Cardiomegaly Prediction Using Deep Learning

Cardiomegaly Prediction Using Deep Learning

Aaditya Sharma

Computer Science and Engineering Meerut Institute of Engineering and Technology Meerut, India (2200680100002)

Abhishek Tyagi

Computer Science and Engineering Meerut Institute of Engineering and Technology Meerut, India (2200680100018)

Vishakha Rohila

Computer Science and Engineering Meerut Institute of Engineering and Technology Meerut, India

Akash Sharma

Computer Science and Engineering Meerut Institute of Engineering and Technology Meerut, India (2200680100035)

Abhinav Gupta

Computer Science and Engineering Meerut Institute of Engineering and Technology Meerut, India (2200680100014)

Abstract – The enlargement of heart is a condition known as cardiomegaly. Signicant diagnostic marker of cardiovascular diseases. Chest X Rays are one of the most readily available and automated Diagnostic tools. But, manual Interpretation is time-consuming and it has a tendency to interobserver variability. The current paper analyzes the functionality of two deep learning techniques DenseNet and EfcientNet architectures – automated diagnosis with cardiomegaly based on chest X-ray. Transfer preprocessing techniques were applied together with learning, such as data augmentation, rescaling, and normalization to enhance feature extraction and model robustness. DenseNet performed more successfully, with high training F1 score and test accuracies and validation respectively while Efcient net model achieved an F1 score of 77.1%. These ndings indicate that the reliabilities of the models given by EfcientNet and medical accuracy. application, and have a possible impact of diagnostic errors reduction, and support radiologists by greater workow efciencies. The current studies underline the increasing role of articial intelligence in medical imaging particularly in resource-constrained environments. The Monograph deals mainly with the above, that concerns the diagnosis of the cardiomegaly, it also focuses on the use of the same technology opportunities for applying the same to other thoracic emergencies. The future research is geared towards ensuring that there are no false positives. as little as may be. Assess the performance of multiple data sets and optimize AI models to wider clinical integrations.

1. Introduction

   One of the major causes of death is cardiovascular diseases. Among the leading causes of death in the world, leading to tens of thousands of deaths per year due to heart failure, arterial of millions obstructions, and related complications of hypertension. In general terms Cardiomegaly: excessive

   enlargement of the heart assumes cardiomegaly as signicant clinical indicator of a profound pathology or pathophysiology of the heart. The early and accurate diagnosis of cardiomegaly is vital, because the development of cardiomegaly can be avoided by timely treatment of heart diseases will result in better patient outcomes. One of the most common is x-ray imaging. Common techniques to scan the chest area in use, diagnostic methods of identifying cardiomegaly as a result of its portability, availability and appropriateness in healthcare. But traditional diagnostic practices are based on heavy reliance on manual interpretation by radiologists, which may be highly slow-moving and prone to variability. These may be visual fatigue, differences in expertise, judgment, which may induce a degree of incongruities, and which point out the necessity, to a higher level of reliability and a diagnostic process that is not so human dependent. The advances in articial intelligence can provide a Chance in machine learning addressing all the challenges. The CNNs are used in convolutional neural networks (CNNs) specic, demonstrated remarkable medical image capabilities. They are considered to be analysis because they have the ability of feature learning. Here we explore the two most recent and modern architectures: DenseNet, Efcient automated cardiomegaly detection. DenseNet is characterized by its super-networked pattern which enhances feature reuse and assists gradient ow and EfcientNet model uses a com-pound scaling strategy to deal with network depth and width and solution to better computational Efciency. Both transfer learning is used in models to the extent that it would speed up acclimatization to the cardiomegaly classication task. The purpose of this study is to evaluate these models in the respect of diagnostic accuracy, computational efciency and usability.

