Determination of Anteroposterior and Posteroanterior Imaging Positions on Chest X-Ray Images Using Deep Learning ^†

Abstract

This study proposes a deep learning framework to classify anteroposterior (AP) and posteroanterior (PA) chest X-ray projections automatically. Multiple convolutional neural networks (CNNs), including ResNet18, ResNet34, ResNet50, DenseNet121, EfficientNetV2-S, and ConvNeXt-Tiny, were utilized. The NIH Chest X-ray Dataset, with 112,120 images, was used with strict patient-wise splitting to prevent data leakage. ResNet34 achieved the highest performance: 99.65% accuracy, 0.9956 F1 score, and 0.9994 ROC-AUC. Grad-CAM visualized model decisions, and expert-reviewed misclassified samples were removed to enhance dataset quality. This methodology highlights the importance of robust preprocessing, model interpretability, and clinical applicability in radiographic view classification tasks.

1. Introduction

In recent years, deep learning-based approaches have made significant progress in the field of medicine and have become widely used, especially in radiological image analysis. More than 1 billion X-ray imaging protocols are performed worldwide each year, the majority of which are chest X-rays [1]. The correct classification of image acquisition positions is of great importance for the effective use of chest X-rays in diagnostic processes. Differences between anteroposterior (AP) and posteroanterior (PA) projections can lead to differences in the size and shape of the heart and mediastinal structures, which can increase the risk of misdiagnosis and evaluation [2]. Therefore, the automatic classification of AP and PA images and detection of mislabeled data play a critical role in increasing the accuracy of clinical decision support systems [1,3,4].

Various studies have been conducted on PA/AP classification in the literature [1,3]. For example, AP and PA images were classified in a deep learning model developed by Kim et al. (2019) using the NIH ChestX-ray14 dataset [3]. However, it was not clearly stated in this study whether patient-based data separation was performed. If different images belonging to the same patient were distributed to the training and test sets, the model might have learned patient-based features and provided an artificial success increase. This may reduce the generalization performance of the model in the real world.

In this study, patient-based data separation was carefully performed to eliminate deficiencies in the literature and increase the clinical applicability of the model. Data leakage was prevented by ensuring that images belonging to the same patient were assigned to only one dataset (training, validation, or testing). Thus, the model was evaluated on patients it had never encountered before and was tested in a way that was more suitable for real clinical scenarios.

In addition, age is known to be an important factor when determining the AP and PA imaging positions. In pediatric patients (especially those under 5 years of age), the AP imaging technique is used more frequently because young children are usually unable to stand and may have difficulty following breath-holding instructions. In this study, unlike previous studies, pediatric patients (under 18 years of age) in the dataset were analyzed without separation. The generalization ability of the model was increased by ensuring consistent performance across different age groups.

To analyze the performance of the model, the images with the highest loss values were determined in the validation dataset. These images were evaluated by an expert radiologist, and examples containing mislabeling, acquisition errors, or technical problems were removed from the dataset. In addition, images that the model had difficulty with but were medically meaningful were included in the training process. This process was performed to increase the generalization ability of the model and improve its clinical applicability.

Our proposed model differs from previous studies in terms of not only patient-based data separation but also performance evaluation methods. The success of our model, accuracy, error rate, F1 score, Cohen’s kappa, recall, specificity, precision, Brier score, balanced accuracy, and ROC-AUC were analyzed comprehensively using various metrics. This detailed evaluation demonstrates the reliability and clinical applicability of our model for PA/AP classification more strongly than in previous studies. The main objective of this study was to develop a reliable deep learning model that can accurately classify the shooting positions in chest X-rays and detect mislabeled or incorrectly captured images. With the proposed methodology, data quality is increased, the generalization ability of the model is improved, and a more reliable contribution is made to clinical decision support systems.

2. Method

2.1. Dataset

In this study, the NIH Chest X-ray Dataset [4] was used for AP and PA position detection. This dataset contained 112,120 frontal chest X-ray images from 30,805 unique patients. Each image had labels associated with 14 thoracic diseases.

All images in the dataset had 1024 × 1024 pixels, grayscale, and 8-bit color depth. In addition, the resolution of the images was specified to be 96 dpi.

This dataset was publicly shared and did not require special permission for use. In this study, data were obtained by observing patient privacy and ethical rules [5].

