{"id":62821,"date":"2025-02-20T11:38:56","date_gmt":"2025-02-20T11:38:56","guid":{"rendered":"https:\/\/biomedpharmajournal.org\/?p=62821"},"modified":"2025-03-18T10:57:38","modified_gmt":"2025-03-18T10:57:38","slug":"decoding-challenges-using-mathematics-of-fuzzy-theory-in-interpretability-shifts-adaptation-and-trust","status":"publish","type":"post","link":"https:\/\/biomedpharmajournal.org\/staging\/vol18marchspledition\/decoding-challenges-using-mathematics-of-fuzzy-theory-in-interpretability-shifts-adaptation-and-trust\/","title":{"rendered":"Decoding Challenges using Mathematics of Fuzzy Theory in Interpretability, Shifts, Adaptation and Trust"},"content":{"rendered":"\n

Introduction<\/strong><\/p>\n\n\n\n

The Amalgamation of\nArtificial Intelligence into medical imaging streak an everchanging era in\ndiagnostic medicine where technology meets healthcare which promises enhanced\npatient outcomes and improved clinical workflows. The Combined & crucial\nrole of Fuzzy and AI\u2019s in medical imaging can be seen through their competence\nin analysing complex data with precision and speed that were previously\nimpassable, heralding a new age of diagnostic precision and efficiency1<\/sup>.However\ncontinuous advancement brings some threats, too. The clinical deployment of AI\nsystems in medical imaging accommodates a multifaceted approach that marks\ninterpretability, domain shift, adaption, and trustworthiness to ensure their\nefficacy and assimilation into patient care (RSNA,2023)<\/p>\n\n\n\n

In the examination for\ninterpretability, the black-box nature of AI systems often poses a compelling\nbarrier, and we use fuzzy logic for the same. The acceptance of clinicians to\ninterpret and trust AI decision-making processes is preeminent, mainly when\nsuch decisions encounter patient diagnosis and treatment (NCBI, 2023). This\ntrust can only be settled through transparent AI models that provide intuition\ninto their reasoning, thereby providing an environment of informed clinical DM\nand adequate patient communication  <\/p>\n\n\n\n

Moreover, the domain\nshift presents a hefty obstacle. AI models, often trained on peculiar datasets,\nmay fail when applied to new datasets from various clinical environments, a\ncertainty that is prevalent in the distinct landscape of healthcare2,3<\/sup>.\nAddressing this problem requires a robust adaptation procedure that enables AI\nsystems to maintain steady performance across a spectrum of medical imaging\nmodalities and patient populations, conflicting the need for wide-ranging\nretraining4<\/sup>.<\/p>\n\n\n\n

Lastly, the\ntrustworthiness of AI systems in medical imaging extends beyond accuracy. It\nencompasses the reliability, fairness, and ethical considerations that are\nintegral to clinical acceptance. Ensuring that AI systems adhere to these\nprinciples is not only a technical challenge but also a moral imperative, as\nthe main goal of AI in healthcare is to benefit patient safety and well-being5<\/sup>\u200b\u200b.<\/p>\n\n\n\n

As the field of medical\nimaging continues to evolve, the need for AI systems that are technologically\nadvanced and the application of interdisciplinary approaches is much\nrequired.  In the coming section of this\narticle, the historical context of AI is presented. The next section deals with\ninterpretability, presenting the case study using fuzzy sets where the\nparameters taken as pixel intensity, greyscale, and texture coefficient, and we\ntook the threshold value as 0.6. Further, we analyzed the domain shift with\nfour features. The article concludes with the final conclusion and future\nstudies.<\/p>\n\n\n\n

Historical Context of AI in Medical Imaging<\/strong><\/p>\n\n\n\n

Artificial Intelligence\n(AI) in medical imaging is not a cutting-edge phenomenon but rather an growing\none, with its inception dating back to the mid-20th century. The thought of AI\nwas first described in 1950, and it aimed to mimic human cognitive functions.\nHowever, the practical implementation of AI in medicine was stalled due to\ntechnological inadequacies of initial models. These limitations persisted until\nthe early 2000s, when the advent of deep learning significantly propelled the\ncapabilities of AI systems, thus overcoming many previous barriers and setting\nthe stage for their integration into medical imaging.6<\/sup><\/p>\n\n\n\n

The 1980s marked a\nresurgence of AI due to international competition, but it was also a period\nknown as the ‘AI winter’ from 1983 to 1993, characterized by a collapse of the\nmarket for the computational power needed at the time, leading to a withdrawal\nof funding. However, research and development in AI did not come to a complete\nhalt and picked up pace thereafter, setting the groundwork for its eventual\nresurgence and integration into the medical field7<\/sup>\u200b\u200b.<\/p>\n\n\n\n

Current State of the Art and its Limitations<\/strong><\/p>\n\n\n\n

Today, AI applications in\nmedical imaging are widespread and continuously growing, with the field widely\nrecognizing that AI will completely transform medical diagnostics. Current AI\nmodels have shown remarkable success in the interpretation of medical images,\nand their use has been extended to various applications, including the\ndetection of abnormalities and quantification of disease processes8,9<\/sup>.<\/p>\n\n\n\n

However, despite these\nadvances, AI in medical imaging is not without its limitations. One such\nlimitation is the ‘black box’ nature of many AI systems, where the reasoning\nbehind AI decisions is not transparent, posing a significant challenge for\nclinical acceptance. Additionally, while AI aims to replicate human\ndecision-making, it still struggles with issues such as domain shift, where\nmodels trained on one set of data fail to generalize to other datasets, often\nseen in the diverse clinical environments encountered in healthcare10<\/sup>.<\/p>\n\n\n\n

The Need for Improved AI Interpretability and Trust in Clinical Settings<\/strong><\/p>\n\n\n\n

The necessity for AI\ninterpretability in medical imaging is paramount, as it is crucial for\nclinician trust. Clinicians need to comprehend the AI decision-making process\nto make informed decisions and communicate effectively with patients. The\nparadigm change that AI is bringing to healthcare is driven by the enlarging\navailability of healthcare data and enhancements in analytics methods. The\nfuture of AI in healthcare is envisioned to address these interpretability\nissues, thereby increasing trust and reliability in clinical settings11<\/sup>\u200b\u200b.<\/p>\n\n\n\n