2. Literature Review

   There has been considerable advancement in this eld over the years. contributed signicantly in automating analysis of cardiac images, transitioning from traditional machine learning methods to more advanced deep learning approaches. The earlier studies were algorithm-based such as Support Vector Machines, Decision Trees, and random Forests which normally required hand-crafted features derived from chest X-rays. Most of these techniques required the use of calculating metric such as Cardiothoracic ratio to evaluate the possibility of cardiomegaly. Although these methods offered useful base-lines, these methods offered results, their dependence on manually engineered features made them highly sensitive to variations in image quality, patient position, and acquisition conditions, which may lead to inconsistent performance.The advent of deep learning brought a huge change in medical im-age analysis. Traditional techniques, Deep learning algorithms learn these methods automatically, They learn signicant fea-tures from the data itself, thereby overcoming the need for manual feature extraction. Convolutional Neural Networks in particular, the following have proven to be highly effective in analyzing complex patterns in medical images. DenseNet improves feature reuse through linking every layer to all the succeeding layers, enhanced gradient ow, and enabling strong performance even on relatively small datasets. Rajpurkar et al[1]. (2017 CheXNet, conducted utilizing DenseNet-121, achieved AUC of 0.848 for cardiomegaly, illustrating the effectiveness of the architecture for large-scale analysis of chest X-rays. However, other have pointed out in their studies, such as Nguyen et al. [8]. such that DenseNets compu-tational requirements can be limiting factor in real-time or resources/constrained environments. These challenges can be overcome by EfcientNet through introducing a compound scaling strategy which achieves a good balance of depth, width, and resolution in This enables the model to retain its levels of accuracy. Prior to the introduction of ,while utilizing far fewer computations, was able to resources that traditional CNNs. Research by Hashir et al.[7]. (2021) demonstrated outperformed other architectures such as ResNet and VGG in both accuracy and efciency, making it a strong potential candidate for clinical application. Gupta et al. additional inference times, which is especially very useful for real-time decision support in healthcare settings. Recent developments in the area have also focused on ensemble learning, attention mechanisms, mechanisms, and hybrid architectures. Rehman et al.[9]. (2022) demonstrated how the use of machine models such as Resnet and VGG to improve classication perfor-mance. Rubin etal,[6]. showed how attention mechanisms might improve feature localization, resulting in more elperar femen´ accurate predictions. Work by Pradeep et al[10 (2023) proposed Hybrid CNN-LSTM Models capable of leveraging contextual information, providing a novel outlook on car-diomegaly detection. Despite these achievements there are still challenges,particularly with respect to model generaliz-ability on various patient groups and robustness under different

   imaging conditions. Deep Learning has helped to solve some of these problems, as was highlighted by Baltruschat et al.[5]. found that pre-trained models generally perform better than randomly initialized models on medical datasets. Although these progress have been made, challenges existparticularly with respect to model generalizability across diverse patient populations and robustness under different imaging conditions. Deep Learning has helped to tackle some of these challenges, who found that there are often perform better on than models trained from scratch on medical datasets. In conclusion, the current body of work indicates a continued momentum toward automated cardiomegaly detection. Eff in particular, is special for its excellent balance of accuracy, computational efciency, and scalability. Future studies could focus on improving over-all generalization, reducing computational constraints, and the investigation into next-generation hybrid architecture to further enhance the diagnosis reliability.

3. Proposed Methodology

   In this paper, I will discuss a systematic approach to work-ing on an automated system for identication *Cardiomegaly using two deep learning models DenseNet and EfcentNet. include data preparation in the workow, Preprocessing, transfer learning, model training, and performance appraisal, all directed at building a reliable and clinically applicable solution.

   1. Dataset Acquisition and Preparation

      This research utilizes a selected subset of chest X-ray images. images obtained from publicly available repositories. Each image was labeled to * Cardiomegaly cases versus nor-mal instances cases, thereby facilitating equal representation during training and evaluation. Poor-quality, low-resolution images or images affected by errors were not taken into account to ensure the integrity of the dataset

   2. Preprocessing Pipeline

      To enhance model performance and ensure uniformity, several preprocessing steps were applied:

      • Resizing: All images were resized to 224 * 224 pixels.