The number of patients in this dataset and their distributions are listed in

Table 1. Distribution of patients in the dataset used.

Feature	Value
Total Number of Images	112,120
Number of PA Images (%)	67,310 (60.1%)
Number of AP Images (%)	44,810 (39.9%)
Number of Male Patient Images	63,340
Number of Female Patient Images	48,780
Number of Unique Patients	30,805
Average Number of Images per Patient	3.64

.

When the dataset was examined, it was observed that there was a total of 112,120 images, with 67,310 (60.1%) in the PA position and 44,810 (39.9%) in the AP position. In binary classification problems, class balance is an important factor that affects the learning performance of a model. When the class ratios (60.1–39.9%) in this dataset are considered, it is seen that although it is not completely balanced (homogeneous), it does not contain any serious imbalance. Therefore, metrics such as precision, recall, and F1 score were taken into account when examining the results. When the dataset was examined, it was determined that more than one image belonged to a patient. It was observed that there was more than one PA or AP image belonging to the same patient, as well as images taken in both the PA and AP positions belonging to the same patient. The distribution of the number of images per patient is presented graphically in Figure 1.

The distribution of patient images with the most images is presented graphically in Figure 2. The colors in the bars represent a sequential color gradient to visually distinguish patients and do not correspond to any additional variable. Such repeated images require attention to the principle of patient-based independence during training. Splitting images belonging to the same patient between the training and test datasets is a critical step in reducing the risk of overfitting the model and increasing its generalizability to the real world. In addition, the presence of more than one image belonging to a patient in the dataset can lead to the risk of data leakage. This is especially true if the test dataset is included in the training dataset or if the test data are indirectly learned by the model during the training process. Data leakage can cause the model to overfit, and the results are inconsistent with the real-world performance. In addition, using more than one image belonging to the same patient can cause the model to “see” the test dataset, misleading its true performance. Although there were 112,120 images in the dataset used in this study, the number of patients with a single image was 17,503. The number of images belonging to a patient reached 184, indicating that data leakage should be carefully considered.

Sample radiographs of the AP and PA projections are shown in Figure 3. These images allow for visual comparison of the significant anatomical differences between the two viewing positions.

2.2. Fastai

Fastai is defined as a library that enables fast and easy development of deep learning models [6]. This library provides high-level components for practitioners and flexible and modular low-level components for researchers. Aiming to strike a balance between ease of use, flexibility, and performance, Fastai (version 2.8.2) has an architecture that optimizes deep learning and data processing by leveraging Python’s dynamism (version 3.12.8) and PyTorch’s flexibility (version 2.5.1).

2.3. Classification Studies

In this study, a computer with an i7 10th generation processor, an Nvidia RTX 2060 6 GB graphics card, and 16 GB RAM was used.

As mentioned above, considering the risk of data leakage, images from the same patient were included in only one dataset (training, validation, or test). The dataset was randomly divided into three subsets: 70% training (78,566 images), 15% validation (17,063 images), and 15% testing (16,491 images). In addition, the distribution of AP and PA positions was maintained at 40–60% in each dataset. This helped minimize the imbalance in the dataset so that the model could provide more generalizable and accurate results.

In this study, ResNet18, ResNet34, ResNet50, DenseNet121, EfficientNetV2-S, and Con-vNeXt-Tiny models were applied for classification using the Fastai library. During the training process, the models resized the images from 1024 to 224 pixels. Data augmentation techniques were used during the training and testing stages. However, these techniques were carefully selected based on the structure of radiological images. In particular, methods such as horizontal and vertical flipping were not applied because the images were contrary to body symmetry, and it was not meaningful to flip an X-ray image. Instead, more limited data augmentation techniques, such as slight rotation (maximum 5°), slight zoom (maximum 10%), and slight pan (maximum 10%), were used. Fastai randomly applied these data augmentation methods in each training stage, which allowed the model to encounter a wider variety of images and increased its generalization ability. In addition, considering the class imbalances in the dataset, such careful data augmentation techniques allow for more robust and generalizable models to be obtained.

In this study, the employed deep learning architectures and the resulting performance metrics are presented in

Table 2. AP/PA classification performance of different CNN architectures without data augmentation.

