Explainable AI in Diagnostic Radiology for Neurological Disorders – a Review

1. Introduction

1.1. Neurological Disorders – Morbidity and Mortality

According to World Health Organization (WHO), neurological disorders are among the top three killers in the world, even in developed countries with adequate healthcare infrastructure [1]. Neurodegenerative disorders like Alzheimer’s disease (AD) and Parkinson’s disease (PD) can drastically degrade the quality of life. Unlike other parts of the body, brain cells do not regenerate, and hence early diagnosis is of paramount significance to contain the progression of such diseases [2,3,4,5,6]. However at the early stage of such disorders, the diagnosis is challenging [2], due to very subtle changes in medical imaging data. In addition, some diseases in their infancy may present similar or overlapping findings, specially in case of neurodegenerative disorders, thereby rendering the differential diagnosis even more tedious [7,8,9]. This would require highly skilled and experienced medical professionals for accurate diagnoses to choose the right course of treatment and contain the prognoses. The unavailability of such experts and lack of infrastructure, specially in under-developed countries can produce catastrophic outcomes.

1.2. AI in CAD

AI has revolutionized every walk to life, and healthcare and medicine are no exceptions [6,10,11]. Over the past couple of decades, tremendous research has been observed for the design and development of CAD tools for multi-modal data to act as assistants to domain experts in reaching fast and concrete diagnoses. From old-school Computer Vision (CV) algorithms, to Machine Learning (ML), to the more recent DL architectures have been observed making their way in this field and producing outstanding results [6,10,12,13,14,15]. Radiological imagining has become an inseparable part of the diagnosis process with technologies like X-ray, ultrasound, MRI, CT, PET scans playing a vital role in assisting experts in accurate diagnoses. MRI, being a non-invasive and highly informative modality is found to be widely employed [13,16,17]. DL powered machines have the ability to look into the most intricate features even up to pixel levels, which the human-eye might overlook. In this domain, Convolutional Neural Networks (CNNs) have recently shown unprecedented promise [10,18,19,20] and have been employed in systems with multiple disease diagnosis capabilities [16]. Some researchers have even observed that the performance of such DL powered systems are comparable to human in real-world tasks, and in some cases might surpass human domain experts’ performance in terms of speed and accuracy [10,20,21]. But this increased performance and accuracy of DL powered systems comes at a price.

Before taking the discussion any further, we would like to take a moment here to introduce some standard terminologies pertaining to XAI.

• Interpretability: Interpretability refers to the ability to understand the decision-making process of an AI model. The operation of an interpretable model is transparent and provides details about the relationships between inputs and outputs.
• Explainability: Explainability refers to the ability of an AI model to provide clear and intuitive explanations of the decisions made to the end user. In other words, an explainable AI model provides justification of the decisions made.
• Transparency: Transparency refers to the ability of an AI model to provide a view into the inner workings of the system, from inputs to inferences.
• Black box: Block box model in AI is one whose operations are not visible to the user. Such models arrive at decisions without providing any explanation as to how they were reached. Such models lack transparency, and therefore are frowned upon, and not trusted in applications like diagnostic medicine, where precious human lives are on the line.

The journey with AI started with simpler rule-based algorithms in ML, like decision trees, which provided clear rules for end-users to understand the reasons of classifications. The features in ML powered systems were hand-crafted by developers [17] and hence such systems offered higher levels of transparency in their inferences and decisions. The accuracy of such systems was relatively low. In order to increase accuracy and performance, complex DL architectures were developed with many hidden layers and millions of trainable parameters. The features in such systems were extracted implicitly and hence the opacity increased drastically [22]. Although there were leaps in accuracy, the decision-making process got wrapped in a “Black Box” [6,10,11,15,21,23,24,25,26,27,28], resulting in the decrease in trust specially in highly sensitive fields like diagnostic medicine, where a single wrong decision can be a matter of life and death [6,20,28,29,30,31]. This trade-off between accuracy and explainability [6,14,15,32] in the evolution of AI is shown in Figure 1. This has been one of the leading factors that despite the tremendous performance peaks achieved by such DL powered CAD tools, they have not yet been able to find their way in the routine medical practices since both the doctors and the patients by all means demand their right to know the reasons of a particular diagnosis/ inference generated by such CAD tools [20,29]. This gave birth to a recent sub-domain of AI called XAI.

