The National Center for Health and Science claims that 150,000 human deaths are due to lung cancer [1]. In 2020, the US had 228,820 new cases of lung cancer were 135,270 lung cancer-related fatalities [2]. They described the current advancement and assessment of computer-aided diagnosis (CAD) for detecting lung nodules in CT [3]. A CT scan, however, provides a high number of images. To determine the nodule’s diameter, the doctor visualizes it with his naked eye. In some cases, the results could be incorrect and even lead to death. A CAD system can assist in diagnosing pulmonary nodules. In modern medicine, CAD technology has become a major research topic. To assist doctors in the early detection of cancer [4]. Lung cancer requires automatic lung nodule detection, which can provide classification and 3D information [5]. In CAD systems, image segmentation plays a very important role in the detection of nodules. It extracts the area of interest and location of the nodule. This increases the accuracy of medical data and faster the result to detect nodule detection based on multiscale feature fusion [6]. The significance of chest CT scans in the process of diagnosing lung cancer and correcting anomalies discovered in chest radiography [7]. The procedure of identifying lung nodules can be broken down into three primary stages: first enhancing and processing the images second extracting features, and third carrying out the actual detection. During the first stage of the process, operators are used on CT scans to emphasize particular elements that are contained within the area of interest (ROI), with the primary goal of determining the curvature of the lung while avoiding any unnecessary regions. Because of this, the search area is effectively reduced, which speeds up the nodule detection process. After that, the lungs are cut up into a series of tiny rectangular pieces so that the features may be extracted from each specific portion. After the traits have been retrieved, they are entered into a decision-making system so that it can determine whether or not there is a tumor present in that particular unit [8]. Some research has been conducted to investigate a wide variety of nodule detection methods. Some people looked at how successful hierarchical bit plane histograms and cumulative distribution functions were, but they ran into difficulties when trying to determine the ideal number of bins for the different kinds of images. Others have presented methods for image segmentation that apply morphological operators and make use of histograms and wavelet coefficients as effective descriptors. In addition, some individuals utilized local kernel regression models, often known as LKRMs, for the purpose of classifying lung nodule [9]. In one study, an active counter modeling strategy was presented for the purpose of segmenting and detecting lung nodules. This method utilized a support vector machine (SVM) for the purpose of classifying the nodules found in each lung. This article provides a detailed review of contemporary nodule detection approaches, with the goal of summing up current trends and forecasting upcoming issues in this domain [10]. The contribution of our proposed research work is:

A new pulmonary nodule detection method was developed by combining a histogram of surface normal vectors with a scale-dependent textural description and improved local binary pattern features.

The manuscript is organized into five chapters, the second of which presents a detailed Literature Review. Suitable research and theories are discussed in detail. The third chapter offers a discussion of the Methodology, prefacing research approaches, and techniques; Chapter four enters into an examination of results with concise data analysis. The final fifth chapter draws together the main points, leaving readers with four key conclusions of this research and further areas for future exploration.

One common intensity-based detection method is the use of a threshold to define a nodule region of interest. The threshold is typically selected manually or automatically using a training set of images [11]. To further segment the nodule region from the surrounding lung parenchyma, region-growing or watershed segmentation methods are used. The study found that the 3D Automatic Lung Parenchyma Extraction and Border Repair (ALPEBR) method, which includes an Automatic Single Seeded Region Growing (ASSRG) and a 3D Spherical region-growing method (SPRG) [12]. There have been several traditional approaches to lung tumor segmentation before deep learning techniques were developed, such as thresholding, region-growing, and machine learning-based techniques [13] Nodule identification has been presented by [14] using single feed-forward approaches. Convolutional neural networks (CNNs) with dense connections were designed as one end-to-end 3D network. Neither post-processing nor user guidance is required to refine detection results using S4ND. Using 3D Deep Convolutional Neural Network (DCNNs) in computer vision and lung nodule detection, they show how the network compares with the current state-of-the-art object detection networks using U-Net and LSTM networks [15]. A decision-level technique was used to improve the classification of nodules using support vector machine (SVM) and Random Forest CNN architectures were used for the classification of nodules [16]. They presented a multi-task deep learning framework in which segmentation and physical location were identified using image bands [17]. Furthermore, they used the baseline approach and deep learning methods like Region-based Convolutional Network (RCNN) and U-net for the analysis of solar images and hole identification [18]. To solve many computational problems a probabilistic technique is required. ANT colony optimization algorithm solves problems like multitargeting and vehicle targeting [19]: The proposed method involves three steps: converting images into 3D, applying a blocking technique for informative block detection, segmenting the image for candidate selection of nodules, and performing classification [20]. They proposed the rolling ball algorithm to minimize the loss of juxtapleural nodules by applying the gray-level thresholding technique. Segmentation has been used to classify the nodule by removing the background which is separate from the thorax and lungs from the thorax to identify the candidate nodule detection [21]. They proposed the use of deep learning models, such as DeepLabV3+ and U-shaped networks, for lung segmentation and vessel segmentation in CT images [22]. They presented an ensemble method, which involves many classifiers to create and evaluate the predictive model. By employing the dataset on coronary artery disease, the present investigation showcases that the ensemble technique exhibits superior performance compared to other established methods for predicting the disease [23]. According to the evaluation of deep learning-based methods for multi-modal medical picture segmentation, the later fusion strategy has the potential to yield more accurate results than the early fusion strategy, provided that the fusion method is efficient enough [24]. They proposed an efficient approach for the categorization of gastrointestinal diseases by utilizing a pre-trained deep convolutional neural network [25]. This study presents a novel approach for the classification of gastrointestinal disorders by leveraging deep transfer learning and optimizing characteristics. The proposed technique demonstrates a notable enhancement in accuracy, obtaining a commendable rate of 96.43%. The utilization of Mask R-CNN for disease localization involves the implementation of fine-tuned ResNet-50 and ResNet-152 models to extract relevant features [26]. Firstly, classic techniques like thresholding and region-growing may face limitations in effectively delineating complex nodule boundaries or distinguishing them from surrounding tissues. Moreover, some strategies discussed might necessitate manual intervention or parameter adjustments, potentially yielding varied outcomes. The effectiveness of specific methods, particularly those predating deep learning advancements, may be influenced by factors like image quality and variations in nodule properties. Considering these limitations is imperative in real-world clinical applications, prompting a need to explore new directions for further refinement and improvement.

