IAES Inter national J our nal of Articial Intelligence (IJ-AI) V ol. 15, No. 3, June 2026, pp. 2041 ∼ 2052 ISSN: 2252-8938, DOI: 10.11591/ijai.v15.i3.pp2041-2052 ❒ 2041 Pneumothorax detection using a lear ning f ocal point ar chitectur e Salah-Eddine Mansour 1 , Bouabid Qabliyane 2 , Abdelhak Sakhi 3 , Zakaria Khoudi 4 , Mohamed Baslam 2 1 LaRI Laboratory , F aculty of Sciences Oujda, Mohammed I Uni v ersity , Oujda, Morocco 2 TIAD Laboratory , F aculty of Sciences and T echnology , Sultan Moulay Slimane Uni v ersity , Beni Mellal, Morocco 3 Department of Computer Science, F aculty of Sciences Ain Chock, Hassan II Uni v ersity , Casablanca, Morocco 4 LM2I, F aculty of Sciences and T echnology , Sultan Moulay Slimane Uni v ersity , Beni Mellal, Morocco Article Inf o Article history: Recei v ed Sep 16, 2024 Re vised Apr 29, 2026 Accepted May 11, 2026 K eyw ords: Con v olutional neural netw ork Deep learning Learning focal point algorithm Perceptron Pneumothorax ABSTRA CT Automatic image se gmentation and feature analysis play a crucial role in impro ving the accurac y and ef cienc y of disease diagnosis and treatment within modern medical practice. This study propose the use of t he learning focal point (LFP) architecture, which is based on the LFP algor ithm, to perform ef fecti v e se gmentation of medical images by di viding each image in the dataset into multiple meaningful zones. This zonal se gmentation strate gy enables the precise e xtraction of critical re gions of interest that are most rele v ant for pathological analysis. The proposed approach is specically applied to the detection of common pneumothorax in lung imaging, a condition tha t requires timely and accurate diagnosis. By concentrating on essential lung zones, the LFP architecture enhances the reliability and rob ustness of pneumothorax identication. The results demonstrat e that this method has the potential to signicantly assist clinicians by pro viding more accurate diagnostic support and f acilitating earlier medical interv ention, ultimately impro ving patient outcomes. This is an open access article under the CC BY -SA license . Corresponding A uthor: Salah-Eddine Mansour LaRI Laboratory , F aculty of Sciences Oujda, Mohammed I Uni v ersity Mohammed VI P alace, P .O. Box 524, Oujda 60000, Morocco Email: s.mansour@ump.ac.ma 1. INTR ODUCTION Medical imaging refers to the nonin v asi v e acquisition and processing if internal tissue images of the human body or specic body parts, essential for medical diagnosis and research purposes. This eld emplo ys in v erse computational methods to generate images crucial for v arious medical disciplines including diagnostic imaging, radiology , endoscop y , medical thermograph y , medical photograph y , and microcop y [1]. Additionally , techniques such as brainw a v e mapping and magnetoencephalograph y , despite primarily focusing on measurement and data recording rather that visual imagery , pro vide v aluable positional information that can be cate gorized as a distinct form of medical imaging due to its inherent locational characteristics. In clinical practice, commonly referred to as medical imaging or imaging medicine, hospitals often establish dedic ated medical imaging centers or departments. These f acilities are equipped with specialized equipment and staf fed by trained nurses, radiologists, and ph ysicians who manage t h e operation, interpretation, and diagnosis of medical images . This role dif fers signicantly from radiation therap y used in radiology , focusing on precise imaging and diagnostic aspects essential for medical assessments and treatments [2]. J ournal homepage: http://ijai.iaescor e .com Evaluation Warning : The document was created with Spire.PDF for Python.
2042 ❒ ISSN: 2252-8938 In the realms of medical science, medical engineering, medica l ph ysics, and biomedi cal information science, the term “medical imaging” generally pertains to the scientic e xplorati on and adv ancement of imaging technologies, including the de v elopment of i mage capture, storage, and processing methods. Research endea v ors in interpreting and diagnosing medical images represent a complementary eld within specialties such as radiology , neurology , and cardio v ascular diseases [3]. In this study , chest X-ray images serv e as the primary dataset for identifying v arious lung-related health issues. Specically , our approach in v olv es utilizing the learning focal point (LFP) architecture as the initial method for classifying