1.3. Unravelling the Mystery!

XAI is an attempt by engineers to demystify the otherwise secretive working of complex DL architectures popularly referred to as the “Black Boxes”, which is the leading cause of mistrust of people [29,33]. Figure 2 shows the end-to-end working of an XAI powered DL regimen (from training to deployment), along with the tentative expressions and comfort level of domain experts at various stages, in a block diagram.

Many theories, methods and frameworks have been devised to provide plausible explanations to the outcomes of such models.

1.4. XAI Methods and Frameworks

XAI methods have been classified into various categories and many respective frameworks have been developed over a brief period of time [29], but going in their details is beyond the purview of this work and hence we will state them succinctly.

The generated explanations can be textual, numerical, visual or example based. In the case of radiological imaging, most researchers have worked with visual explanations in the form of heatmaps highlighting the regions in the input images contributing towards a particular inference [26]. From the point of view of scope, the explainability of a model can either be local or global. A global explanation explains the behavior of an entire model on all the input dataset, whereas a local explanation might just use a couple of examples to help explain why certain decisions were made. From the point of view of the stage of implementation, XAI methods can be categorized as ante-hoc and post-hoc, depending on whether the explanations were generated during or after training. From the point of view of applicability of XAI methods, we have model-specific and model-agnostic approaches. Model-agnostic approaches are generally easy to implement, since their applicability is independent of the underlying AI model, and are also sometimes referred to as plug-and-play [17]. The remaining part of this section briefly introduces the most commonly employed XAI frameworks.

Local Interpretable Model-Agnostic Explanations (LIME) generates explanations by perturbing the input data to a model, and observing the changes in the output. It can thus highlight the significant features in input data responsible for a particular decision. SHapley Additive exPlanations (SHAP) assigns weights to all input features and observes the outcomes of all weighted input feature combinations. Gradient-weighted Class Activation Mapping (Grad-CAM) works with CNNs to identify important regions in an input image which are responsible for the inference. It is applied using gradient information of the output layer to produce a heatmap for the input image. Layer-wise Relevance Propagation (LRP) also generates similar heatmaps by assigning relevance scores to all neurons in the output layer of a CNN, and then backpropagates to the input layer, while computing scores for every neuron. Occlusion Sensitivity Analysis (OSA) and Saliency Maps (SM) are frequently used schemes to generate visual explanations for image input data. Their mode of operation is similar. In OSA, patches of images are occluded periodically and the corresponding effects on the outputs are observed. If the probability of a certain prediction drops drastically by occluding a certain input image patch, it would signify that patch to hold important information for that particular prediction. SM also generates heatmaps in a similar manner. The difference is that in case of SMs, each pixel from the input is removed iteratively and the corresponding drop in probability of inference defines its importance. Hence, the heatmap generated by SM contains all the significant pixels responsible for a certain prediction. In case of a CAD tool for neurological disorders, for example, all brain MRI regions responsible for the diagnosis of a certain disease would light up for experts to visualize the reason/ explanation of that particular diagnosis by the tool. In order to elaborate visual explanations of the XAI frameworks for clear understanding, a simple brain tumor dataset was downloaded from Kaggle [34], and a 2 class DL model was trained for classification of MRI images as ‘Tumor’ or ‘No tumors’. The visual explnations generated by different XAI techniques are shown in Figure 3.

1.5. Would XAI Be the Matchmaker~

The next important questions are:

“Would the integration of explainability to the otherwise opaque DL architectures (rejected by doctors and patients [35]) bridge the gap and develop the trust of domain experts in using CAD tools~”

“Would only visual explanations be enough for experts in diagnostic radiology~”

“What else should be done to pave the way for the large scale incorporation of AI in diagnostic medicine~”

With the help of literature and medical experts, we will be looking for answers to these and other such questions in this humble endeavor.

