing the lungs. Infiltrates are white areas that are shown
by red arrows in the sample image, they distinguish a pneumonic condition from a healthy lung.
However, chest X-ray examinations for pneumonia detection are subject tosubjective variability.
Therefore, an automated system for pneumonia detection is necessary. In this study, the
researcher developed a computer-aided diagnosis (CAD) system that utilises an ensemble of deep
transfer learning models for the accurate classification of chest X-ray images to detect
pneumonia.

2
Fig 1. Examples of two X-ray plates that display (a) a healthy lung and (b) a pneumonic lung.
The red arrows in (b) indicate white infiltrates, a distinguishing feature of pneumonia. The images
were taken from the Kermany dataset [2].
Convolutional neural networks (CNNs), in particular, are potent artificial intelligence tools
frequently employed in deep learning to solve challenging computer vision problems. But for
these models to function at their best, a lot of data is needed, and this can be difficult to come
by for biomedical image classification tasks because each image must be classified by a team of
highly qualified clinicians, which is costly and time-consuming. One method to overcome this
challenge is transfer learning, which involves taking a model that was trained on a massive
dataset—like ImageNet, which has over 14 million images—and using the learned network
weights to solve a problem and make accurate predictions a final prediction for a test sample by
combining the decisions of numerous classifiers is a popular approach known as ensemble
learning. It seeks to extract the discriminative information from every base classifier to produce
more accurate predictions. Average probability, weighted average probability, and majority
voting are examples of common ensemble approaches. Although the average probability-based
ensemble gives every basic learner equal weight, it is a better idea to give the base classifiers
weights because some may be better at capturing information than others for a given task.
Nonetheless, it guarantees improved performance, it is essential to ascertain the ideal weight
values for every classifier.
For this work, I devised a unique weight allocation technique based on four assessment metrics:
accuracy, recall, f1-score, and area under the receiver operating characteristic (ROC) curve
(AUC). The ideal weights were assigned to three basic CNN models: GoogLeNet, ResNet-18,
and DenseNet-121. Previous studies have largely concentrated on classification accuracy when
figuring the base learner weights, which might not be enough, particularly when working with
datasets that aren't dispersed uniformly throughout the class. Furthermore, other criteria might
provide more relevant information for prioritizing the basic learners. The full process of the
recommended ensemble structure is shown in Figure 2.

3
Fig 2. Representation of the proposed pneumonia detection framework.
Pre = Precision score, Rec = Recall score, F1 = F1-score, AUC = AUC score, and A(i) = {Prei,
Reci, F1i, AUCi}; w(i) is the weight generated for the ith
base learner to compute the ensemble,is
the probability score for the jth
sample by the ith
classifier, and enjoy is the fused probability score
for the jth
sample; and the arg max function returns the position having the highest value in a 1D
array, i.eIn this case, it generates the predicted class of the sample.
2. RELATED WORK
Table 1. Existing methods for pneumonia detection

4
3. MOTIVATIONS AND CONTRIBUTIONS
Many people, particularly children, suffer greatly from pneumonia. This condition is most
common in developing and impoverished nations when risk factors such as overcrowding, poor
hygiene, hunger, and a lack of proper medical facilities are present. It takes an early diagnosis to
fully recover from pneumonia. The most popular diagnostic technique is X-ray examination,
however, it depends on the radiology's interpretive skills, which frequently causes radiologists to
disagree. For an accurate diagnosis, then, a generalisation-capable automated computer-aided
diagnosis (CAD) system is required. The majority of earlier research ignored the possible
advantages of ensemble learning in favor of creating a single convolutional neural network
(CNN) model for the categorization of pneumonia. Better predictions are made possible by
ensemble learning, which combines discriminative data from several base learners. Ensemble
learning was used in this study to address a lack of medical data by using transfer learning models
as base learners and ensembling their decision scores.
By using a weighted average ensemble approach, an ensemble framework was created to improve
the performance of basic CNN learners in the categorization of pneumonia cases. Rather than
relying just on classifier performance or experimental results, the weights assigned to the
classifiers were determined by integrating four assessment metrics: precision, recall, f1-score, and
area under the curve (AUC), using a hyperbolic tangent function. The RSNA Pneumonia
Detection Challenge dataset and the Kermany dataset, two publically available chest X-ray
datasets, were used to evaluate the proposed model using five-fold cross-validation.The results
outperformed the state-of-the-art methods, suggesting that the method may be applied in practical
settings.
4. PROPOSED METHOD
In this study, I designed an ensemble framework consisting of three classifiers: GoogLeNet,
ResNet-18, and DenseNet-121, using a weighted average ensemble scheme. The weights
allocated to the classifiers were generated using a novel scheme, as explained in detail below.
4.1. Googlenet
The GoogLeNet design, which was proposed by Szegedy et al., is a 22-layer deep network that
employs "inception modules" as opposed to layers that are uniformly progressive. An inception
block may accommodate many units at each level by supporting parallel convolution and pooling
layers. Nevertheless, the extra parameters lead to an unmanageable computing complexity. The
GoogLeNet model employs inception blocks with dimension reduction, as shown in Fig. 3(b), to
control the computational complexity as opposed to the naïve inception block (Fig. 3(a)) used in
previous work. GoogLeNet's performance, which introduced the inception block, shows that an
optimal sparse architecture made from easily obtainable dense building blocks improves the
performance of artificial neural networks for computer vision applications. Design of the
GoogLeNet model shown in Fig. 4.