pneumothorax. LFP ret rie v es the most important di visions within each image by treating patches, whereas con v olutional neur al netw ork (CNN) processes indi vidual pix els. Therefore, the quality of the dataset signicant ly inuences the learning process. W e conducted a comparati v e analysis to e v aluate the ef fecti v eness and ef cienc y of both architectures in accurately detecting and classifying pneumothorax from the chest X-ray images [4]. Using the LFP architecture, we ha v e achie v ed a precision rate of 87%. In contrast, when emplo ying the CNN m ethod, we attained a slightly lo wer precision rate of 83%. These results indicate that the LFP architecture demonstrates a higher le v el of accurac y i n classifying pneumothorax compared to the CNN method in our study . The 4% dif ference in precision suggests that the LFP architecture may of fer certain adv antages or be better suited for this specic classication task using chest X-ray images. In this paper , section 2 pro vides a comprehensi v e re vie w of related w orks in pneumothorax detection using chest X-ray images. Section 3 then details the implementation and comparison of tw o distinct architectures: the LFP architecture, which focuses on e xtracting signicant image patches, and CNN, which operates at the pix el le v el. These architectures are e v aluated using a pneumothorax dataset to assess their ef fecti v eness in classication. Section 4 discusses and compares the results and performance metrics obtained from both methods, including precision, recall, and accurac y . This comparati v e analysis aims to highlight the strengths and weaknesses of each approach in detecting pneumothorax. Finally , section 5 concludes by summarizing the ndings, discussing implications for medical image analysis, and suggesting directions for future research in enhancing pneumothorax detection methodologies. 2. RELA TED W ORKS Articial intelligence (AI) has indeed been le v eraged in v arious w ays to aid in the detection of pneumothorax. Pneumothorax is a condition where air l eaks into the space between the lungs and chest w all, leading to lung collapse. AI techniques ha v e been applied to medical imaging, particularly in analyzing chest X-rays and computed tomograph y (CT) scans, to assist radiologists in identifying pneumothorax accurately and swiftly [5]. In recent research, there has been a propos al for a sophisticated computerized system le v eraging deep learning techniques. This system utilizes a Mask re gion-based con v olutional neural netw ork (Mask RCNN) frame w ork inte grated with ResNet101 as a feature p yramid netw ork, specically designed to enhance the detection capabilities for medical imaging tasks. T rained meticulously on a dataset sourced from Society for Imaging Informatics in Medicine–American Colle ge of Radiology (SIIM-A CR), the system sho wcased signicant adv ancements o v er pre vious methods. In direct comparison with a ResNet50-based approach, the proposed system demonstrated superior performance metrics, e xhibiting notably reduced class loss, bounding box loss, and mask loss during e v aluation stages. The ef fecti v eness of the models w as further scrutinized through rigorous testing in v olving distinct learning rates, specically 0.0004 and 0.0006, implemented o v er v arying epochs - 10 epochs for one conguration and 12 epochs for another . This comprehensi v e e v aluation underscores the system’ s potential to re v olutionize the detection and analysis of comple x medical conditions, promising enhanced accurac y and ef cienc y in clinical settings [6]. In another research study , support v ector machine (SVM) is applied as a pi v otal tool for the identication of pneumothorax. This in v olv es e xtracting detailed features from lung images usi ng local binary pattern (LBP), a technique kno wn for its ef fecti v eness in capturing local te xture information. The e xtracted features are then utilized to train the SVM model, enabling it to classify whether pneumothorax is present or not based on the pattern of features identied [7]. Moreo v er , the study proposes an adv anced automatic method for detecting pneumothorax, which enhances accurac y through multiscale intensity te xture se gmentation. This inno v ati v e approach focuses on remo ving background noise and rening the se gmentation of abnormal lung re gions within chest images. Int J Artif Intell, V ol. 15, No. 3, June 2026: 2041–2052 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Artif Intell ISSN: 