1.6. Study Selection

Google scholar was used to search and collect relevant studies. The search phrases used include “explainable AI in brain MRI based computer-aided diagnoses”. Research conducted was scanned collectively from 2017 to 2023, other than that for the year 2024, which was scanned individually using filters, and the titles of articles included in first 10 pages were examined and the most relevant were downloaded, only those with full-text access. After the removal of duplicates, these articles were later screened by abstracts. This was followed by thorough examination of full texts. The Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines were followed in the process, as shown in Figure 4. From the articles reviewed in this study, the following information was extracted and summarized:

• Year of study
• Diseases researched
• Modalities employed
• AI techniques used
• Accuracy of developed systems
• Algorithms used for Explainability
• Datasets used

The rest of this paper is organized in such a way that prior art pertaining to brain MRI based CAD of neurological disorders using both AI and XAI is given in section II. The challenges ahead are presented in section III, and the paper is concluded in section IV which presents the answers we seek.

2. AI and XAI in CAD of Neurological Disorders

Despite the next to negligible penetration of AI in current routine healthcare regimens (due to its multitude of limitations and unreliability), the recent massive research, progress and development in CAD from data science perspective is sincerely praiseworthy. This section highlights the marvels achieved by this technology in diagnostic medicine (radiology, to be specific) in its journey from black box AI to the more recent and transparent XAI, for CAD of neurological disorders.

Numerous systems have been developed ranging from recognizing MRI sequences and view planes [36], preprocessing [37], segmentation of brain regions or anomalies [38] like tumors [39], to diagnosing disorders from a given MRI in multiclass problems [40]. This has only been possible mainly due to a couple of factors. The most important among them is the availability of massive, open access, publicly available, labelled datasets for data scientists and engineers to develop and train complex frameworks, rendering them capable of producing accurate inferences from unseen data in real-time. In addition, the very recent boost in the storage and processing capabilities of our machines to perform rapid calculations with millions of trainable parameters in the deep and complex models has been a gamechanger for this technology [21,30]. Hats off to the drastic evolution of Graphics Processing Units (GPUs) that made this possible. DL is known to be data hungry [11]. That is, it can produce better results if trained on massive annotated/ labelled data [41]. In case of brain MRI, the annotations/ labels coming from experts (consultant radiologists/ neurologists) are considered to be the “Gold Standard”. However, this can become an extremely time-consuming and tedious job [42] given the magnitude of data available online [43]. Moreover, incorrectly labelled data can obviously lead to poor training which in turn results in poor accuracy of models. Some pathologies are better visualized in specific MRI sequences as compared to others. For example, demyelinating diseases like Multiple Sclerosis (MS) and Neuromyelitis Optica (NMO) produce plaques/ lesions on the brain which are prominent as hyper-intense regions on a FLAIR (Fluid attenuated inversion recovery) sequence [43,44]. So, a study on the CADD of MS and NMO should technically focus on FLAIR MRI. To assist such applications, various systems have been developed to automatically identify the sequence and view planes of MRI scans [45,46,47,48,49]. In addition, the extra-cranial tissues including the skull, eyes, neck, etc. can be a source of noise for an AI system being designed to assist in differential diagnosis from the brain. To help in such scenarios, various brain extraction tools (skull-strippers) have been developed with extraordinary (radiologist-like) capabilities to handle all MRI sequences and orientations. Such preprocessing techniques (Synthstrip [37], NeuroImaging Volumetric Extractor – NIVE [50]) have been found to increase the accuracies of CAD tools. Two main problems that AI is generally found to be working on are segmentation and classification. From the literature, AD and Brain Tumors were the most widely researched disorders in developing AI powered CAD tools using brain MRIs as inputs. DL architectures have been widely used to segment tumors and lesions from MRI scans. The analysis of texture and morphology of such tumors and lesions can further lead to accurate diagnoses. For example, the lesions appearing in MS and NMO on the brain can appear very similar [51]. A concrete differential diagnosis from this modality alone can hence be extremely challenging [52,53,54], resulting in delays incurred due to additional testing. AI can assist in reducing such overheads of cost and time. 2D/ 3D DL architectures, specially CNNs have been found to be vastly employed in multiple disease CAD systems [55]. Systems have been found claiming to classify even up to 35 diseases [56]. The major issue in such systems reported by literature is the generalizability. That is, although some of such systems have been reported to have an accuracy even up to 100% for classifying multiple diseases and their sub-types [57,58,59,60,61,62], these systems fail to be as accurate when tested on unseen data, from different sources, not used in training. In addition, the research discussed thus far has no embedded explainability, which means that the end-user (doctor) has no idea of what is going on within the DL model and what was the reason for a particular diagnosis. The opacity of such systems is also one of the leading reasons of mistrust between the healthcare professionals and the AI systems, and hence demands massive attention. From this point onwards, this section discusses XAI powered CAD tools which is an attempt by researchers and data scientists to bridge the gap between AI and healthcare, by making CAD systems transparent, thereby building the confidence of medical experts in these assistive tools.