Recent advancements in\nartificial intelligence (AI) have demonstrated the potential to transform\nmedical diagnostics, particularly through techniques like fuzzy logic and\nexplainable AI 12, 13<\/sup> frameworks. Fuzzy soft set theory, for example, has been extensively\nreviewed for its applications in medical diagnosis and decision making14<\/sup>, offering a structured approach to handle uncertainty in complex\nmedical data15<\/sup>. Furthermore, explainability and transparency are critical to the\nadoption of AI in clinical settings, as many AI models, particularly in\nradiology, struggle with the “black box” problem, which undermines\nclinicians’ trust16, 17<\/sup>. To address these issues, recent research has emphasized the\nimportance of harmonized data infrastructures and federated learning models,\nwhich facilitate secure and efficient data sharing across healthcare systems,\nenabling the development of robust AI models for medical imaging18<\/sup>. Additionally, innovations in fuzzy logic, such as the intelligent\nsaline control valve, highlight the practical applications of AI in patient\ncare, providing solutions to optimize clinical procedures19<\/sup>. <\/p>\n\n\n\n

Interpretability<\/strong><\/p>\n\n\n\n

In machine learning, interpretability is specified as the degree to\nwhich a human can envision the reasons behind a model\u2019s decision. This is not\nonly a technical prerequisite but an ethical imperative, especially in sectors\nwhere decisions have deep implications on human lives, such as healthcare.\nTrust in these systems comes from their capacity to bring transparent reasoning\nfor their decisions, which arouses higher acceptance and fidelity to clinical\ndecision-making facilitated by AI.<\/p>\n\n\n\n

Approach for Model Interpretability <\/strong><\/p>\n\n\n\n

The quest for interpretability has given acceleration to different\ntechniques. Inherently interpretable models like logistic regression and\ndecision trees offer clarity through their elimination and the direct way they\ncan be charted to human-understandable rules. However, with the arrival of\ncomplex models like deep neural networks, the field has moved toward a post-hoc\ninterpretability approach. Model agnostic approaches like LIME quip local\ninterpretability, providing explanations for certain indicators heedless of the\nmodel\u2019s complexity. SHAP values boost this by allocating each attribute a value\nfor certain predictions, depiction from cooperative game theory to establish\nconsistency and precision in featuring crucial attribution <\/p>\n\n\n\n

Visual techniques are also instrumental; for instance, Class Activation\nMapping (CAM) and its variants allow for the envision of regions in the input\nimage that are essential for predictions by a convolutional neural network,\nproviding clinicians with a visual rationale for the AI’s decision.<\/p>\n\n\n\n

Interpretability in Medical Imaging<\/strong><\/p>\n\n\n\n

The utilization and\nsignificance of interpretability in medical imagining are illustrated through\nspecific case studies. For instance, a study on interpreting AI decisions in\nmammography has shown that using heatmaps to indicate areas of interest helps\nradiologists to quickly focus on potential issues and corroborate the AI’s\nfindings with their expertise. In another case, the use of AI to diagnose\ndiabetic retinopathy was greatly enhanced by interpretability techniques that\nallowed ophthalmologists to understand the basis of the AI\u2019s diagnostic\nsuggestions, thus integrating AI assistance seamlessly into their clinical\nworkflow. Fig.1 simulate the appearance of an AI-generated heatmap on a medical\nimage for illustrative purposes.<\/p>\n\n\n\n\n\n
\"\"<\/td>\n\n

Figure 1: A medical scan with a heatmap overlay, to illustrate AI interpretability in medical diagnostics<\/strong><\/p>\n

<\/p>\n

Click here to view Figure<\/a><\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n\n\n

Materials and Methods <\/strong><\/p>\n\n\n\n

Case Study 1<\/strong><\/p>\n\n\n\n

Consider the problem of\nan MRI scan analysis for tumor detection. The challenge is to accurately\nidentifying tumor tissues amidst a variety of other factors. <\/p>\n\n\n\n

In this example, an MRI scan with 10 distinct pixels, each characterized by four parameters: Membership Grade, Pixel Intensity, Grey Scale, and Texture Coefficient is taken into consideration. The task is to analyze these pixels using Fuzzy Logic to determine the likelihood of each pixel being part of a tumor. <\/p>\n\n\n\n

We assume 10 pixels, each with four parameters.<\/p>\n\n\n\n

Each pixel\u2019s membership grade (0.1 to 0.9) reflects its likelihood of being tumor tissue.<\/p>\n\n\n\n

A threshold of 0.6 indicates higher tumor likelihood.<\/p>\n\n\n\n

Pixel Intensity, Grayscale, and Texture Coefficient are integrated with membership grades for tumor detection.<\/p>\n\n\n\n

The Weighted Average defuzzification method computes a single value for each pixel.<\/p>\n\n\n\n

Higher defuzzified values indicate a greater likelihood of tumor presence.<\/p>\n\n\n\n

Table 1: Parameter influence on the final defuzzified value<\/strong><\/p>\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
\n