5
Fig 3. Inception modules in the GoogLeNet architecture.
(a) The naive inception block is replaced by (b) the dimension reduction inception block in
the GoogLeNet architecture to improve computational efficiency.
Fig 4. The architecture of the GoogLeNet model was used in the study
The inception block is shown in Fig 3(b).
4.2. Resnet-18
Deep network training is made more successful by Huang et al.'s ResNet-18 model, which is
based on a residual learning methodology. The residual blocks of ResNet models aid in network
optimization, improving model accuracy overall. This is distinct from the initial unreferenced
mapping present in inversely continuing convolutions. These residuals, or links, provide identity
mapping without adding parameters or increasing computing complexity. The architecture of the
ResNet-18 model is shown in Figure 5.

6
Fig 5. The architecture of the ResNet-18 model used in this study
4.3.Densenet-121
According to Huang et al., DenseNet topologies provide a rich feature representation and are
computationally efficient. The primary rationale is because each layer of the DenseNet model's
feature maps are concatenated with feature maps from all preceding levels, as seen in Fig. 6.
Because the convolutional layers can accommodate fewer channels, the model becomes
computationally efficient when the number of trainable parameters decreases. Concatenating the
feature maps from previous layers with the current layer further enhances the feature
representation capacity.
Fig 6. The basic architecture of the DenseNet convolutional neural network model.

7
The values of the hyperparameters used for training the learning algorithms (base learners) were
set empirically and are shown in Table 2.
Table 2. Hyperparameters are used for training the convolutional neural network base learners.
5. PROPOSED ENSEMBLE SCHEME
Better predictions than any of its base learners are produced by the ensemble learning model,
which assists in incorporating the discriminative information from all of its constituent models.
The weighted average ensemble is an effective method for classifier fusion. But one of the most
important factors in guaranteeing the ensemble's success is the selection of weights given to the
corresponding base learners. The majority of methods in the literature either experiment or just
consider the accuracy of the classifier when determining the weights. If there is a class imbalance
in the dataset, this method might not be appropriate. Other assessment metrics, such f1-score, area
under the curve (AUC), recall (sensitivity), and accuracy, may offer more reliable data for
establishing the base learners' priority. This study came up with a unique plan to achieve this goal
for weight allocation, which is explained below.
First, the probability scores obtained during the training phase by the base learners are utilised to
calculate the weights assigned to each base learner using the proposed strategy. These generated
weights are used in the formation of an ensemble trained on the test set. This strategy is
implemented to ensure that the test set remains independent for predictions. The predictions of
the ith
model are generated and compared with the true labels (y) to generate the corresponding
precision score (prei
), recall score (reci
), f1-score (f1i
), and AUC score (AUCi
). Assume that this
forms an array Ai
= {prei
, reci
, f1i
, AUCi
}. The weight (wi
) assigned to each classifier is then
computed using the hyperbolic tangent function, as shown in Eq 1. The range of the hyperbolic
tangent function is [0, 0.762] because x represents an evaluation metric, the value of which is in
the range [0, 1]. It monotonically increases in this range; thus, if the value of a metric x is high,
the tanh function rewards it by assigning it a high priority; otherwise, the function penalises it.
These weights (w(i)) computed by Eq 1 are multiplied by the decision scores of the
corresponding base learners to compute the weighted average probability ensemble, as shown in
Eq 2, where the probability array (for a binary class dataset) of the jth
test sample by the ith
base
classifier is, where a ≤ 1 and the ensemble probability for the sample is ensemble_probj ={b, 1 −
b}. Finally, the class predicted by the ensemble is computed by Eq 3, where predictionj denotes
the predicted class of the sample.
6. RESULTS AND DISCUSSION
This section displays the evaluation results of the recommended approach. I utilized two freely
available datasets of chest X-rays for pneumonia. 5856 chest X-ray pictures that are unequally
distributed between the "Normal" and "Pneumonia" classifications make up the first dataset,
known as the Kermany dataset. The images feature a diverse range of people and kids. The
second dataset was released as a Kaggle challenge for pneumonia detection and made available