XAI has very recently emerged as a sub-domain of AI to assist domain experts in diagnoses and prognoses. From statistical point of view of incorporation of explainability in CAD systems using medical imaging data, X-rays have been widely researched, closely followed by MRI [28]. Among MRI studies, structural/ anatomical MRI have been widely used followed by functional MRI [17]. Chest and brain are the top two researched anatomical locations [17], followed by eyes and breasts [28]. From the point of view of explanations, over a period from 2017 to 2020, visual explanations have been employed the most with a constantly increasing trend, followed by textual and example-based explanations [28]. In visual explanations domain, perturbation methods have been found dominating the XAI algorithms [10]. Among XAI techniques, CAM and Grad-CAM are found to be leading, followed by LRP, guided backpropagation, LIME and SHAP, among others [17]. In XAI powered CAD research, AD and brain tumors are widely researched, followed by PD and others.

The study in [3] employs CNN and LIME for the diagnosis of AD. The system has been trained using ADNI dataset and claims to have a classification accuracy of 94.96%. Many similar studies [4,5,19,63,64,65,66,67,68,69] were found using different DL and XAI architecture for the diagnosis of AD, some handling its sub-types as well. The common problem with these systems was relying on ADNI dataset only for training and testing their systems. Such systems in practical scenarios are bound to suffer from drastic accuracy reduction due to poor generalizability. Moreover, they fail to provide any concrete quantitative or qualitative analysis of the explanations generated by their systems. Additionally, no doctors (domain experts and the ultimate end-users) were found to be onboard for evaluation. Similar issues and limitations were found in brain tumor research [12,16,22,70,71,72,73,74,75]. Camacho et.al. in [2] present their work on explainable classification of PD. They use a large multi-center database of T1-weighted (T1w) MRIs to train their CNN model with saliency maps to identify the regions responsible for inferences. Employment of data from about 14 centers caters to the generalizability issue, but the other questions pertaining to explainability still remain unanswered.

Few studies propose the use of multimodal data for prediction and management of disorders. This offers extra diagnostic features which may support and enhance the accuracy of diagnoses. Jahan et.al. in [33] use clinical, physiological, and MRI data for five class classification of AD using OASIS dataset. They compare the performance of 9 popular ML models and employ SHAP for explainability. Their study finds Random Forest (RF) as the best suited classifier for this job with a 10-fold cross-validation accuracy of 98.81%. The research in [13] proposes concatenation of PET and MRI images for the diagnosis of AD using ResNet18. There explainable system claims to have an accuracy of 73.90% using ADNI dataset. Kamal et.al. in [76] present fusion of brain MRI with Gene Expression Data for AD classification into 4 categories. They use CNN and LIME to achieve an accuracy of 97.6%. Similarly in the study conducted in [77], integration of 11 modalities, including PET, MRI, cognitive scores, genetic data, demographic data, patient history, CSF, neuropsychological battery, lab tests, etc., for classification of AD can be observed. The use of multimodal data may certainly enhance the potential of accurate diagnoses by these CAD tools, yet no evaluation criteria of explanations generated in these systems have been found.