Pixel<\/strong><\/p>\n<\/td>\n

\n

Membership Grade<\/strong><\/p>\n<\/td>\n

\n

Pixel Intensity<\/strong><\/p>\n<\/td>\n

\n

Grey Scale<\/strong><\/p>\n<\/td>\n

\n

Texture Coefficient<\/strong><\/p>\n<\/td>\n

\n

Threshold (Grade > 0.6)<\/strong><\/p>\n<\/td>\n

\n

Defuzzified Value<\/strong><\/p>\n<\/td>\n<\/tr>\n

\n

1<\/p>\n<\/td>\n

\n

0.10<\/p>\n<\/td>\n

\n

35<\/p>\n<\/td>\n

\n

120<\/p>\n<\/td>\n

\n

0.30<\/p>\n<\/td>\n

\n

0.00<\/p>\n<\/td>\n

\n

0.000000<\/p>\n<\/td>\n<\/tr>\n

\n

2<\/p>\n<\/td>\n

\n

0.30<\/p>\n<\/td>\n

\n

45<\/p>\n<\/td>\n

\n

130<\/p>\n<\/td>\n

\n

0.35<\/p>\n<\/td>\n

\n

0.00<\/p>\n<\/td>\n

\n

0.000000<\/p>\n<\/td>\n<\/tr>\n

\n

3<\/p>\n<\/td>\n

\n

0.50<\/p>\n<\/td>\n

\n

60<\/p>\n<\/td>\n

\n

145<\/p>\n<\/td>\n

\n

0.40<\/p>\n<\/td>\n

\n

0.00<\/p>\n<\/td>\n

\n

0.000000<\/p>\n<\/td>\n<\/tr>\n

\n

4<\/p>\n<\/td>\n

\n

0.20<\/p>\n<\/td>\n

\n

20<\/p>\n<\/td>\n

\n

110<\/p>\n<\/td>\n

\n

0.20<\/p>\n<\/td>\n

\n

0.00<\/p>\n<\/td>\n

\n

0.000000<\/p>\n<\/td>\n<\/tr>\n

\n

5<\/p>\n<\/td>\n

\n

0.40<\/p>\n<\/td>\n

\n

40<\/p>\n<\/td>\n

\n

125<\/p>\n<\/td>\n

\n

0.25<\/p>\n<\/td>\n

\n

0.00<\/p>\n<\/td>\n

\n

0.000000<\/p>\n<\/td>\n<\/tr>\n

\n

6<\/p>\n<\/td>\n

\n

0.60<\/p>\n<\/td>\n

\n

55<\/p>\n<\/td>\n

\n

150<\/p>\n<\/td>\n

\n

0.45<\/p>\n<\/td>\n

\n

0.00<\/p>\n<\/td>\n

\n

0.000000<\/p>\n<\/td>\n<\/tr>\n

\n

7<\/p>\n<\/td>\n

\n

0.80<\/p>\n<\/td>\n

\n

80<\/p>\n<\/td>\n

\n

200<\/p>\n<\/td>\n

\n

0.60<\/p>\n<\/td>\n

\n

0.80<\/p>\n<\/td>\n

\n

0.141933<\/p>\n<\/td>\n<\/tr>\n

\n

8<\/p>\n<\/td>\n

\n

0.70<\/p>\n<\/td>\n

\n

75<\/p>\n<\/td>\n

\n

190<\/p>\n<\/td>\n

\n

0.55<\/p>\n<\/td>\n

\n

0.70<\/p>\n<\/td>\n

\n

0.101391<\/p>\n<\/td>\n<\/tr>\n

\n

9<\/p>\n<\/td>\n

\n

0.90<\/p>\n<\/td>\n

\n

90<\/p>\n<\/td>\n

\n

210<\/p>\n<\/td>\n

\n

0.65<\/p>\n<\/td>\n

\n

0.90<\/p>\n<\/td>\n

\n

0.204334<\/p>\n<\/td>\n<\/tr>\n

\n

10<\/p>\n<\/td>\n

\n

0.85<\/p>\n<\/td>\n

\n

85<\/p>\n<\/td>\n

\n

205<\/p>\n<\/td>\n

\n

0.70<\/p>\n<\/td>\n

\n

0.85<\/p>\n<\/td>\n

\n

0.191607<\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n\n\n

Pixels 7, 8, 9, and 10,\nwith defuzzified values of 0.141933, 0.101391, 0.204334, and 0.191607,\nrespectively, suggest a likelihood of being part of a tumor.<\/p>\n\n\n\n

Case Study 2<\/strong><\/p>\n\n\n\n

Consider two domains of\nmedical imaging data for tumor detection, Domain A (Source) and Domain B\n(Target), each domain comprises data points with four features: Feature 1\n(e.g., tumor size), Feature 2 (e.g., texture), Feature 3 (e.g., shape\nirregularity), and Feature 4 (e.g., presence of specific markers). <\/p>\n\n\n\n

The challenge is to adapt\nan AI model trained on Domain A to maintain its diagnostic accuracy when\napplied to Domain B, addressing the domain shift numerically represented by\ndifferences in feature distributions and associations with tumor\ncharacteristics.<\/p>\n\n\n\n

In this example\n(hypothetical data), we use fuzzy logic for Feature1 classification, and apply\ndomain adaptation techniques for Feature 2, Feature3, and Feature4.<\/p>\n\n\n\n

The Fuzzy Logic\nParameters used are: benign threshold for Feature 1 is 5; malignant threshold\nfor Feature 1 is 7. The domain shift is shown in Fig. 2.<\/p>\n\n\n\n

Table 2: Domain A (Source)<\/strong><\/p>\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
\n