8
by the RSNA. Table 3 displays the distribution of images between the two datasets as well as the
picture descriptions for the training and testing sets for each fold of the five-fold cross-validation
approach employed in this investigation. Additionally, the implications of the obtained results are
discussed. A comparative examination was done to show how much better the proposed method
is over other models and frequently used ensemble techniques published in the literature.
Table 3. Description of images in the training and testing sets in each fold of five-fold cross-validation in the
two datasets used in this study.
7. EVALUATION METRICS
Four common evaluation measures were applied to the two pneumonia datasets to assess the
suggested ensemble method: f1-score (F1), accuracy (Acc), precision (Pre), and recall (Rec).
First, I define the phrases "True Positive," "False Positive," "True Negative," and "False
Negative" to define these evaluation measures. Now, the four evaluation metrics can be defined as:
Fig 7. The different evaluation metrics of the pneumonia detection ensemble model using components of the
confusion matrix make up the evaluation metrics.
The accuracy rate provides a broad idea of the proportion of the model's predictions that were
realized. A model's high accuracy rate does not, however, imply that it can differentiate between
several classes equally if the dataset is imbalanced. More specifically, medical image
categorization requires a globally applicable model. In these cases, looking at the "precision" and
"recall" variables will help you understand how well the model performs. The accuracy of the
positive label prediction made by the model is displayed. This is the ratio between all of the
model's predictions and the accurate forecasts. Conversely, "recall" measures how much of the
positive ground truth data the model correctly predicted. FN and FP quantities can be reduced by
the model based on these two evaluation criteria. The accuracy rate provides a broad idea ofthe
proportion of the model's predictions that were realized. A model's high accuracy rate does not,
however, imply that it can differentiate between several classes equally if the dataset is
imbalanced. More specifically, medical image categorization requires a globally applicable

9
model. In these cases, looking at the "precision" and "recall" variables will help you understand
how well the model performs. The accuracy of the positive label prediction made by the modelis
displayed. This is the ratio between all of the model's predictions and the accurate forecasts.
Conversely, "recall" measures how much of the positive ground truth data the model correctly
predicted. FN and FP quantities can be reduced by the model based on these two evaluation
criteria.
8. IMPLEMENTATION
This work employed a 5-fold cross-validation technique to evaluate the performance of the
proposed ensemble model in detail. Tables 4 and 5 display the findings for the RSNA challenge
dataset and the Kermany dataset, respectively, along with the average and standard deviation
values for all five folds. The outstanding accuracy and sensitivity (recall) ratings show the
reliability of the recommended strategy. Additionally, Figures 7 and 8 display the confusion
matrices on the RSNA and Kermany datasets, and Figure 8 displays the ROC curves generated by
the recommended approach for each of the two datasets' five cross-validation folds.
Fig 8. Confusion matrices were obtained on the Kermany pneumonia chest X-ray dataset by the proposed
method by 5-fold cross-validation. a) Fold-1. (b) Fold-2. (c) Fold-3. (d) Fold-4. (e) Fold-5.