Some studies were found to focus on guidelines for evaluation of clinical XAI in medical image analysis. Jin et.al. in [27] propose guidelines for choosing an XAI technique based on understandability, clinical relevance, truthfulness, informative plausibility and computational efficiency. They implemented and evaluated 16 commonly used post-hoc heatmap XAI techniques including gradient-based and perturbation-based techniques. With focus on two tasks, (i) to classify gliomas as low-grade (LGG) and high-grade gliomas (HGG) using 3D CNN architecture and BraTs 2020 dataset, and (ii) knee lesion identification task, they conclude that all 16 XAI techniques were inadequate for clinical use due to their failure in truthfulness and informative plausibility. Another study in [31] propose the modality-specific feature importance (MSFI) metric to evaluate feature attribution maps on multi-modal medical images. They also highlight that literature review indicates 35% studies evaluated the explanations with computational metrics only, 8% involved medical experts to verify explanation plausibility either quantitatively or qualitatively, whereas only a meager 5% employed both modes of evaluation.

Very few articles were found from literature incorporating medical experts’ feedback to assess their proposed systems’ clinical utility. The system proposed in [24] employs hybrid vision transformers and CNNs for glioma segmentation in brain MRI scans. They also claim to provide surgeon-understandable heatmaps to make their system transparent. In addition, they conducted structured interviews with two medical experts from Neurosurgery department at Ulm university hospital to evaluate the practical utility of their developed system. Their discussions included evaluation of model performance with actual patient cases, and interpretability of the model’s decision-making process with respect to the clinical experience of the neurosurgeons. According to the authors, experts found Grad-CAM to be a valuable tool to introduce explainability/ transparency in the otherwise opaque DL regimens. Another study in [78] employs Subtractive Spatial Lightweight CNN to classify malignant tumors as Medulloblastomas, Ependymomas, Meningiomas, Lymphomas, and Anaplastic Astrocytomas. They included CAM for explainability. After achieving a reasonable classification accuracy of up to 93.33% in the first evaluation phase, they took 10 doctors onboard to judge the meaningfulness of the generated explanations in the second phase. Each doctor went through a total of 120 MRI images and colored the tumor regions manually. These colorings were then compared with the heatmaps generated by CAM. According to the authors, the overlap never went below 98%, with it being 100% in the majority of the cases. They also conducted a survey from the doctors towards the end of their study with 11 statements like “The system is trustworthy in terms of diagnosis of brain tumors”. I am able to understand well about the detection by looking at the heatmaps”. I am able to make decision faster thanks to XAI view of the system” and “I want to use this system for auto-decision-making in brain tumor diagnosis”, which were rated by doctors from a scale of 1 to 5; with 5 being the strongest agreement. The average rating from all 10 doctors who participated in this study is above 4.5, with 4.8 being the highest. We think that the input from medical domain experts on XAI powered CAD research is an absolute nessicity, but is severly lacking at the moment as evident from literature. The requirements of doctors specially from the point of view of mode of explanations (visual, textual, etc.), among others, must be identified in the first phase, and incorporated in such CAD tools in the subsequent phases followed by rigourous testing and validation from medical experts in clinical settings.

A summary of the recent XAI powered CAD systems is given in Table 1. This table summarizes information regarding the disease diagnosing capability, the modality used as input, the AI and XAI technologies employed, the datasets used for training, and the accuracy achieved by the CAD tools.

3. The Challenges Ahead

Besides explainability, XAI offers many other advantages including improved error analysis capabilities [6], verification of results and prospects of model refinement [21]. But despite all that, it does not seem to be typically designed for clinical purposes [31]. This section sheds light on the limitations and current challenges standing between the field of XAI (and the CAD tools powered by it) and routine healthcare.