2252-8938 ❒ 2043 The process be gins with meticulously s e gme nting out the re gions of interest, le v eraging te xture analysis computed o v er multiple o v erlapping blocks to ensure comprehensi v e co v erage [8]. Furthermore, to rene the se gmentation process and delineate the boundaries more accurately , t he study emplo ys Sobel edge detection. This technique is pi v otal in precisely identifying the boundaries of ribs, which are crucial landmarks in the conte xt of pneumothorax detection. By ef fecti v ely locating these edges, the study enhances the accurac y of se gmenting abnormal lung re gions [9]. In addition to addressing the critical need for rapid and accurate pneumothorax diagnosis using frontalvie w chest X-ray images, crucial in emer genc y settings, other study emphasizes the limitations of manual radiograph re vie w , which is time-consuming and prone to human error . The research proposes a tw o-stage deep learning approach based on ResNet, inte grating local feature learning (LFL) and global multi-instance learning (GMIL). This method is designed to prioritize discriminati v e features while e xcluding non-lesion re gions, thereby enhancing diagnostic accurac y . V alidation is performed rigorously using tw o e xtensi v e datasets: a pri v at e dataset comprising 27,955 images and the NIH ChestX-ray14 dataset containing 112,120 images. The model under goes comprehensi v e e v al uation with metrics such as accurac y , precision, recall, specicity , F1-score, recei v er operating characteristic (R OC), and area under the curv e (A UC), through rob ust v efold cross-v alidation. Res ults on the NIH dataset demonstrate state-of-the-art performance, achie ving notable metrics including accurac y of 94.4% ± 0.7%, A UC of 97.3% ± 0.5%, precision of 94.2% ± 0.3%, recall of 94.6% ± 1.5%, specicity of 94.2% ± 0.4%, and F1-score of 94.4% ± 0.7%. These ndings underscore the ef cac y of the proposed deep learning approach in signicantly impro ving diagnostic ef cienc y and accurac y for pneumothorax detection in clinical practice, promising substantial benets for patient care and emer genc y management [10]. Another study aimed to e v aluate the diagnostic performance of fully-connected small articial neural netw orks (ANNs) trained using the Kim-Monte Carlo algorithm for localizing pneumothorax in chest X-rays. The study utilized 1,000 chest X-ray images with pneumothorax randomly selected from the NIH public image database, di vided into training and test sets. Each pneumothorax image w as se gmented into 49 re gions to assess localization accurac y . The ANN achie v ed an impressi v e A UC of 0.882 on the test set, with corresponding sensiti vity and specici ty rates of 80.6% and 83.0%, respecti v ely . Furthermore, the study compared the ANN’ s performance with that of a CNN, a widely used deep-learning method, on the same dataset. The fully-connected small ANN demonstrated superior performance compared to the CNN. Notably , among the CNN models tested with dif ferent acti v ation functions, the sigmoid acti v ation function for fully-connected hidden nodes sho wed the best results, surpassing the rectied linear unit (ReLU) acti v at ion function. This research underscores the potential of the proposed approach in accurately localizing pneumothorax in chest X-rays, thereby mitig ating diagnostic delays in critical conditions and enhancing o v erall clinical ef cac y and patient care [11]. Based on the ndings from pre vious research, it is e vident t hat signicant adv ancements ha v e been made in the realm of precision and accurac y , particularly in the domain of medical imaging and diagnostic systems. Ho we v er , b uilding upon these achie v ements, our forthcoming research aims to introduce and implement the LFP algorithm to further enhance performance be yond e xisting methodologies. The LFP algorithm is poised to augment the current capabilities observ ed in recent studies by optimizing feature e xtraction and inte gration processes, thereby potentially rening the detection and classication accurac y of medical conditions. By incorporating LFP into our methodology , we aspire to contrib ute no v el insights and adv ancements that surpass the benchmarks set by prior approaches, ultimately adv ancing the state-of-the-art in medical image analysis and diagnostic accurac y [12]. 