10
Fig 9. Confusion matrices obtained on the Radiological Society of North America pneumoniachallenge
chest X-ray dataset by the proposed method by five-fold cross-validation. a) Fold-1. (b) Fold-2. (c) Fold-3.
(d) Fold-4. (e) Fold-5.
Fig 10. Receiver operating characteristic curves obtained by the proposed ensemble method on the two
pneumonia chest X-ray datasets used in this research:(a) Kermany dataset [2]. (b) RSNA challenge dataset
[16].

11
Table 4. Results of five-fold cross-validation of the proposed ensemble method on the pneumonia Kermany
dataset [2].
Table 5. Results of five-fold cross-validation of the proposed ensemble method on the pneumonia
Radiological Society of North America challenge dataset.
Figure 10 showcases the accuracy rates achieved by the base learners in transfer learning using
different optimizers on the Kermany dataset. The Adam optimizer yielded the best results for all
three base learners and was consequently chosen as the optimizer for training the base learners in
the ensemble framework.
Fig 11. Variation of accuracy rates on the Kermany dataset [2]) was achieved by the three base learners,
GoogLeNet, ResNet-18, and DenseNet-121, and their ensemble, according to the optimizers chosen for
fine-tuning.

12
The results of many ensembles using three different base learners are displayed in Table 6, which
also includes newly proposed architectures on the Kermany dataset: GoogLeNet, ResNetvariants,
DenseNet variations, MobileNet v2, and NASMobileNet. The results validate the choice of base
learner combinations used in this study, namely GoogLeNet, ResNet-18, and DenseNet-121. This
ensemble combination has a 98.2% accuracy rating. The ensemble consisting of GoogLeNet,
ResNet-18, and MobileNet v2 yielded the second-best result, with anaccuracy rate of 98.54%. In
addition, the models were trained to find the optimal configuration once a few layers were fixed
for the chosen group of base learners. The findings, which are shown in Figure 12, demonstrate
that the ensemble worked best on both datasets when every layer was trainable (0 layers frozen).
Because of this, the precise setting was chosen for the ensemble framework.
Fig 12. Variation in performance (accuracy rates) of the ensemble concerning the number of fixed non-
trainable layers in the base learners on the two datasets used in this study:(a) Kermany dataset [2]. (b)
RSNA challenge dataset [16].

13
Table 6. Results of extensive experiments performed to determine the base learners for forming the
ensemble in this study.
8.1. Comparison with State-of-the-Art Methods
On the Kermany pneumonia dataset, Table 7 presents a performance comparison between the
suggested ensemble framework and the approaches that are already available in the literature. It
should be mentioned that the suggested approach performed better than any other approach.
Moreover, it is noteworthy that the proposed ensemble framework outperformed all of the
previous methods (Mahmud et al. [18], Zubair et al. [8], Stephen et al. [15], Sharma et al. [14],
Liang et al. [6]) that relied on using a single CNN model for the classification of pneumonic lung
X-ray images. This suggests that the ensemble technique developed in this study is a dependable
method for the image classification task at hand. As far as we are aware, no research has been
done on the categorization of pictures in the RSNA pneumonia dataset exist. Hence, for this
dataset, I compared the performance of the proposed model to that of several baseline CNN
models.
Table 7. Comparison of the proposed method with other methods in the literature on the Kermany
pneumonia dataset [2] and the Radiological Society of North America challenge dataset [16].

14
Table 8 compares the assessment results of the proposed technique with those of the basic CNN
models used to construct the ensemble and various other conventional CNN transfer learning
models on both datasets utilized in this work. On both datasets, it is evident that the suggested
ensemble approach did rather well compared to alternative transfer learning models and the base
learners. In addition, Table 9 compiles the findings to demonstrate the superiority of the
suggested ensemble scheme over conventional popular ensemble strategies. For both the
Kermany and RSNA challenge datasets, the average results across the five folds of cross-
validation are displayed. The ensembles employed the same three basic CNN learners,
GoogLeNet, ResNet-18, and DenseNet-121. Popular ensemble techniques were outperformed by
the suggested ensemble approach. It is evident from both datasets that the weighted average
ensemble that uses the accuracy metric as the sole weighting factor produced the best results,
nearly matching the suggested ensemble approach. The class that received the most votes from
the base learners is expected to be the sample class in the majority voting-based ensemble. In the
maximum probability ensemble, all base learners' probability scores are added up, and the class
with the highest probability is designated as the sample's predicted class. In the average
probability ensemble, on the other hand, each contributing classifier is given the same weight.
Table 8. Comparison of the proposed ensemble framework with several standard convolution neural network
models in the literature on both the Kermany and the Radiological Society ofNorth America challenge
datasets.
The same base learners were used in all the ensembles: GoogLeNet, ResNet-18, and DenseNet-
121.
9. CONCLUSION AND FUTURE WORK
To treat pneumonia appropriately and keep the patient's life from being in danger, early
recognition of the illness is essential. The most common method for diagnosing pneumonia is a
chest radiograph; however, there can be inter-class variation in these images, and the diagnosis
relies on the doctor's skill to identify early signs of pneumonia. In this work, an automated