Feature 1<\/strong><\/p>\n<\/td>\n

\n

Benign Grade<\/strong><\/p>\n<\/td>\n

\n

Malignant Grade<\/strong><\/p>\n<\/td>\n

\n

Feature 2<\/strong><\/p>\n<\/td>\n

\n

Feature 3<\/strong><\/p>\n<\/td>\n

\n

Feature 4<\/strong><\/p>\n<\/td>\n<\/tr>\n

\n

4.1<\/p>\n<\/td>\n

\n

1.00<\/p>\n<\/td>\n

\n

0.00<\/p>\n<\/td>\n

\n

2.08<\/p>\n<\/td>\n

\n

7.42<\/p>\n<\/td>\n

\n

7.85<\/p>\n<\/td>\n<\/tr>\n

\n

4.42<\/p>\n<\/td>\n

\n

0.89<\/p>\n<\/td>\n

\n

0.11<\/p>\n<\/td>\n

\n

2.59<\/p>\n<\/td>\n

\n

7.64<\/p>\n<\/td>\n

\n

7.99<\/p>\n<\/td>\n<\/tr>\n

\n

4.74<\/p>\n<\/td>\n

\n

0.78<\/p>\n<\/td>\n

\n

0.22<\/p>\n<\/td>\n

\n

3.11<\/p>\n<\/td>\n

\n

7.85<\/p>\n<\/td>\n

\n

8.14<\/p>\n<\/td>\n<\/tr>\n

\n

5.07<\/p>\n<\/td>\n

\n

0.67<\/p>\n<\/td>\n

\n

0.33<\/p>\n<\/td>\n

\n

3.62<\/p>\n<\/td>\n

\n

8.07<\/p>\n<\/td>\n

\n

8.29<\/p>\n<\/td>\n<\/tr>\n

\n

5.39<\/p>\n<\/td>\n

\n

0.56<\/p>\n<\/td>\n

\n

0.44<\/p>\n<\/td>\n

\n

4.13<\/p>\n<\/td>\n

\n

8.29<\/p>\n<\/td>\n

\n

8.44<\/p>\n<\/td>\n<\/tr>\n

\n

5.71<\/p>\n<\/td>\n

\n

0.44<\/p>\n<\/td>\n

\n

0.56<\/p>\n<\/td>\n

\n

4.65<\/p>\n<\/td>\n

\n

8.50<\/p>\n<\/td>\n

\n

8.58<\/p>\n<\/td>\n<\/tr>\n

\n

6.03<\/p>\n<\/td>\n

\n

0.33<\/p>\n<\/td>\n

\n

0.67<\/p>\n<\/td>\n

\n

5.16<\/p>\n<\/td>\n

\n

8.72<\/p>\n<\/td>\n

\n

8.73<\/p>\n<\/td>\n<\/tr>\n

\n

6.36<\/p>\n<\/td>\n

\n

0.22<\/p>\n<\/td>\n

\n

0.78<\/p>\n<\/td>\n

\n

5.67<\/p>\n<\/td>\n

\n

8.94<\/p>\n<\/td>\n

\n

8.88<\/p>\n<\/td>\n<\/tr>\n

\n

6.68<\/p>\n<\/td>\n

\n

0.11<\/p>\n<\/td>\n

\n

0.89<\/p>\n<\/td>\n

\n

6.19<\/p>\n<\/td>\n

\n

9.15<\/p>\n<\/td>\n

\n

9.02<\/p>\n<\/td>\n<\/tr>\n

\n

7.00<\/p>\n<\/td>\n

\n

0.00<\/p>\n<\/td>\n

\n

1.00<\/p>\n<\/td>\n

\n

6.70<\/p>\n<\/td>\n

\n

9.37<\/p>\n<\/td>\n

\n

9.17<\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n\n\n

Table 3: Domain B (Target)<\/strong><\/p>\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
\n

Feature 1<\/strong><\/p>\n<\/td>\n

\n

Benign Grade<\/strong><\/p>\n<\/td>\n

\n

Malignant Grade<\/strong><\/p>\n<\/td>\n

\n

Feature 2<\/strong><\/p>\n<\/td>\n

\n

Feature 3<\/strong><\/p>\n<\/td>\n

\n

Feature 4<\/strong><\/p>\n<\/td>\n<\/tr>\n

\n

7.1<\/p>\n<\/td>\n

\n

0.0<\/p>\n<\/td>\n

\n

1.0<\/p>\n<\/td>\n

\n

13.07<\/p>\n<\/td>\n

\n

13.96<\/p>\n<\/td>\n

\n

8.18<\/p>\n<\/td>\n<\/tr>\n

\n

7.37<\/p>\n<\/td>\n

\n

0.0<\/p>\n<\/td>\n

\n

1.0<\/p>\n<\/td>\n

\n

13.25<\/p>\n<\/td>\n

\n

14.03<\/p>\n<\/td>\n

\n

8.90<\/p>\n<\/td>\n<\/tr>\n

\n

7.63<\/p>\n<\/td>\n

\n

0.0<\/p>\n<\/td>\n

\n

1.0<\/p>\n<\/td>\n

\n

13.44<\/p>\n<\/td>\n

\n

14.11<\/p>\n<\/td>\n

\n

9.61<\/p>\n<\/td>\n<\/tr>\n

\n

7.9<\/p>\n<\/td>\n

\n

0.0<\/p>\n<\/td>\n

\n

1.0<\/p>\n<\/td>\n

\n

13.62<\/p>\n<\/td>\n

\n

14.18<\/p>\n<\/td>\n

\n

10.33<\/p>\n<\/td>\n<\/tr>\n

\n

8.17<\/p>\n<\/td>\n

\n

0.0<\/p>\n<\/td>\n

\n

1.0<\/p>\n<\/td>\n

\n

13.80<\/p>\n<\/td>\n

\n

14.25<\/p>\n<\/td>\n

\n

11.05<\/p>\n<\/td>\n<\/tr>\n

\n

8.43<\/p>\n<\/td>\n

\n

0.0<\/p>\n<\/td>\n

\n

1.0<\/p>\n<\/td>\n

\n

13.99<\/p>\n<\/td>\n

\n

14.33<\/p>\n<\/td>\n

\n

11.76<\/p>\n<\/td>\n<\/tr>\n

\n

8.7<\/p>\n<\/td>\n

\n

0.0<\/p>\n<\/td>\n

\n

1.0<\/p>\n<\/td>\n

\n

14.17<\/p>\n<\/td>\n

\n

14.40<\/p>\n<\/td>\n

\n

12.48<\/p>\n<\/td>\n<\/tr>\n

\n

8.97<\/p>\n<\/td>\n

\n

0.0<\/p>\n<\/td>\n

\n

1.0<\/p>\n<\/td>\n

\n

14.35<\/p>\n<\/td>\n

\n

14.47<\/p>\n<\/td>\n

\n

13.20<\/p>\n<\/td>\n<\/tr>\n

\n

9.23<\/p>\n<\/td>\n

\n

0.0<\/p>\n<\/td>\n

\n

1.0<\/p>\n<\/td>\n

\n

14.54<\/p>\n<\/td>\n

\n

14.55<\/p>\n<\/td>\n

\n

13.91<\/p>\n<\/td>\n<\/tr>\n

\n

9.5<\/p>\n<\/td>\n

\n

0.0<\/p>\n<\/td>\n

\n

1.0<\/p>\n<\/td>\n

\n

14.72<\/p>\n<\/td>\n

\n

14.62<\/p>\n<\/td>\n

\n

14.63<\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n\n\n\n
\"\"<\/td>\n\n

Figure 2: Domain shift across all four features, highlighting the differences in distributions and relationships that characterize the two domains. The blue dots (Domain A) and red crosses (Domain B)<\/strong><\/p>\n

<\/p>\n

Click here to view Figure<\/a><\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n\n\n

Case Study 3 : Application of Intuitionistic Fuzzy Sets in Medical Imaging for Tumor Detection<\/strong><\/p>\n\n\n\n

Step 1: Data Preprocessing<\/strong><\/p>\n\n\n\n

Normalization of Pixel Intensity Values<\/strong><\/p>\n\n\n\n

Normalize pixel intensity values I to a standard range, typically [0, 1], using the formula:<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

Segmentation of Medical Images<\/strong><\/p>\n\n\n\n

Apply segmentation algorithms such as Otsu’s method or k-means clustering to identify Regions of Interest (ROIs) within the medical images.<\/p>\n\n\n\n

Step 2: Fuzzy Logic for Interpretability<\/strong><\/p>\n\n\n\n

Definition of Fuzzy Membership Functions<\/strong><\/p>\n\n\n\n

Define membership functions \u03bc for key features x (e.g., pixel intensity, grayscale, texture coefficient). For example, a Gaussian membership function can be used:<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

Calculation of Membership Grades<\/strong><\/p>\n\n\n\n

Compute membership grades \u03bc(xi<\/sub>) for each pixel i<\/em> based on the defined fuzzy membership functions.<\/p>\n\n\n\n

