Diagnostic imaging's importance is extensively acknowledged in the medical literature. Computed Tomography (CT) is a widely used diagnostic imaging method that is critical in the diagnosis of chest disorders such as COVID-19 [1–2]. Nonetheless, significant concerns have been raised concerning the quantity of ionizing radiation absorbed by patients during CT exams [3–5]. On the other hand, ionizing radiation has biological consequences, including the induction of chromosomal abnormalities and the development of cancer [6]. The extent of cell damage is dependent on the radiation dosage received [7] and the radiosensitivity of the exposed organ [8]. Although quantifying radiation-induced hazards for individual patients is difficult, 0.3 percent (1/330) of females aged 20 years who have CT pulmonary angiography acquire radiation-induced cancer [9]. Similarly, cumulative exposure to CT radiation from repeated CT examinations raises the risk of cancer [10]. CT dose can be reduced using a variety of techniques, including adjusting technical parameters and employing iterative reconstruction [11], in addition to establishing local diagnostic reference levels (DRL) for specified patient groups [12]. CT methods, as well as technical factors and radiation dosages, vary considerably between nations [3]. CT dosage variations have been attributed to machine usage and technical parameter selection depending on patient characteristics [3]. The term of patient characteristic has a significant effect on the amount of radiation absorbed. Obese patients or those with a high body mass index (BMI) often get a high dose of CT radiation [13]. Artificial Intelligence (AI) algorithms are currently being investigated for their potential to minimize the amount of radiation exposure [14]. The process of building computer systems capable of doing activities that need human intellect is referred to as artificial intelligence [15–16]. On the other hand, machine learning (MI) is a subset of artificial intelligence (AI) that comprises training algorithms to accomplish tasks through the discovery of patterns and characteristics in data. Additionally, deep learning (DL) is a subset of machine learning in which tasks are completed using deep neural networks with numerous layers of mathematical equations [15–16].
Numerous uses of AI in radiology are being researched, including the use of deep learning approaches to get high-quality images and reduce dosage [17–21]. Previous research has established the presence of Artificial Intelligence (AI) in the field of diagnosis. In a recent study, Meineke et al. have used a machine-learning algorithm to assess dosage optimality in CT efficiency [22]. In another investigation, Sinha et al. have evaluated the use of artificial neural networks in paediatric radiographs considering its potential in the identification of CT findings [23]. Their results showed that the artificial neural networks (ANN) model is more sensitive (82.2%) against physician estimation (62.2%). Their results also concluded that dose optimisation along with machine learning (ML) would comprehensively and rapidly detects CT scans [23]. McCollough and Leng have investigated the utilization of artificial intelligence for CT dose optimisation [24]. Their findings revealed that using AI to simplify the organ and pathologic characterisation had successfully migrated from a simulation to clinical practice and gave the rationale for imaging of pathologies in the right direction. Currently, machine learning approaches are being used to create promising elements of modern radiology, such as cancer prediction in patients with incidental lung nodules [24], radionomics [25], biomarker extraction [26], and assessment of cardiac-CT-derived fractional flow reserve [23]. With advancements in software-based technologies, machine learning-based algorithms are being utilized to analyse images, forecast outcomes [27], and as a quality assurance tool [28–29]. Recent research emphasizes the field's contribution of machine learning approaches. In a study [30], the authors conducted a thorough analysis of three machine learning models (logistic regression, support vector machines, and neural networks) to predict whether a patient will experience symptoms following a certain radiation dosage. In another study [31], dosage prediction in CT testing was performed using machine learning techniques such as linear regression, regression trees, Gaussian processes, support vector machines, and neural networks, with the ANN outperforming the other approaches. It has been established that a neural network regression model may be used to analyse large-scale CT dose data in order to detect insufficient and inefficient CT dosage application [22].
To the best of our knowledge, dose length product (DLP) estimation using ten different scan-related input parameters for ANN for chest CT scan was not performed in literature. Therefore, this study aimed to generate a novel DLP estimation algorithm using advanced artificial learning methods. Accordingly, we utilized ANN to create a new learning algorithm for estimation of DLP values in chest CT scan considering the total number of 1897 examinations and their extended scan parameters. We used a fairly big data set in our research, and the R-value, or the correlation coefficient between the test data and the real system, was determined to be 0.97328, 0.82263, and 0.90574 for the Levenberg-Marquardt, Bayesian, and Scaled-Conjugate Gradient algorithms, respectively. The model with Levenberg-Marquardt is highly satisfactory in terms of classification accuracy and R value compared with former literature studies. A variation of 18.3 percent between projected (anticipated) and actual DLP was shown to be ideal for detecting exams with poor dosage administration, and our findings highlight the approximation performance of the algorithm. We hypothesized that ANNs may be beneficial in today's CT dose assessment and quality assurance systems.
The outcomes and proposed DLP estimation method would help to specialists for advanced risk estimation and dose optimization in daily chest CT routines. It is worth mentioning that the using of retrospective data was main flaw of our research. Moreover, the patient number could be higher in terms of sample size. However, this study can be considered as a preliminary study and may create a promising motivation for similar future studies.