3. METHOD 3.1. Dataset description The dataset for pneumothorax consists of a modest collection comprising 2,027 images, speci cally curated for a binary classication task focused on distinguishing between images that depict pneumothorax and those that do not. This data set serv es as a foundational resource for training and e v aluating machine learning models task ed with automated detection and diagnosis of pneumothorax from medical images, such as chest X-rays or CT scans. The binary nature of the classication task simplies the learning process, aiming to accurately identify the presence or absence of pneumothorax with high precision and recall. Such datasets are essential in de v eloping rob ust algorithms that can potentially aid healthcare professionals in making timely and accurate clinical decisions, thereby impro ving patient care outcomes in medical settings [13]. Pneumothor ax detection using a learning focal point ar c hitectur e (Salah-Eddine Mansour) Evaluation Warning : The document was created with Spire.PDF for Python.
2044 ❒ ISSN: 2252-8938 Re g arding the tar get labels in our dataset, where a v alue of 1 indicates the presence of the disease (pneumothorax), and 0 signies its absence, we observ e a distrib ution of 430 images labeled as tar get 0 (no pneumothorax) and 1,597 images labeled as tar get 1 (sho wing pn e umothorax). This class distrib ution underscores the imbalance between the tw o cate gories, with a predominance of images depicting pneumothorax. This distrib ution is crucial for training machine learning models ef fecti v ely , ensuring the y can properly learn and dif ferentiate between images with and without pneumothorax. Addressing this class imbalance will be critical in de v eloping a rob ust and accurate classication [14]. 3.2. Model ar chitectur e This paper utilized the LFP architecture, which is designed to enhance the ef cienc y and accurac y of data processing tasks. The architecture comprises three fundamental components: i) the LFP algorithm, which serv es as the foundational frame w ork for data e xtraction and manipulation; ii) the selector , which aids in the precise identication and selection of rele v ant data subsets; and ii i) a dedicated block of neural netw orks, emplo yed for adv anced learning and processing tasks. Here is a corrected and impro v ed v ersion: together , these components form a cohesi v e sys tem that optimizes both data e xtraction and subsequent analysis, thereby enhancing the o v erall ef fecti v eness of the methodology proposed in this study , as illustrated in Figure 1. Figure 1. Flo w diagram of methodology utilizing coordinate-based image partitioning This method re v olv es around a no v el concept centered on coordinates. Specically , partition each image into multiple di visions, and assign unique indices, referred to as coordinates, to each di vision. These coordinat es are pi v ot al as t he y guide the selector in e xtracting pat ches of pix els that are most re le v ant to the analysis. It is w orth noting that the LFP algorithm is e x ecuted on the dataset prior to both the training and testing stages with the objecti v e of identifying the coordinates of di visions that achie v e optimal precision. This process in v olv es a systematic e xploration of v arious di visional congurations within the dataset to determine which coordinates yield the highest le v els of accurac y . By meticulously e v aluating these di visions based on precision met rics, the algorithm aims to enhance its ability to accurately partition and analyze data, thereby impro ving o v erall performance in subsequent stages of training and testing [15]. 3.3. Pr oposed lear ning f ocal point-based ar chitectur e 3.3.1. Lear ning f ocal point algorithm Understanding the method of classication requires an e xploration of the LFP algorithm, which plays a critical role in identifying k e y areas within images. Based on perceptron principles, LFP utilizes an initial training layer to detect signicant features in each image dataset. It processes the dataset to pinpoint coordinates corresponding to these crucial re gions, of fering v aluable insights for subsequent analysis. The o wchart as in Figure 2 visually outlines the operational w orko w of the LFP algorithm. This gure illustrating its systematic approach to image analysis and feature e xtraction [16]. Figure 2. Flo wchart of LFP algorithm: visualizing image analysis and feature e xtraction Int J Artif Intell, V ol. 15, No. 3, June 2026: 2041–2052 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Artif Intell ISSN: 2252-8938 ❒ 2045 As depicted in the o wchart, the LFP algorithm initiates by di viding