      • Normalization: Pixel values were scaled to the 0-1 range for consistent input across the models.

      • Data Augmentation: Techniques such as horizontal ip-ping, small rotations, zoom variations, and brightness adjustments were incorporated. These steps increased data diversity and reduced the risk of overtting.

   3. Model Selection

      Two state-of-the-art CNN architectures were selected for evaluation:

      • DenseNet:It is a deep learning architecture used for image recognition. It connects each layer with every other layer within a dense block. It uses concatenation instead of summation.

      • EfcientNet: It is a family of highly accurate, lightweight convolutional neural network models designed for com-puter vision tasks like image classication.

   4. Transfer Learning

      Transfer learning was used for the transfer of models for the detection of cardiomegaly: The rst layers of the convolutional network were frozen to guarantee preservation of the abilities of feature extraction. To this network, a custom classication head was added on top. facilitating the model to learn task-related features relevant to cardiomegaly

   5. Model Training Conguration

      Model training was conducted with carefully tuned hyper-parameters:

      • Optimizer:Adam optimizer for adaptive learning rate adjustments.

      • Loss Function: Categorical cross-entropy for multi-class classication.

      • Batch Size: 32 or 64, chosen to balance memory usage and training stability.

      • Epochs:Models were trained for 30-50 epochs.

   6. Performance Evaluation

      The trained models were assessed using standard evaluation metrics:

      • Accuracy: Measures overall prediction correctness.

        TP + TN

   7. Comparative Analysis

      Comparisons were made between the DenseNet and Ef-cientNet using training sets, validation sets and test sets. Then with accuracy of diagnosis, computational efciency, and robustness through the imaging differences, a model was created that would be the most suitable for integration.

   8. Clinical Relevance

      The approach also focuses on the practical impact of automation systems on diagnostic variability and radiologist workload. Time-efcient networks such as EfcientNet par-ticularly useful in healthcare scenarios with limited resources. This structured and holistic approach uses recent methods from deep learning to practical constraints, contributing to the creation of an automated system for detection system that is bot accurate and scalable.

4. BACKGROUND AND DISCUSSION AND COMPARISION

   Cardiomegaly refers to an abnormal enlargement of the heart, and it is a signicant clinical sign of a range of heart conditions, which are the worlds leading cause of death. Early and accurate diagnosis of cardiomegaly is essential because it can help slow the progression of cardiovascular diseases and enhance the lives of patients. Chest X-rays are the primary imaging method used due to their low cost,

   Accuracy =

   TP + TN + FP + FN

   (1)

   availability and widespread use in clinical practice. But the interpretation of these images requires signicant radiologist

   • Precision and Recall: Evaluate the models capability to

     correctly identify true positive cardiomegaly cases.

     expertise, making it prone to human error, bias, and incon-sistencies in expertise. These issues underline the need for computerised diagnostic systems that can help doctors make

     TP

     Precision =

     TP + FP

     TP

     Recall =

     TP + FN

     (2)

     (3)

     quicker and more accurate diagnoses. The rise of articial intelligence technologies, especially deep learning models, has revolutionised the way medical images are interpreted. Convolutional Neural Networks (CNNs) have been particularly successful in this area as they are able to automatically

   • F1-Score: Provides a balanced measure of precision and recall.

     F 1 = 2 × Precision × Recall (4)

     Precision + Recall

   • Validation Loss: Used to track generalization perfor-mance and detect overtting during training.