15
CAD system was created to aid medical professionals. It classifies chest X-ray pictures into two
groups, "Normal" and "Pneumonia," using deep transfer learning-based classification. A
weighted average ensemble is formed by an ensemble framework that takes into account the
decision scores from three CNN models: GoogLeNet, ResNet-18, and DenseNet-121. An
innovative technique was used to calculate the weights assigned to the classifiers, whereby five
evaluation parameters, accuracy, precision, recall, f1-score, and AUC, were fused using the
hyperbolic tangent function. The framework, evaluated on two publicly available pneumonia
chest X-ray datasets, obtained an accuracy rate of 98.2%, a sensitivity rate of 98.19%, a precision
rate of 98.22%, and an f1-score of 98.29% on the Kermany dataset and an accuracy rate of
86.7%, a sensitivity rate of 86.62%, a precision rate of 86.69%, and an f1-score of 86.65% on the
RSNA challenge dataset, using a five-fold cross-validation scheme. It outperformed state-of-the-
art methods on these two datasets. Statistical analyses of the proposed model using McNemar’s
and ANOVA tests indicate the viability of the approach. Furthermore, the proposed ensemble
model is domain-independent and thus can be applied to a large variety of computer vision tasks.
But as was previously mentioned, there were instances in which the ensemble architecture was
unable to produce precise estimates. To improve the quality of the photos, I could investigate
techniques like picture contrast enhancement or other pre-processing steps in the future. Before
categorizing the lung picture, I recommend segmenting it to assist the CNN models extract
additional characteristics from it. Furthermore, since three CNN models are required to train the
recommended ensemble, its computational cost is higher than that of the CNN baselines
developed in works published in the literature. To try to reduce the processing needs in the future,
I may consider employing strategies like snapshot ensembling increasing overall performance.
ACKNOWLEDGEMENTS
I would like to acknowledge my parents for buying me the computer on which I did all of my
research.
REFERENCES
[1] WHO Pneumonia. World Health Organization. (2019), https://www.who.int/news-room/fact-
sheets/detail/pneumonia
[2] Kermany D., Zhang K. & Goldbaum M. Labeled Optical Coherence Tomography (OCT) and Chest
X-Ray Images for Classification. (Mendeley,2018)
[3] Dalhoumi S., Dray G., Montmain J., Derosière, G. & Perrey S. An adaptive accuracy-weighted
ensemble for inter-subjects classification in brain-computer interfacing. 2015 7th International
IEEE/EMBS Conference On Neural Engineering (NER). pp. 126-129 (2015)
[4] Albahli S., Rauf H., Algosaibi A. & Balas V. AI-driven deep CNN approach for multi-label
pathology classification using chest X-Rays. PeerJ Computer Science. 7 pp. e495 (2021)
pmid:33977135
[5] Rahman T., Chowdhury M., Khandakar A., Islam K., Islam K., Mahbub Z., et al. Transfer learning
with deep convolutional neural network (CNN) for pneumonia detection using chest X-ray. Applied
Sciences. 10, 3233 (2020)
[6] Liang G. & Zheng L. A transfer learning method with deep residual network for pediatric
pneumonia diagnosis. Computer Methods And Programs In Biomedicine. 187 pp. 104964 (2020)
pmid:31262537
[7] Ibrahim A., Ozsoz M., Serte S., Al-Turjman F. & Yakoi P. Pneumonia classification using deep
learning from chest X-ray images during COVID-19. Cognitive Computation. pp. 1–13 (2021)
pmid:33425044
[8] Zubair S. An Efficient Method to Predict Pneumonia from Chest X-Rays Using Deep Learning
Approach. The Importance Of Health Informatics In Public Health During A Pandemic. 272 pp. 457
(2020)