Metric	EfficientNetV2-S	ConvNeXt-Tiny	DenseNet-121	ResNet-18	ResNet-34	ResNet-50
Train Loss	0.016856	0.033227	0.024416	0.000312	0.000239	0.013660
Valid Loss	0.016391	0.021557	0.017844	0.030812	0.023618	0.023737
Error Rate	0.003985	0.005157	0.004395	0.004044	0.003458	0.004747
Accuracy	0.996015	0.994843	0.995605	0.995956	0.996542	0.995253
F1 Score	0.996668	0.995690	0.996325	0.996620	0.997109	0.996035
Cohen’s Kappa	0.991711	0.989271	0.990857	0.991588	0.992808	0.990121
Recall	0.997646	0.996959	0.997254	0.997842	0.998038	0.997940
Precision	0.995693	0.994423	0.995398	0.995401	0.996182	0.994137
Brier Score Loss	0.003985	0.005157	0.004395	0.004044	0.003458	0.004747
Balanced Accuracy	0.995620	0.994330	0.995205	0.995499	0.996180	0.994602
ROC-AUC	0.999510	0.998567	0.999384	0.999151	0.999552	0.999357

. Multiple convolutional neural network (CNN) models—including ResNet18, ResNet34, ResNet50, DenseNet121, EfficientNetV2-S, and ConvNeXt-Tiny—were evaluated for the task of distinguishing AP and PA chest X-ray projections.

Each model was trained under two distinct conditions.

Table 3. AP/PA classification performance of different CNN architectures with data augmentation.

Metric	EfficientNetV2-S	ConvNeXt-Tiny	DenseNet-121	ResNet-18	ResNet-34	ResNet-50
Train Loss	0.024058	0.012695	0.009012	0.027934	0.022322	0.025057
Valid Loss	0.017230	0.020557	0.017776	0.021910	0.014919	0.022696
Error Rate	0.004220	0.004806	0.004161	0.005333	0.003751	0.005392
Accuracy	0.995780	0.995194	0.995839	0.994667	0.996249	0.994608
F1 Score	0.996473	0.995982	0.996520	0.995542	0.996863	0.995496
Cohen’s Kappa	0.991223	0.990004	0.991347	0.988906	0.992199	0.988781
Recall	0.997548	0.996959	0.997057	0.996665	0.997548	0.997254
Precision	0.995400	0.995007	0.995983	0.994422	0.996180	0.993745
Brier Score Loss	0.004220	0.004806	0.004161	0.005333	0.003751	0.005392
Balanced Accuracy	0.995352	0.994767	0.995544	0.994183	0.995935	0.993967
ROC-AUC	0.999479	0.998876	0.999429	0.999303	0.999726	0.998885

summarizes the results obtained with data augmentation techniques such as slight rotation, zoom, and pan, while Table 2 presents the outcomes without any augmentation. Key classification metrics—accuracy, error rate, F1 score, Cohen’s kappa score, and the area under the receiver operating characteristic curve (ROC-AUC)—were used to comprehensively evaluate model performance.

Multiple metrics were used to evaluate the model success. Accuracy indicates the overall success rate of the model, whereas the error rate indicates the percentage of incorrectly predicted examples. However, because the dataset is not fully balanced and may contain some class imbalance, the balanced accuracy score metric, which evaluates the accuracy of both classes by assigning equal weights, was also used.

Additional metrics were also used to evaluate the classification performance more comprehensively. Recall/sensitivity measures the rate of correct prediction of the positive class, while precision shows the accuracy of the model’s positive predictions. The F1 score, which provides a balance between these two metrics, was calculated to determine whether the model exhibited balanced performance in terms of both sensitivity and precision.

In addition, Cohen’s kappa score, which measures how successful the model is compared to random prediction, was used as an important performance indicator in cases of data imbalance. Finally, the ROC-AUC score is a critical measure that shows how well the model can distinguish between positive and negative classes at different decision thresholds. This metric was analyzed to evaluate the overall discrimination power of the model.

The validation loss was examined to evaluate the success of the model using the validation data. A decrease in the validation loss during the training process may indicate that the generalization ability of the model has increased, and an increase after a certain point may indicate signs of overfitting. In this context, while the validation loss in models without data augmentation remained at higher levels, particularly for ResNet50, it was observed that the validation loss decreased in models with data augmentation. This indicates that data augmentation improves the generalization ability of the model and reduces the risk of overfitting.