Threshold Application for Classification<\/strong><\/p>\n\n\n\n

Apply a threshold \u03c4 to classify pixels as part of the tumor or not:<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

Step 3: Domain Adaptation<\/strong><\/p>\n\n\n\n

Identification of Source and Target Domains<\/strong><\/p>\n\n\n\n

Identify the source domain Ds<\/sub> <\/em>(training dataset) and target domain DT<\/sub><\/em>  (new clinical dataset).<\/p>\n\n\n\n

Transfer Learning<\/strong><\/p>\n\n\n\n

Fine-tune the pre-trained model Ms<\/sub>, from the source domain using labeled data from the target domain:<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

Domain-Invariant Feature Learning<\/strong><\/p>\n\n\n\n

Employ domain-invariant feature learning techniques to learn features f that are robust across both domains Ds<\/sub><\/em> and DT<\/sub><\/em> :<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

where D is a domain discriminator.<\/p>\n\n\n\n

Step 4: Model Evaluation<\/strong><\/p>\n\n\n\n

Cross-Validation on Target Domain<\/strong><\/p>\n\n\n\n

Perform k-fold cross-validation on the target domain DT<\/sub><\/em> to evaluate model performance. Calculate metrics such as accuracy, precision, recall, and F1 score.<\/p>\n\n\n\n

Interpretability Assessment<\/strong><\/p>\n\n\n\n

Use techniques like Class Activation Mapping (CAM) or LIME for visual explanations.<\/p>\n\n\n\n

Step 5: Incremental Data Integration:<\/strong><\/p>\n\n\n\n

Continuously integrate new data D(new)<\/sub><\/em> into the model to update its parameters:<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

Results and Discussion <\/strong><\/p>\n\n\n\n

The results from the case studies illustrate the effectiveness of the fuzzy logic framework in addressing key challenges in AI-based medical imaging, such as interpretability and domain shift. Fuzzy logic provides a transparent way to manage complex datasets, making it easier for clinicians to trust AI outputs. The weighted average defuzzification method ensured the accurate classification of tumor pixels, with higher defuzzified values correlating with tumor likelihood.<\/p>\n\n\n\n

For case study 1, the analysis helps in\nunderstanding and enhancing interpretability in AI for medical imaging,\ndemonstrating a practical application of fuzzy logic in a complex, real-world\nscenario. Intuitionistic fuzzy sets extend\ntraditional fuzzy sets by incorporating a degree of hesitancy, which offers a\nmore nuanced way to handle uncertainty and improve the robustness of AI models.\nThe following pictorial representation in Fig. 3 and 4 gives a visual\nunderstanding of case study 3.<\/p>\n\n\n\n

Normalization\nof pixel intensity values to a standard range [0, 1].<\/p>\n\n\n\n\n\n
\"\"<\/td>\n\n

Figure 3: Normalization of Pixel Intensity Values<\/strong><\/p>\n

<\/p>\n

Click here to view Figure<\/a><\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n\n\n

Segmentation\nof MRI scans to identify Regions of Interest (ROIs).<\/p>\n\n\n\n\n\n
\"\"<\/td>\n\n

Figure 4: Segmentation of Medical Images<\/strong><\/p>\n

<\/p>\n

Click here to view Figure<\/a><\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n\n\n

Intuitionistic\nfuzzy membership, non-membership, and hesitancy functions for key features.<\/p>\n\n\n\n

Calculation\nof membership, non-membership, and hesitancy grades for each pixel.<\/p>\n\n\n\n

Table 4: Calculation of Membership, Non-membership, and Hesitancy Grades<\/strong>.<\/p>\n\n\n\n\n\n\n\n\n\n\n
\n

Pixel Value<\/strong><\/p>\n<\/td>\n

\n

Membership (\u03bc)<\/strong><\/p>\n<\/td>\n

\n

Non-membership (\u03bd)<\/strong><\/p>\n<\/td>\n

\n

Hesitancy (\u03c0)<\/strong><\/p>\n<\/td>\n<\/tr>\n

\n

0.1<\/p>\n<\/td>\n

\n

0.0<\/p>\n<\/td>\n

\n

1.0<\/p>\n<\/td>\n

\n

0.0<\/p>\n<\/td>\n<\/tr>\n

\n

0.3<\/p>\n<\/td>\n

\n

0.14<\/p>\n<\/td>\n

\n

0.86<\/p>\n<\/td>\n

\n

0.0<\/p>\n<\/td>\n<\/tr>\n

\n

0.5<\/p>\n<\/td>\n

\n

1.0<\/p>\n<\/td>\n

\n

0.0<\/p>\n<\/td>\n

\n

0.0<\/p>\n<\/td>\n<\/tr>\n

\n

0.7<\/p>\n<\/td>\n

\n

0.14<\/p>\n<\/td>\n

\n

0.86<\/p>\n<\/td>\n

\n

0.0<\/p>\n<\/td>\n<\/tr>\n

\n

0.9<\/p>\n<\/td>\n

\n

0.0<\/p>\n<\/td>\n

\n

1.0<\/p>\n<\/td>\n

\n

0.0<\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n\n\n

Application\nof thresholds to classify pixels as tumor or non-tumor.<\/p>\n\n\n\n

Table 5: Threshold Application for Classification<\/strong><\/p>\n\n\n\n\n\n\n\n\n\n\n
\n

Pixel Value<\/p>\n<\/td>\n

\n

Membership (\u03bc)<\/p>\n<\/td>\n

\n

Non-membership (\u03bd)<\/p>\n<\/td>\n

\n

Hesitancy (\u03c0)<\/p>\n<\/td>\n

\n

Classification<\/p>\n<\/td>\n<\/tr>\n

\n

0.1<\/p>\n<\/td>\n

\n

0.0<\/p>\n<\/td>\n

\n

1.0<\/p>\n<\/td>\n

\n

0.0<\/p>\n<\/td>\n

\n

Non-Tumor<\/p>\n<\/td>\n<\/tr>\n

\n

0.3<\/p>\n<\/td>\n

\n

0.14<\/p>\n<\/td>\n

\n

0.86<\/p>\n<\/td>\n

\n

0.0<\/p>\n<\/td>\n

\n

Non-Tumor<\/p>\n<\/td>\n<\/tr>\n

\n

0.5<\/p>\n<\/td>\n

\n

1.0<\/p>\n<\/td>\n

\n

0.0<\/p>\n<\/td>\n

\n

0.0<\/p>\n<\/td>\n

\n

Tumor<\/p>\n<\/td>\n<\/tr>\n

\n

0.7<\/p>\n<\/td>\n

\n

0.14<\/p>\n<\/td>\n

\n

0.86<\/p>\n<\/td>\n

\n

0.0<\/p>\n<\/td>\n

\n

Non-Tumor<\/p>\n<\/td>\n<\/tr>\n

\n

0.9<\/p>\n<\/td>\n

\n

0.0<\/p>\n<\/td>\n

\n

1.0<\/p>\n<\/td>\n

\n

0.0<\/p>\n<\/td>\n

\n

Non-Tumor<\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n\n\n

Domain\nadaptation process to ensure robustness across different clinical environments.\nCross-validation and evaluation of model performance using various metrics.\nContinuous learning and performance monitoring to maintain optimal model\naccuracy. Classification results of pixels in the MRI scan using Intuitionistic\nfuzzy sets is shown in Fig. 5.<\/p>\n\n\n\n\n\n
\"\"<\/td>\n\n