The number of pulmonary nodules in the right and left lungs are shown in Fig. 1. The proposed method is described below in a step-by-step manner:

To segment the volume of the lungs, two-dimensional histograms are applied.

In many applications, good threshold values have been selected to achieve high accuracy. Pulmonary nodules are classified into four components which are proposed in the methodology as shown in Fig. 2.

In the initial stage of the process, differential evolution optimization techniques are employed to obtain automatic threshold values for nodules, walls, and arteries within the lung volume. The method uses two-dimensional histogram technology, combining the steps of geometric feature extraction and arc delineation, along with extended arc extraction and ellipse detection to find all kinds of shapes in nodules from 3 mm up until about 20 mm. In addition, texture descriptions such as the histogram-oriented surface normal vectors (HOSNV) are introduced to differentiate between them. This makes use of intensity deviations in histograms from labeled CT scan images and plays into proposed intelligence-based statistical features.

We evaluate the proposed system using the LIDC database [34]. The evaluation was based on 59 CT scans. Per scan, there are approximately 200 slices, each with 512 pixels by 512 pixels and 4096 gray levels. The database has pixels between 0.5 and 0.76 mm in size and reconstruction intervals between 1 and 3 mm. A total of 389 nodules and 10,712 slices are included. There are nodules with diameters ranging from 3 to 30 millimeters and 3 to 30 mm. Some ratios have been tested, as well as different features, and then, classification has been conducted using the XGBoost classifier. To make sure that there is no bias toward overtraining or undertraining, we divided the data into 50–50 ratios that were used as training data and testing data to demonstrate our results. XGBoost and other classifiers were used to test the proposed features, as shown in Table 1. Compared to all other classifiers, XGBoost performs better. The performance of classifiers is assessed using three key measures: Accuracy, Sensitivity, and Specificity, calculated based on the counts of true positives (TP), false positives (FP), false negatives (FN), and true negatives (TN) in the classification of nodules. This shows that the proposed feature is best for classifying nodules as shown in Fig. 4.

Using all three feature types and different classifiers, compared to earlier existing techniques, the proposed method has a sensitivity of 98.49 for nodules sized 3–30 mm. As a result of the HOSNV and LBP features proposed in the proposed method and random forest, the proposed method performs better as shown in Fig. 4.

Based on proposed NESGF features, random forest-based classification is used for comparison with other existing methods. From the results in Table 2 and Fig. 5, segmentation clearly plays a crucial role in improving performance. This proposed method distinguishes lung nodules from walls and vessels with fuzzy c-means (FCM) clustering by integrating 2D and 3D images. With a hybrid feature descriptor based on shape information, combined with gradient features in an NESGF format it is more effective than existing methods.

A novel method for detecting pulmonary nodules was developed by combining a histogram of surface normal vectors with a scale-dependent textural description and improved local binary pattern features. The proposed paradigm facilitates easier medical and diagnostic decisions. An optimal thresholding method is employed to segment lung volume, followed by Ant Colony optimization, label masking, background removal, and contour correction. 2D and 3D Regions of Interest (ROIs) are utilized for detecting candidate nodules. A classifier based on XGBoost is employed to classify proposed features NESGF. Testing involved classifiers, extracted features, and extracted nodules from the LIDC dataset for segmentation, nodule extraction, and feature extraction. The classifier results indicate 98.49% accuracy in category classification, with 98.82% sensitivity. In comparison with conventional descriptors, our proposed feature descriptor performs superiorly, as demonstrated by the results. We plan to explore deeper learning techniques in the future and test the proposed method on other datasets.