16
[9] Rajpurkar P., Irvin J., Zhu K., Yang B., Mehta H., Duan T., et al. & Others Chexnet: Radiologist-
level pneumonia detection on chest x-rays with deep learning. ArXiv Preprint ArXiv:1711.05225.
(2017)
[10] Albahli S., Rauf H., Arif M., Nafis M. & Algosaibi A. Identification of thoracic diseases by
exploiting deep neural networks. Neural Networks. 5 pp. 6 (2021)
[11] Chandra T. & Verma K. Pneumonia detection on chest X-Ray using machine learning paradigm.
Proceedings Of 3rd International Conference On Computer Vision And Image Processing. pp. 21-33
(2020)
[12] Kuo K., Talley P., Huang C. & Cheng L. Predicting hospital-acquired pneumonia among
schizophrenic patients: a machine learning approach. BMC Medical Informatics And Decision
Making. 19, 1–8 (2019) pmid:30866913
[13] [13] Yue H., Yu Q., Liu C., Huang Y., Jiang Z., Shao C., et al. & Others Machine learning-
based CT radiomics method for predicting hospital stay in patients with pneumonia associated with
SARS-CoV-2 infection: a multicenter study. Annals Of Translational Medicine. 8 (2020)
pmid:32793703
[14] Sharma H., Jain J., Bansal P. & Gupta S. Feature extraction and classification of chest x-ray images
using cnn to detect pneumonia. 2020 10th International Conference On Cloud Computing, Data
Science & Engineering (Confluence). pp. 227-231 (2020)
[15] Stephen O., Sain M., Maduh U. & Jeong D. An efficient deep learning approach to pneumonia
classification in healthcare. Journal Of Healthcare Engineering. 2019 (2019) pmid:31049186
[16] Wang X., Peng Y., Lu L., Lu Z., Bagheri M. & Summers R. Chestx-ray8: Hospital-scale chest x-ray
database and benchmarks on weakly-supervised classification and localTocommon thorax diseases.
Proceedings Of The IEEE Conference On Computer Vision And Pattern Recognition. pp. 2097-
2106 (2017)
[17] Selvaraju R., Cogswell M., Das A., Vedantam R., Parikh D. & Batra D. Grad-cam: Visual
explanations from deep networks via gradient-based localization. Proceedings Of The IEEE
International Conference On Computer Vision. pp. 618-626 (2017)
[18] Mahmud T., Rahman M. & Fattah S. CovXNet: A multi-dilation convolutional neural network for
automatic COVID-19 and other pneumonia detection from chest X-ray images with transferable
multi-receptive feature optimization. Computers In Biology And Medicine. 122 pp. 103869 (2020)
pmid:32658740
[19] Dietterich T. Approximate statistical tests for comparing supervised classification learning
algorithms. Neural Computation. 10, 1895–1923 (1998) pmid:9744903
[20] Cuevas A., Febrero M. & Fraiman R. An anova test for functional data. Computational Statistics &
Data Analysis. 47, 111–122 (2004)
AUTHOR
Prithvi is a driven high school student set to graduate in 2025 with an impressive academic
record in computer science and STEM fields. His diverse pursuits, ranging from founding a
global AI education platform to pioneering research in AI-powered medical imaging
analysis, exemplify his innovative mindset and passion for developing socially impactful
technology. With an unwavering determination to be at the forefront of ethical AI
innovation, Prithvi aims to continue pushing boundaries and creating lasting positive
impacts in the field.

AN INVESTIGATION INTO DETECTING PNEUMONIA THROUGH IMAGE PROCESSING AND OBJECT DETECTION

More Related Content

Similar to AN INVESTIGATION INTO DETECTING PNEUMONIA THROUGH IMAGE PROCESSING AND OBJECT DETECTION (20)

Recently uploaded (20)

AN INVESTIGATION INTO DETECTING PNEUMONIA THROUGH IMAGE PROCESSING AND OBJECT DETECTION