   1 X

   M

   L = L(yval, yval) (5)

   learn complex features from image data, removing the need for manual feature extraction. Their efciency in handling large datasets and ability to achieve high accuracy makes them ideal candidates for medical image classication and detection, such as identifying cardiomegaly. In the world of CNN, DenseNet and EfcientNet have gained signicant at-tention for their effectiveness and design principles. DenseNet employs dense connections, with each layer connected to all previous layers. This enables effective feature reuse, enhances training efciency and enables the network to reach high

   where:

   val

   M i i

   i=1

   accuracy with a moderate number of parameters. However, the design can be computationally expensive for deep networks. EfcientNet, on the other hand, is more efcient. Its com-

   • TP = True Positives

   • TN = True Negatives

   • FP = False Positives

   • FN = Fale Negatives

   • M = number of validation samples

     i

   • yval = ground truth label

     i

   • yval = predicted label

   pound scaling method that proportionally scales the network width, depth, and resolution, achieving high accuracy with a manageable level of computational cost. EfcientNet is, therefore, well-adapted to resource-constrained environments. Previous studies have noted specic strengths of the mod-els in detecting cardiomegaly. DenseNets dense connectivity enables high diagnostic accuracy, but can be limited by its

   computational demands. EfcientNet, meanwhile, provides a scalable approach to performance metrics, and is thus suitable for clinical applications, particularly in resource-poor settings. Rajpurkar et al. found that DenseNet had an AUC of 0.848 for cardiomegaly detection, while Hashir et al. showed that EfcientNet was more accurate and efcient than ResNet and VGG for this task. These results indicate that both networks are capable of producing satisfactory diagnostic outcomes, but their application will depend on the computational and clinical setting for which they will be used. Overall, both DenseNet and EfcientNet can be used as reliable tools for automatic detection of cardiomegaly in chest X-rays. DenseNet is highly efcient in feature reuse and provides high accuracy, whereas EfcientNet is a lightweight alternative without compromising performance. The decision to choose one over the other should be determined by the needs of the clinical setting, such as computational resources and inference time.

5. Research Gaps

   While there are papers showing the high The po- gaps still exist to improve the acceptance of this technology is lagging in its acceptance is not catching on with the doctors and dentists. Another important area that needs to be explored is the performance of of these model performance on different data and cohorts. Many studies, such as Rajpurkar etal[1]. 2017, rely on specic datasets. that may not reect real-world variety in imaging imaging conditions, demographics, or prevalence. These would need to be more generalisable to to ensure models can be applied to all health systems. Another demand from deep learning efciency of deep learning models. Although It is efcient net, which aims to improve efciency, its compound scal- ound scaling method can increase computational density, hardware resources, making it challenging to deploy with limited hardware resources. Gupta etal[11] 2021 However, there is a demand for more lightweight approaches which dont sacrice and reducing the memory and processing re-quirements. Future research should be focused on the search to optimise These networks or to build composite models to the aim of catering to specic needs. for the low-resource clinical environment. Overtting is also a challenge. particularly for small and unbalanced data. datasets. This will severely impact model robustness and reducing generalization. when exposed to new data. Strategies such as sophisticated data augmentation methods such as increased Regularization; Better leveraging of transfer this process may be conducive to learning for over-tting, as observed. by Nguyen etal[8], 2020. Another challenge of crucial importance is false positives. and false negatives. A risk of false positives may cause unnecessary further up tests, while This may also lead to false negatives, which may result in delayed necessary treatment. Specicity and elaboration by classication thresholds that are major steps toward increasing Diagnostic reliability. Hashir etal[7] (2021) highlighted the need for architectural changes and better tuning to obtain clinically acceptable rates of accuracy. Finally, the key to integrating AI tools into Clinical practice is not just about tech- nical performance-it requires smooth

   workow compatibility. Rehman et al. say, The AI system should also provide work- work-ow compatibility. diagnosis without interfering diagnostic routines to make a diagnosis. Knowing how they can be integrated into the real world. healthcare settings is crucial. It can serve as a source of inspiration for this research. Taken together, these gaps suggest the need for further work to improve the The accuracy, speed and practicality of AI-based cardiomegaly detection systems.