each image within the datas et into multiple areas, ef fecti v ely generating subdataset s. Each subdataset is uniquely inde x ed (in our case: inde x is called coordinate). Mo ving forw ard, the algorithm progresses t hrough se v eral k e y steps: the algorithm rst trains the perceptron on 80% of each subdataset to learn k e y image features , then e v aluates precision on the remaining 20%. It ranks subdatasets by accurac y and selects the coordinates of the subset with the highest precision, highlighting the most signicant image areas. In summary , the o wchart presents a systematic approach where image datasets are se gmented, models are trained and e v aluated, and optimal coordinat es are determined based on performance metrics. The LFP algorithm ini tializes dataset, estimates latent f actors, and iterati v ely renes them until con v er gence, ensuring ef fecti v e feature e xtraction and analysis, as in Figure 3. Figure 3. A dataset containing pix el v alues and tar get labels for each image Algorithm in v olv es partitioning each image into distinct areas. This se gmentati on is achie v ed by di viding the image dataset into groups of columns, where each group is identied by a coordinate inde x. Based on the original dataset, we generate additional subsets or sub datasets to focus on specic aspects or se gments of the dat a. These subsets may be created by applying lters, sampling techniques, or partitioning methods that e xtract subsets based on certain criteria or characteristics present in the original dataset. Each sub dataset thus represents a rened vie w or subset of the original data, tailored to address particular analyses, e xperiments, or modeling requirements. This approach allo ws for tar geted e xploration and analysis of dif ferent f acets or components within the dataset, enabling more focused insights and conclusions to be dra wn from the data as a whole. This process is illustrated in Figure 4. Figure 4. Generating subdatasets: fening insights from the main dataset As sho wn in Figure 5, each sub-dataset under goes indi vidual training using the perceptron model, with 80% of its data lines dedicated to training. This phase is essential for the model to learn and discern signicant features inherent in the images. Subsequently , the algorithm e v aluates the precision of each trained perceptron model using the remaining 20% of d a ta lines from each subdataset, serving as a v alidation set of unseen data. This e v aluation step critically assesses the accurac y and reliability of the perceptron’ s predictions on data it did not encounter during training, pro viding insights int o the model’ s rob ustness and its ability to generalize across dif ferent subsets of the dataset. Pneumothor ax detection using a learning focal point ar c hitectur e (Salah-Eddine Mansour) Evaluation Warning : The document was created with Spire.PDF for Python.
2046 ❒ ISSN: 2252-8938 Figure 5. Perceptron model training and e v aluation on subdatasets 3.3.2. Selector component Based on the coordinates identied by the LFP algorithm, the selector block utilizes this information to retrie v e specic patches of pix els from the image dataset. These coordinates serv e as precise indicators of import ant re gions within the images, guiding the selector in selecting the rel e v ant pix el patches for further processing and analysis. By e xtracting these patches, the selector block focuses on capturing detailed information from the designated areas identied by the LFP algorithm. This tar get ed approach ensures that the subsequent stages of data processing and analysis are conducted with a high de gree of accurac y and rele v ance, le v eraging the insights g ained from the initial feature detection phase performed by t he LFP algorithm, as illustrated in Figure 6. Figure 6. Enhanced pix el patch selection using LFP algorithm coordinates The LFP algorithm rst establishes foundational coordinates during the preprocessing stage, pro viding an initial spatial frame w ork for image analysis. Subsequently , the selector component dynamically identies and renes the most crucial pix els during both the trai ning and testing phases. This dynamic selection process enables the model to focus on the most informati v e re gions of the image. Both components were applied to chest X-ray images to ef fecti v ely e xtract essential pix els that contrib ute to accurate feature representation and decision-making. The e xtracted pix els and their spatial distrib ution are illustrated