Figure 5: Tumor Regions Highlighted<\/strong><\/p>\n

<\/p>\n

Click here to view Figure<\/a><\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n\n\n

Alternative Methods and Hypotheses<\/strong><\/p>\n\n\n\n

While fuzzy logic enhances interpretability, other\nAI techniques, such as explainable neural networks and decision trees, also aim\nto provide transparent decision-making. Methods like SHAP (Shapley Additive\nExplanations) and LIME (Local Interpretable Model-Agnostic Explanations) offer\nfeature-level insights but may struggle with generalization across diverse\ndatasets. Additionally, non-fuzzy models may outperform fuzzy logic in cases\nwhere pixel-level precision is less critical, such as in broad pattern\nrecognition tasks.<\/p>\n\n\n\n

Limitations<\/strong><\/p>\n\n\n\n

One limitation of our approach is the reliance on\npredefined membership functions and threshold values, which may need adjustment\nfor different imaging modalities. Moreover, the framework has been tested\nprimarily on a single case study with simulated data, meaning further\nvalidation on real-world clinical datasets is required for broader\napplicability.<\/p>\n\n\n\n

Uncertainties and Sensitivity of Results<\/strong><\/p>\n\n\n\n

The sensitivity of the fuzzy logic framework to\nchanges in membership grades and defuzzification parameters presents a\npotential source of variability. Small variations in these parameters may alter\nthe classification outcome, highlighting the need for robust parameter\noptimization. Additionally, domain shift remains a challenge, as the model\u2019s\naccuracy may decrease in clinical settings with significant differences in\nequipment or patient demographics. Continuous fine-tuning and adaptive learning\nmechanisms are necessary to mitigate these effects.<\/p>\n\n\n\n

Implications for Clinical Practice<\/strong><\/p>\n\n\n\n

The application of fuzzy logic provides a promising pathway for improving the interpretability and trustworthiness of AI models in medical imaging. However, collaboration between data scientists and clinicians will be essential to tailor these models to specific clinical workflows. Future work should focus on integrating fuzzy logic with other adaptive AI techniques to further enhance model robustness and generalizability across diverse clinical environments.<\/p>\n\n\n\n

Domain Shift<\/strong><\/p>\n\n\n\n

Understanding Domain Shift in AI<\/strong><\/p>\n\n\n\n

Domain shift refers to\nthe switch in data distribution that a machine learning model encounters when\napplied to new environments or scenarios different from the training data. This\nphenomenon is critical in AI, as it can remarkably impact a model’s performance\ndue to the distinction between the source (training) and target (application)\ndomains.<\/p>\n\n\n\n

Impacts on Medical Imaging Analysis<\/strong><\/p>\n\n\n\n

In medical imaging,\ndomain shift can be extremely challenging, as models trained on data from one\nset of equipment or demographic might not execute well when applied to data\nfrom another, due to difference in image acquisition protocols, patient\npopulations, or disease prevalence. This can lead to decreased precision in\nautomated diagnosis systems and likely impact patient outcomes.<\/p>\n\n\n\n

To address domain shift,\nvarious domain adaptation strategies have been developed. These include\ntransfer learning, where a model trained on one domain is redesign to another,\nand domain-invariant feature learning, which targets to learn aspect that are\nrobust to the switch between domains. Additionally, data augmentation and\nsynthetic data generation can be employed to simulate a variety of domain\nshifts during the training process, strengthen the model’s generalizability.<\/p>\n\n\n\n

Adaptive AI Systems in Medical Imaging<\/strong><\/p>\n\n\n\n

Continuous Learning and Model Updating<\/strong><\/p>\n\n\n\n

In medical imaging,\nadaptive AI systems must incorporate machine learning techniques such as online\nlearning, where the model is incrementally trained on new data, or transfer\nlearning, where a pre-trained model is fine-tuned with data from a new domain.\nThis enables the model to reshape to new patterns in the data, such as novel\nimaging biomarkers or advancing disease presentations.<\/p>\n\n\n\n

Cross-Modality and Cross-Institutional Transformation <\/strong><\/p>\n\n\n\n

These arrangements require deploying domain\nvariation techniques to ease issues resulting from contrast in imaging\nmodalities example CT, MRI, PET and institutional practices. This may encompass\nthe use of GANs- generative adversarial networks to execute image-to-image\ntranslation, authenticating model robustness and transferability across\nnumerous imaging technologies and healthcare framework. <\/p>\n\n\n\n

Trustworthiness<\/strong><\/p>\n\n\n\n

Reliability in AI\nsystems, particularly in medical imaging, needs a multifaceted approach, as\ndepicted by Fig 6.<\/p>\n\n\n\n\n\n
\"\"<\/td>\n\n

Figure 6: A reliability framework for developing trustworthiness<\/strong><\/p>\n

<\/p>\n

Click here to view Figure<\/a><\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n\n\n

Current Regulatory Landscape<\/strong><\/p>\n\n\n\n

A complex international\nguidelines, national laws, and industry standards defines the ongoing\nregulatory landscape for AI in medical imaging. In the U.S., the Food and Drug\nAdministration (FDA) give oversight through its regulatory framework for\nSoftware as a Medical Device (SaMD), which includes AI and machine\nlearning-based software. The FDA’s risk-based approach target on the software’s\nintended use. In Europe, the European Union’s Medical Device Regulation (MDR)\nclassifies and regulates AI as a medical device, with an identical risk-based\napproach. Both frameworks demand strict clinical evaluation, post-market\nsurveillance, and a quality management system aligned with standards such as\nISO 13485. Various factors affecting this is shown in Fig 7. AI is not a threat\nbut a tremendous opportunity to assist radiologists in quickening the backend\nprocesses, improving workflow, increasing accuracy, and quantification of\nfindings20<\/sup>. Borys K studied a common ground for cross-disciplinary\nunderstanding and exchange across disciplines between deep learning builders\nand healthcare professionals21<\/sup>.<\/p>\n\n\n\n\n\n
\"\"<\/td>\n\n