6. Conclusion and Future Research Directions

This paper shows that the reality of using deep learning models, particularly DenseNet and EfcientNet, regarding the automated detection of cardiomegaly from chest X-ray images. Both models showed promising results, with Offering a more computationally efcient and accurate option for clinical use: application. The results indicate that this kind of AIdriven tool is capable of assisting radiologists through lowering diagnostic errors, speeding up decisions, and workow optimization – efciency, and especially environments. Paradoxically enough, despite these advances in robotics warrant further research. A major challenge is the generalization of these models with varying imaging conditions and the broader patient popula-tions. Such models are trained on small or limited uniform datasets could potentially fare poorly against applied to images from different institutions or regions. Large and diverse datasets, and techniques intended for the enhancement of resilience and generalizability to be feasible in practical – deployment Enhancing the efciency of computation is also considered a priority. Though EfcientNet is already optimized for computational speed, more to some extent, adapting these models for real-time applications in low-power conditions re-quires Additional architectural simplication or hybridization solutions that balance accuracy with efciency. A method of interpreting models is important because future work. Even though such models are able to perform tasks well, understand-ing why a prediction was made is a very good starting point is absolutely crucial for clinical trust and acceptance. Devel-oping a trustful relationship explainable AI (XAI) techniques that communicate model reasoning in a clear and clinically meaningful way will strengthen condence among healthcare professionals. Moreover, false positives and false negative results must also be reduced to ensure safe and reliable For diagnostic use: Fine-tuning model thresholds, class separation, as well as improving specicity through architectural changes can however, assist in resolving this situation. In Conclusion In summary, DenseNet and EfcientNet demonstrate a lot of promise in detecting cardiomegaly, future work needs to aim at improving Generalization, computational efciency, inter-pretability, and diagnostic accuracy. These challenges will thus help create robust AIdriven systems for diagnostics: Scalable and very benecial in modern healthcare environments.

References

1. Rajpurkar, P., etal[1] (2017). CheXNet: Radiologist-Level Pneumonia Detection on Chest X-Rays with Deep Learning.

2. J. Irvin, J., etal[2] (2019). CheXpert: A Large Chest Radiograph Dataset with Uncertainty Labels and Expert Comparison..

3. Wang, X., etal[3] (2018). ChestX-ray8: Hospital-Scale Chest X-ray Database and Benchmarks on Weakly-Supervised Classication and Localization of Common Thorax Diseases.

4. Yan, C., etal[4] (2020). Medical Image Classication with Ensemble Learning and Data Augmentation.

5. Baltruschat, I. M., etal[5] (2019). Comparison of Deep Learning Ap-proaches for Multi-label Chest X-ray Classication.

6. Rubin, J., etal[6] (2018). Visualizing Anatomical Patterns in Chest Radiographs with Attention Mechanisms.

7. Hashir, M., etal[7] (2021). EfcientNet for Cardiomegaly Detection: An Evaluation on Public Chet X-ray Datasets.

8. Nguyen, H., etal[8] (2020). DenseNet Architectures for Medical Image Diagnosis

9. Rehman, Z., etal[9] (2022). Improving Cardiomegaly Detection through Ensemble Learning Techniques with Attention Mechanisms.

10. Pradeep, R., etal[10] (2023). Combining CNN and LSTM for Car-diomegaly Prediction in Chest X-rays

11. Gupta, A., etal[11] (2021). Lightweight Architectures for Medical Image Analysis: A Study on Cardiomegaly.

12. Khan, S., etal[12] (2023). Vision Transformers for Medical Image Analysis: A Case Study on Chest X-rays

13. Simonyan, Karen, and Andrew Zisserman., etal[13] (2013). Very deep convolutional networks for large-scale image recognition.

14. Mrigakshi Gupta, Akash Singh and Yatender Kumar., etal[14](2024) Deep learning for prediction of cardiomegaly using chest X-rays.

15. Lee, Kang-Hee, Jun-Woo Choi, Chun-Oh Park, Dong-Hun Han, and Min-Soo Kang. A Development and Validation of an AI Model for Cardiomegaly Detection in Chest X-rays ., etal[15](2024)