in Figure 7, highlighting the ef fecti v eness of the proposed approach in isolating clinically rele v ant re gions. The selector functions as an algorithmic lter designed to manage groups of pix els. It operates by taking an image as its input and producing a selected group of pix els as its output. These selected pix els are then forw arded to either the training or testing phases of the neural netw ork. The selector algorithm relies on coordinates as parameters to determine which specic areas or patches of pix els within the image should be e xtracted. This parameterization ensures that the selector ef fecti v ely tar gets and processes the most rele v ant pix el groups, optimizing their utility for subsequent neural netw ork training or testing procedures [17]. When emplo ying the LFP algorithm, we di vided the image into 13 squares, with each square measuring 79×79 pix els, gi v en the o v erall dimensions of 1024×1014 pix els for the image. Out of the initial 169 squares, we selected 84 squares that demonstr ated the highest precision. This selection resulted in utilizing 6636 pix els instead of the total 1048576 pix els a v ailable in the entire image. By focusing on these 84 squares with the best precision, we ef fecti v ely concentrated computati onal ef forts on the most rele v ant areas of the image, optimizing both accurac y and ef cienc y in subsequent analyses or tasks. Int J Artif Intell, V ol. 15, No. 3, June 2026: 2041–2052 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Artif Intell ISSN: 2252-8938 ❒ 2047 Figure 7. W orko w illustration of selector during training and testing stages 3.3.3. T raining and infer ence strategy After emplo ying the LFP algorithm and selector , we e xtracted the coordinates of the most cruc ial pix els. These coordinates were s ubsequently fed into the neural netw ork for both training and testing phases, as sho wn in Figure 8. W e implemented a multilayer perceptron (MLP) neural netw ork for learning, structured with four layers [18]. The rst layer w as congured with 6636 neurons, matching the number of pix els pro vided as input for initial processing and feature e xtraction. The nal layer w as designed with a single neuron, tailored for handling the binary nature of the problem at hand. This architecture allo wed the netw ork to ef fecti v ely process and classify the input data into the desired binary outcomes. Figure 8. T raining and testing stages utilizing the MLP neural netw ork 3.4. Baseline con v olutional neural netw ork models f or comparison Three well-kno wn CNN architectures, ResNet-50, DenseNet-121, and Ef cientNet, were applied on the pneumothorax X-ray dataset as baseline models to compare them with the LFP architecture. Each model w as initialized with ImageNet pre-trained weights and then ne-tuned on our dataset under the same conditions to ensure a f air and consistent e v aluation. Figure 9 presents the three e xperiments for these architectures. Figure 9. Fine-tuning of ResNet-50, DenseNet-121, and Ef cientNet on the pneumothorax dataset ResNet-50 is a deep CNN with residual connections that pre v ent v anishing gradients, allo wing it to learn hierarchical features from chest X-rays. It is ef fecti v e for detecting subtle signs of pneumothorax. DenseNet-121 uses dense connecti on s where each layer recei v es inputs from all pre vious layers, promoting feature reuse and capturing ne details, which helps identify small or hard-to-see pneumothoraces. Ef cientNet balances netw ork depth, width, and resolution to achie v e high accurac y with fe wer resources, making it f ast and suitable for clinical or mobile X-ray applications. T ogether , these models enhance AI-assisted pneumothorax diagnosis by combining accurac y , detail detection, and ef cienc y . 3.5. Explainability analysis: Grad-CAM vs LFP Gradient-weighted class acti v ation mapping (Grad-CAM) is a method that highlights the re gions of an image most inuential in a CNN’ s prediction by producing heatmaps from the nal con v olutional layers. While Grad-CAM generates relati v ely coarse maps for each X-ray , the LFP architecture with the selector precisely identies specic groups of pix els. These groups of pix els are critical for pneumothorax detection. Pneumothor ax detection using a learning focal point ar c hitectur e (Salah-Eddine Mansour) Evaluation Warning : The document was created with Spire.PDF for Python.