Figure 7: Various factors in regulatory and security landscape<\/strong><\/p>\n

<\/p>\n

Click here to view Figure<\/a><\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n\n\n

Conclusion<\/strong><\/p>\n\n\n\n

In this study, we used a fuzzy based framework to\naddress key challenges in AI-driven medical imaging, particularly in improving\ninterpretability, handling domain shifts, and enhancing trustworthiness in\nMRI-based tumor detection. Our case study demonstrated that applying fuzzy\nmembership grades and weighted average defuzzification techniques can\neffectively classify tumor pixels, offering clinicians a more transparent\ndecision-making tool compared to conventional AI methods. This approach\nunderscores the potential of fuzzy logic to bridge the gap between AI\u2019s\n\u2018black-box\u2019 nature and the need for explainability in clinical environments.<\/p>\n\n\n\n

However, it is essential to recognize that this\nresearch was limited to a single case study focusing on MRI scans. The\npredefined fuzzy membership functions and threshold values may need to be\nadjusted for different imaging modalities and clinical datasets. Furthermore,\nwhile the framework showed promise in managing domain shifts between source and\ntarget datasets, its robustness across varied real-world clinical settings\nremains to be tested.<\/p>\n\n\n\n

Future research should focus on extending this framework to larger and more diverse datasets, testing it across other medical imaging techniques such as CT and PET scans, and refining the fuzzy parameters to suit different clinical environments. By integrating fuzzy logic with adaptive learning models and domain-invariant feature extraction, the applicability of this approach could be broadened significantly. These advancements would contribute not only to better AI interpretability but also to the establishment of more trustworthy, reliable AI systems in medical diagnostics. The future of AI in medical imaging holds promising research avenues, including:<\/p>\n\n\n\n

In both case studies, a number of additional features may be given, and the threshold value can be adjusted and varied from case to case.<\/p>\n\n\n\n

Focusing on developing interpretable AI systems that provide transparency in decision-making processes and ethical considerations in deployment.<\/p>\n\n\n\n

Investigating the use of AI in integrating different imaging modalities to provide a comprehensive view of patient pathology.<\/p>\n\n\n\n

While fuzzy logic offers a promising pathway for\naddressing some of the core challenges in AI-based medical imaging, further\nempirical studies are required to validate its generalizability and clinical\nutility. The framework presented here serves as a foundational step toward\ndeveloping interpretable and trustworthy AI solutions that can be seamlessly\nintegrated into clinical workflows, ultimately benefiting patient care and\nclinical decision-making.<\/p>\n\n\n\n

Acknowledgments<\/strong><\/p>\n\n\n\n

The authors would like to express their sincere gratitude to their organization for providing the necessary resources and support to conduct this research. Special thanks are due to Rashmi Singh (Amity Institute of Applied Sciences, Amity University Uttar Pradesh, Noida, India), Aryan Chaudhary (Bio Tech Sphere Research, India), and Samrat Ray (IIMS Pune) for their valuable contributions. <\/strong><\/p>\n\n\n\n

Funding sources<\/strong><\/p>\n\n\n\n

The author(s) received no financial support for the research, authorship, and\/or publication of this article.<\/p>\n\n\n\n

Conflict of Interest<\/strong><\/p>\n\n\n\n

The author(s) do not have any conflict of interest.<\/p>\n\n\n\n

Data Availability<\/strong> Statement<\/strong> <\/p>\n\n\n\n

This statement does not apply to this article <\/strong><\/p>\n\n\n\n

Ethics Statement<\/strong><\/p>\n\n\n\n

This research did not involve human participants, animal subjects, or any material that requires ethical approval<\/p>\n\n\n\n

Informed Consent Statement<\/strong><\/p>\n\n\n\n

This study did not involve human participants, and therefore, informed consent was not required<\/p>\n\n\n\n

Clinical Trial Registration<\/strong><\/p>\n\n\n\n

This research does not involve any clinical\ntrials<\/p>\n\n\n\n

Author Contributions<\/strong><\/p>\n\n\n\n

Rashmi Singh<\/strong>:\nConceptualization, Methodology, Writing \u2013 Original Draft Preparation,\nVisualization, Validation, Data Interpretation.<\/p>\n\n\n\n

Aryan Chaudhary<\/strong>:\nConceptualization, Formal Analysis, Writing \u2013 Review & Editing, Validation,\nResources.<\/p>\n\n\n\n

Samrat Ray<\/strong>: Investigation, Ethics, Writing \u2013 Review & Editing, Resources.<\/p>\n\n\n\n