In addition, the Brier Score Loss metric was used to evaluate the reliability of the probabilistic estimates of the model. This metric is calculated by taking the average of the square of the difference between the probabilities predicted by the model and actual class labels. Lower Brier Score Loss values indicate that the model provides better-calibrated and reliable probability estimates.

In this study, the effects of different deep learning models and data augmentation methods on Brier Score Loss were analyzed. It was observed that the Brier Score Loss values decreased in most of the models with data augmentation, indicating that the model produced more reliable probability estimates. When this result was evaluated together with the ROC-AUC and other performance metrics, it was revealed that the model provided better-calibrated probabilities in the decision-making process.

In the experiments, the performances of used models on the training, validation, and test data were analyzed. The overall success of the models was compared by applying different data augmentation scenarios.

The performance results of the developed models for the test dataset are presented in

Table 4. Test results.

Model	Aug.	TP	FP	FN	TN	Accuracy	Precision	Recall	F1 Score	ROC-AUC
EfficientNetV2-S	No	6569	35	38	9849	0.995573	0.995574	0.995573	0.995574	0.999037
	Yes	6569	35	41	9846	0.995391	0.995392	0.995391	0.995392	0.998984
ConvNeXt-Tiny	No	6564	40	38	9849	0.995270	0.995270	0.995270	0.995270	0.998387
	Yes	6567	37	41	9846	0.995270	0.995271	0.995270	0.995270	0.998546
DenseNet121	No	6568	36	35	9852	0.995695	0.995695	0.995695	0.995695	0.999136
	Yes	6570	34	42	9845	0.995391	0.995393	0.995391	0.995392	0.999081
ResNet18	No	6574	30	33	9854	0.996180	0.995457	0.995005	0.995231	0.995985
	Yes	6564	40	32	9855	0.995634	0.993943	0.995149	0.994545	0.995553
ResNet34	No	6575	29	32	9855	0.996301	0.995609	0.995157	0.995383	0.996111
	Yes	6577	27	38	9849	0.996058	0.995912	0.994255	0.995083	0.995761
ResNet50	No	6573	31	33	9854	0.996119	0.995306	0.995005	0.995155	0.995934
	Yes	6562	42	29	9858	0.995695	0.993640	0.995600	0.994619	0.995679

. In the evaluation, AP was considered the correct class, whereas misclassified PA was regarded as the incorrect class (False PA).

In the experimental studies, the performance of deep learning architectures used for the AP/PA orientation classification task has been analyzed comprehensively. The validation data results presented in Table 2 and Table 3 and the test data results presented in Table 4 show that all models achieved very high and close success levels above 99% in metrics such as accuracy, ROC-AUC, and F1 score. This confirms that all examined architectures have high competence in the mentioned classification task.

However, when a detailed comparison of the metrics is made, the ResNet34 model stands out as the architecture that exhibits the most balanced and consistent performance. The model achieved an ROC-AUC score of 0.999726, an F1 score of 0.996863, and an accuracy rate of 0.996249 on the test data. One of the most important findings supporting this high success is that the model recorded the lowest error rate among all the compared models with a value of 0.003751 in the data augmentation scenario. This extremely low error rate shows that the model has both high overall classification success and is more resistant to erroneous predictions compared to other models. In addition, the low valid loss of 0.014919 obtained during the training process is another important finding that supports the strength of the model’s generalization ability and its resistance to overfitting.

In addition to this consistent success in numerical metrics, practical advantages were also decisive in choosing the ResNet34 architecture. Having fewer parameters compared to deeper and more complex models such as ResNet50 provides significant efficiency advantages such as faster training times and lower computational cost. The lighter structure of the model reduces the risk of overfitting, while facilitating the applicability and deployment of the model, especially in clinical environments or on limited hardware resources.

In conclusion, although all tested models showed high performance, the ResNet34 model was determined to be the most suitable model to be used in this study due to its consistency in numerical metrics, the lowest error rate, and the optimal balance between the computational efficiency it offers. In this regard, the ResNet34 model was taken as basis in all subsequent analysis and evaluation stages of the project.

The images with the highest loss values in the validation phase of the model and examples that were incorrectly predicted in the test data were examined by an expert radiologist. The images that were truly faulty or problematic for the model were removed from the dataset; however, the images that the model had difficulty with but were medically meaningful and correct were included in the training process. The sample images extracted from the dataset are shown in Figure 4.