2048 ❒ ISSN: 2252-8938 The e v aluate model e xplainabil ity on t he pneumothorax X-ray dataset using tw o c o m plementary approaches. In Figure 10, Grad-CAM highlights image-le v el re gions that contrib ute most to the CNN’ s prediction by producing coarse acti v ation heatmaps deri v ed from the last con v olutional layers. In contrast, Figure 11 illustrates the LFP method with the selector , which identies precise pix el-le v el re gions in the form of selected squares that are consistently rele v ant for pneumothorax detection across images. This comparison demonstrates the dif ference between image-specic interpretability (Grad-CAM) and dataset-consistent localization (LFP), highlighting the impro v ed spatial precision and stability of the proposed approach. Figure 10. Grad-CAM heatmaps highlighting re gions most inuential for CNN predictions Figure 11. LFP selector output identifying precise pix el-le v el square re gions associated with pneumothorax detection It is important to understand the k e y dif ference between the LFP algorithm and Grad-CAM. The LFP algorithm identies and returns the coordinates of squares that are common across all images in the pneumothorax dataset, ef fecti v ely capturing the consistent re gions that are critical for classication. In contrast, Grad-CAM focuses on each indi vidual image and highlights the essential pi x el s that most inuence the CNN’ s prediction, producing image-specic heatmaps. This distinction mak es LFP particularly useful for identifying stable, dataset-wide features, while Grad-CAM pro vides insight into image-le v el acti v ations for model e xplainability . 4. RESUL TS AND DISCUSSION All three CNN architectures ResNet-50, DenseNet-121, and Ef cientNet were trained on the pneumothorax X-ray dataset using ImageNet pre-trained weights. The input images were resized to 224×224 pix els, and data augmentation techniques such as random rotation, horizontal ipping, brightness adjustment, and zoom were applied to impro v e generalization. Models were trained for 50–100 epochs with a batch size of 16–32 using the Adam optimizer , starting with a learning rate of 1e-4, which w as reduced on plateau, and binary cross-entrop y as the loss function. Early stopping w as emplo yed to pre v ent o v ertting [19]. ResNet-50 w as ne-tuned using slightly smaller learning rates for its deeper layers, DenseNet-121 beneted from dense connections that enabled f aster con v er gence and impro v ed recall, and Ef cientNet achie v ed the best balance between precision, recall, and o v erall computational ef cienc y . All models were e v aluated on a held-out test set using precision, accurac y , recall, F1-score, and A UC metrics to ensure a f air and consistent comparison. T able 1 reports the results of all models e v aluated in this study [20], [21]. Int J Artif Intell, V ol. 15, No. 3, June 2026: 2041–2052 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Artif Intell ISSN: 2252-8938 ❒ 2049 T able 1. Performance metrics of CNN architectures and LFP on pneumothorax X-ray dataset Method Precision Accurac y Recall F1-score A UC ResNet-50 0.825 0.828 0.827 0.826 0.824 DenseNet-121 0.828 0.830 0.831 0.829 0.827 Ef cientNet 0.831 0.831 0.832 0.831 0.829 LFP 0.8756 0.8756 0.8755 0.8756 0.8745 Figure 12 sho ws the R OC and precision-recall curv es for ResNet-50, DenseNet -121, Ef c ientNet, and LFP on the pneumothorax X-ray dataset. LFP outperforms all CNN models, achie ving the highest A UC and a v erage precision, while Ef cientNet, DenseNet-121, and ResNet-50 follo w in decreasing order . These results highlight the superior accurac y and reliability of the LFP architecture for pneumothorax detection. Figure 12. R OC and precision-recall curv es for CNN architectures and LFP on pneumothorax X-ray dataset T o e v aluate the rob ustness and generalization capability of the LFP architecture on the pneumothorax X-ray dataset, a stratied 5-fold cross-v alidation is performed. The dataset consists of 2027 samples, including 1597 pneumothorax cases (class 1) and 430 non-pneumothorax cases (class 0). Stratication w as applied to ensure that the class distrib ution w as preserv ed in each fold. In each iteration, the model w as trained on four folds and v alidated on the remaining fold, and k e y performance metrics namely accurac y , precision, recall, F1-score, and A UC were computed. This process w as repeated across all folds, and the mean and standard de viation of each metric were reported to pro vide a comprehensi v e e v aluation of the model’ s stability , predicti v e accurac y , and ability to handle class imbalance, as summarized in T able 2. T able 2. 