References<\/strong><\/p>\n\n\n\n

  1. Oren O, Gersh BJ, Bhatt DL. Artificial intelligence in medical imaging: switching from radiographic pathological data to clinically meaningful endpoints. Lancet Digit Health.<\/em> 2020; 2(9): E486-E488.
    CrossRef<\/a><\/li>
  2. Prevedello LM, Halabi SS, Shih G, Carol C. Wu, Kohli M, Chokshi F, Erickson B, Cramer J, Andriole K, Adam E. Flanders Challenges related to artificial intelligence research in medical imaging and the importance of image analysis competitions. Radiol Artif Intell.<\/em> 2019;1(1): 1-7.
     CrossRef <\/a><\/li>
  3. Tang X, The role of artificial intelligence in medical imaging research, BJR Open<\/em>, 2019; 2(1): 1-5.
    CrossRef <\/a><\/li>
  4. Panayides AS, Amini A, Filipovic ND, Sharma A, Tsaftaris SA, Young A, Foran D, Do N, Golemati S, Kurc T, Huang K, Nikita KS, Veasey BP, Zervakis M, Saltz JH and Pattichis CS, AI in medical imaging informatics: current challenges and future directions, IEEE J Biomed Health Inform<\/em>, 2020; 24(7): 1837-1857.
     CrossRef <\/a><\/li>
  5. Wang S, Cao G, Wang Y, Liao S, Wang Q, Shi J, Li C and Shen D, Review and prospect: artificial intelligence in advanced medical imaging, Front Radiol<\/em>, 2021; 1: 781868; 1-18.
    CrossRef <\/a><\/li>
  6. Saw SN and Ng KH, Current challenges of implementing artificial intelligence in medical imaging, Phys Med<\/em>, 2022; 100: 12-17.
    CrossRef <\/a><\/li>
  7. Gupta NS and Kumar P, Perspective of artificial intelligence in healthcare data management: a journey towards precision medicine, Comput Biol Med<\/em>, 2023; 162: 107051: 1-19.
    CrossRef <\/a><\/li>
  8. Al Kuwaiti A, Nazer K, Al-Reedy A, Al-Shehri S, Al-Muhanna A, Subbarayalu AV, Al Muhanna D and Al-Muhanna FA, A review of the role of artificial intelligence in healthcare, J Pers Med<\/em>, 2023; 13(6): 951, 1-22.
     CrossRef <\/a><\/li>
  9. Tekke\u015fin A\u0130, Artificial intelligence in healthcare: past, present and future, Anatol J Cardiol<\/em>, 2019; 22(Suppl 2): 8-9.
    CrossRef <\/a><\/li>
  10. Schaefferkoetter J, Evolution of AI in medical imaging, In: Veit-Haibach P, Herrmann K, eds, Artificial Intelligence\/Machine Learning in Nuclear Medicine and Hybrid Imaging<\/em>, Springer International Publishing, 2022: Chapter 4, 37-56.
    CrossRef <\/a><\/li>
  11. Rajpurkar P and Lungren MP, AI in medicine: the current and future state of AI interpretation of medical images, N Engl J Med<\/em>, 2023; 388(21): 1981-1990.
     CrossRef <\/a><\/li>
  12. Singh R, Bhardwaj N and Islam SMN, Applications of mathematical techniques to artificial intelligence: Mathematical methods, algorithms, computer programming and applications, In: Advances on Mathematical Modeling and Optimization with Its Applications<\/em>, Springer, 2024: 152-169.
    CrossRef <\/a><\/li>
  13. Singh R, Bhardwaj N and Islam SMN, The role of mathematics in data science: Methods, algorithms, and computer programs, Bentham Science Publishers, 2023, 1-23.
    CrossRef <\/a><\/li>
  14. Singh R, Khurana K and Khandelwal P, Decision-making in mask disposal techniques using soft set theory, In Lecture Notes in Electrical Engineering<\/em>, 2023; 968: 649-661.
    CrossRef <\/a><\/li>
  15. Singh R and Bhardwaj N, Fuzzy soft set theory applications in medical diagnosis: A comprehensive review and the roadmap for future studies, New Math Nat Comput<\/em>, 2024; 20(1): 1-23.
     CrossRef <\/a><\/li>
  16. Marey A, Arjmand P, Alerab A, Eslami M, Saad A, Sanchez N & Umair M, Explainability, transparency and black box challenges of AI in radiology: impact on patient care in cardiovascular radiology, Egypt J Radiol Nucl Med<\/em>, 2024; 55(183): 1-14.
    CrossRef <\/a><\/li>
  17. Vimbi V, Shaffi N and Mahmud M, Interpreting artificial intelligence models: a systematic review on the application of LIME and SHAP in Alzheimer\u2019s disease detection, Brain Inform<\/em>, 2024; 11:10, 1-29.
    CrossRef <\/a><\/li>
  18. Kondylakis H, Kalokyri V, Sfakianakis S, Marias K, Tsiknakis M, Jimenez-Pastor A, Camacho-Ramos E, Blanquer E, Segrelles J, L\u00f3pez-Huguet S, Barelle C, Kogut-Czarkowska M, Tsakou G, Siopis N, Sakellariou Z, Bizopoulos P, Drossou V, Lalas A, Votis K, Mallol P, Bonmati L, Alberich L, Seymour K, Boucher S, Ciarrocchi E, Fromont L, Rambla J, Harms A, Gutierrez A, Martijn P. A. Starmans, Fred Prior, Josep Ll. Gelpi , Lekadir   K, Data infrastructures for AI in medical imaging: experiences of five EU projects, Eur Radiol Exp<\/em>, 2024; 7:20, 1-13.
    CrossRef <\/a><\/li>
  19. Ahmed A and Hussein M, Intelligent saline controlling valve based on fuzzy logic, J Eng Appl Sci<\/em>, 2024; 71:163, 1-21.
    CrossRef <\/a><\/li>
  20. Kohli A, AI in medical imaging: current and future status-artificial intelligence or augmented imaging? Indian J Radiol Imaging<\/em>, 2021; 31(3): 525-526.
    CrossRef <\/a><\/li>
  21. Borys K, Schmitt YA, Nauta M, Seifert C, Kr\u00e4mer N, Friedrich CM and Nensa F, Explainable AI in medical imaging: an overview for clinical practitioners – beyond saliency-based XAI approaches, Eur J Radiol<\/em>, 2023; 162: 110786, 1-11.
     CrossRef <\/a><\/li><\/ol>\n","protected":false},"excerpt":{"rendered":"

    Introduction The Amalgamation of Artificial Intelligence into medical imaging streak […]<\/p>\n","protected":false},"author":15,"featured_media":0,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[125],"tags":[],"class_list":["post-62821","post","type-post","status-publish","format-standard","hentry","category-vol18marchspledition"],"_links":{"self":[{"href":"https:\/\/biomedpharmajournal.org\/staging\/wp-json\/wp\/v2\/posts\/62821","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/biomedpharmajournal.org\/staging\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/biomedpharmajournal.org\/staging\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/biomedpharmajournal.org\/staging\/wp-json\/wp\/v2\/users\/15"}],"replies":[{"embeddable":true,"href":"https:\/\/biomedpharmajournal.org\/staging\/wp-json\/wp\/v2\/comments?post=62821"}],"version-history":[{"count":5,"href":"https:\/\/biomedpharmajournal.org\/staging\/wp-json\/wp\/v2\/posts\/62821\/revisions"}],"predecessor-version":[{"id":64824,"href":"https:\/\/biomedpharmajournal.org\/staging\/wp-json\/wp\/v2\/posts\/62821\/revisions\/64824"}],"wp:attachment":[{"href":"https:\/\/biomedpharmajournal.org\/staging\/wp-json\/wp\/v2\/media?parent=62821"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/biomedpharmajournal.org\/staging\/wp-json\/wp\/v2\/categories?post=62821"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/biomedpharmajournal.org\/staging\/wp-json\/wp\/v2\/tags?post=62821"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}