Figure 4a–i present sample images that were misclassified by the model and subsequently removed from the dataset. Below each image, the predicted class label, the true class label, the loss value, and the prediction probability are shown. For instance, in Figure 4a, the model predicted the image as “PA”, whereas the true label is “AP”, with a loss value of 7.18 and a prediction probability of 1.00. These images were analyzed to better understand the challenges in the model’s classification performance and the potential causes of misclassification.

In this study, the binary cross-entropy loss function provided by the Fastai library was employed for binary classification tasks. The loss values were calculated using the following formula:

$$L\left(y,\widehat{y}\right)=-{\displaystyle \frac{1}{N}}\sum _{i=1}^N\left[y_i\mathrm{l}\mathrm{o}\mathrm{g}\right(\sigma \left({\widehat{y}}_i\right))+\left(1-y_i\right)\mathrm{l}\mathrm{o}\mathrm{g}(1-\sigma \left({\widehat{y}}_i\right)\left)\right]$$

(1)

Here, N denotes the total number of samples, $y_i$ ∈ {0, 1} represents the true class label, ${\widehat{y}}_i$ is the logit output produced by the model, σ(${\widehat{y}}_i$) = 1/(1 + e^(−${\widehat{y}}_i$)) refers to the sigmoid activation function, and L($y,\widehat{y}$) denotes the average binary cross-entropy loss. This loss function enhances the sensitivity of the model to misclassifications, thereby enabling it to produce more reliable and robust predictions.

After the removal of the misclassified samples from the dataset, the model was retrained, resulting in a noticeable improvement in the performance metrics. Evaluations conducted on the test dataset showed a statistically significant enhancement in the model performance. The training and test results are listed in

Table 5. Classification results for the AP/PA validation dataset using the ResNet34 model.

Metric	Values
Train Loss	0.009089
Valid Loss	0.004667
Error Rate	0.001176
Accuracy	0.998824
F1 Score	0.999018
Cohen’s Kappa	0.997554
Recall	0.999607
Precision	0.998430
Brier Score Loss	0.998430
Balanced Accuracy	0.998633
ROC-AUC	0.999977

and

Table 6. Test results of the developed model on the test dataset (True AP, False PA accepted).

TP	FP	FN	TN	Accuracy	Precision	Recall	F1 Score	ROC-AUC
6575	29	29	9858	0.99649	0.99561	0.99561	0.99561	0.99935

, respectively.

To enhance the interpretability of the proposed CNN-based model, Gradient-weighted Class Activation Mapping (Grad-CAM) was employed to visualize class-discriminative regions in selected chest X-ray images.

Figure 5 illustrates two representative cases, each corresponding to an AP and PA projection, respectively. In the AP-predicted image (Figure 5a), the model focuses significantly on the lower right thoracic region and the presence of medical devices (e.g., ECG leads or a port on the shoulder), which are frequently associated with supine positioning and thus AP projection. This localization suggests the model leverages clinically relevant visual cues to distinguish the view.

Conversely, in the PA-predicted image (Figure 5b), the attention is primarily distributed over the central thoracic cavity, particularly the heart and lung hilum areas. This pattern is consistent with standard upright PA positioning, where the anatomical structures are more uniformly exposed and less obstructed by external artifacts.

These Grad-CAM results confirm that the model not only achieves high classification accuracy but also bases its decisions on plausible anatomical and contextual features, thereby supporting its reliability and potential clinical applicability.

3. Conclusions

In this study, the ResNet34 deep learning architecture demonstrated high performance in classifying AP and PA chest X-ray positions. Patient-based data separation and carefully selected data augmentation techniques enhanced the generalization ability of the model, supporting its applicability in clinical settings. The high-performance metrics, 99.65% accuracy and 0.99935 ROC-AUC, indicate that the model can serve as a reliable aid in diagnostic processes.

The exclusion of misclassified images following expert radiologist reviews highlights the critical importance of data quality in model performance. However, further evaluation of the model using external datasets and multicenter studies is recommended to better assess its generalizability. Future studies may also focus on incorporating factors, such as image quality and patient demographics, to further enhance the model’s clinical utility.

This study provides a solid foundation for the integration of automated classification systems into routine radiological practices.