5-fold cross-v alidation performance of the LFP architecture on the pneumothorax X-ray dataset F old Accurac y Precision Recall F1-score A UC F old 1 0.876 0.877 0.875 0.876 0.875 F old 2 0.874 0.875 0.874 0.874 0.873 F old 3 0.876 0.876 0.875 0.876 0.874 F old 4 0.875 0.876 0.875 0.875 0.874 F old 5 0.875 0.875 0.876 0.875 0.874 Mean ± Std 0.8756 ± 0.001 0.8756 ± 0.001 0.8755 ± 0.001 0.8756 ± 0.001 0.8745 ± 0.001 In T able 3, where each image w as initially se gmented into 169 di visions, a series of e xperiment s w as conducted to in v estig ate the ef fect of v arying the number of these di visions on system performance. Multiple e xperiments were performed by systematically altering the dimensions of the squares used for se gmentation each time. The ndings underscored the signicant impact of this parameter on the precision of our classication or analysis process. By systematically adjusting the se gmentation, it w as possible to observ e ho w dif ferent le v els of granularity in di viding the images af f ected the accurac y and reliability of the results. This e xploration helped identify optimal settings for achie ving higher precision in the system’ s outcomes [22]–[24]. Pneumothor ax detection using a learning focal point ar c hitectur e (Salah-Eddine Mansour) Evaluation Warning : The document was created with Spire.PDF for Python.
2050 ❒ ISSN: 2252-8938 T able 3. Impact of square dimension v ariation on image se gmentation Number of di visions Precision 25 0.79 100 0.83 169 0.87 196 0.86 255 0.84 The se gmentation of the image into 169 di visions resulted in achie ving the highest precision in this e xperiments, establishing it as the optimal number of di visions for our specic application. scenarios were also e xplored where the number of di visions w as either increased or decreased. These v ariations all o wed the study of ho w dif ferent le v els of se gmentation granularity af fected the precision of classication or analysis outcomes. By systematically adjusting the di vision count, the aim w as to identify whether ner or coarser di visions could potentially impro v e system performance further or if the y led to dim inishing returns. Thi s iterati v e approach pro vided v aluable insights into optimizing the se gmentation process to enhance the accurac y and reliability of the results across dif ferent e xperimental conditions [25]. 5. CONCLUSION The primary approach discussed in the paper in v olv es se gmenting images into multiple di visions using the LFP architecture. This method pro v es ef fecti v e in analyzing lung diseases within chest X-ray images and addresses the challenge o f accurately diagnosing pneumothorax. By se gmenting the images, the method enhances the ability to detect and assess pneumothorax areas more comprehensi v ely . This adv ancement not only increases diagnostic ef cienc y b ut also supports radiologists in reducing their w orkload by pro viding more detailed insights into the e xtent of pneumothorax, thereby f acilitating more informed treatment decisions. A CKNO WLEDGMENTS The author e xpresses sincere gratitude to the Centre National pour la Recherche Scientique et T echnique (CNRST) – Morocco for their nancial and institutional support, which w as essential for the completion of this research. Special thanks are also e xtended to mentors who pro vided v aluable guidance, feedback, and encouragement throughout this study . FUNDING INFORMA TION This research w as funded by Centre National pour la Recherche Scientique et T echnique (CNRST), Morocco. A UTHOR CONTRIB UTIONS ST A TEMENT (MAND A T OR Y) This journal uses the Contrib utor Roles T axonomy (CRediT) to recognize indi vidual author contrib utions, reduce authorship disputes, and f acilitate collaboration. Name of A uthor C M So V a F o I R D O E V i Su P Fu Salah-Eddine Mansour ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ Bouabid Qabliyane ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ Abdelhak Sakhi ✓ ✓ ✓ ✓ ✓ ✓ ✓ Zakaria Khoudi ✓ ✓ ✓ ✓ ✓ ✓ Mohamed Baslam ✓ ✓ ✓ ✓ ✓ ✓ C : C onceptualization I : I n v estig ation V i : V i sualization M : M ethodology R : R esources Su : Su pervision So : So ftw are D : D ata Curation P : P roject Administration V a : V a lidation O : Writing - O riginal Draft Fu : Fu nding Acquisition F o : F o rmal Analysis E : Writing - Re vie w & E diting Int J Artif Intell, V ol. 15, No. 3, June 2026: 2041–2052 Evaluation Warning : The document was created with Spire.PDF for Python.