3.1. Limited Training Datasets and Generalizability Issues

Limited labelled/ annotated datasets have been observed in most of the studies developing CAD tools for neurological disorders [14,29]. Most of the developed systems were found to be employing training and testing data from a single online source. This results in generalizability issues, i.e., such trained models are very highly likely to fail in case of data from unseen sources, not used in training. The open sharing of anonymized neuroimaging data should hence be encouraged and more public grand challenges should be introduced to trigger crowdsourcing [14] for the solutions of problems.

3.2. Current Focus Mostly on Optimizing Performance of CAD Tools

Currently XAI seems to be in its infancy and most of the energy and attention of researchers and data scientists is focused on accuracy enhancement and performance optimization of CAD tools [77]. This might be one of the reasons for the current immaturity of XAI in CAD.

3.3. Absence of Ground Truth Data for Explainability

At present, there is next to negligible annotated ground truth data for explainability, may that be visual, textual, or in any other form. For example, for AD, several neuroimaging and clinical biomarkers labelled datasets can be found, but none exists to validate the heatmaps for AD generated by XAI algorithms [29]. This makes performance evaluation challenging for XAI systems [17].

3.4. Focus on Single Modality

Most of the CAD research found in literature is single modality oriented. Very few multi-modal studies were found. The same was found to be true for XAI powered CAD tools. The correlations between interpretations of different modalities may contribute significantly in the differential diagnosis [29] and hence demands attention. The CAD of neurological disorders is a tedious task, particularly in case of Neurodegenerative disorders, like AD and PD, where no clearly evident findings are present on the brain MRI as opposed to tumors, which can be seen vividly as an abnormal growth. Another set of such diseases is MS and NMO, for which the differential diagnosis is extremely challenging even for medical experts, given the similarities in symptoms and lesion patterns on MRI scans [100]. MS and NMO are demyelinating diseases of the Central Nervous System (CNS) which produce lesions/ plaques on the brain, spinal cord, and optic nerve [101]. Since the treatment and management of both diseases is different [102], and the treatment of MS might have adverse effects on NMO patients [103], early and accurate diagnosis is of paramount significance [104,105]. 50-85% cases in these disorders show lesions on brain MRI [106], which appear as hyper-intense regions on T2w and FLAIR scans [107]. In case of no brain lesions, analysis of spinal lesions and orbits can help [108]. In such cases, a CAD tool trained for the differential diagnosis can perform significantly better if multi-modal patient data is used for training, before deployment in real-time clinical settings.

3.5. Only Visual Explanations Sufficient~

In this review, visual explanations were found to be the most dominant in recent research [26,30,31]. It is understandable since medical imaging is primarily associated with visual tasks. But are they enough~ Non-visual methods were observed to be hardly researched [26]. Some folks might not be contented with only visual explanations, and be more interested in explanations akin to a professor teaching his trainees. Textual XAI approaches might bring additional baggage of Natural Language Processing (NLP) with them, but that’s ok, as long as our end-users/ healthcare professionals are satisfied.

To explore the need for more doctors onboard, we requested input from Dr. Danesh Kella, Assistant Professor, Department of Medicine, Mayo Clinic Florida, USA. We believe that medical domain experts, being the ultimate end-users of such XAI powered CAD tools, would be the best sources to get this information from. The following comments of Dr. Kella have been presented here offering a sneak peek into the mind of medical domain experts and their expectations from XAI powered CAD tools:

“Yes, the explanation of the AI’s decision into AI models would likely increase confidence and comfort in using such systems. However, it would be even more helpful if the AI explains it in a manner akin to how a professor would explain to their trainees how a certain characteristic of the brain mass on MRI indicates the likelihood of a certain tumor. By providing human-like insights into how AI models arrive at their decisions, they offer transparency and clarity to users, including healthcare professionals.”

3.6. How to Judge XAI Performance~

Many researchers have proposed and developed XAI powered tools but very few have worked in the direction of their performance evaluation, either quantitative or qualitative [20,31]. The existing very scanty engineer-centered performance evaluation paradigm needs to shift, and more involvement of the medical experts (who are the ultimate users) needs to be ensured. There is an absolute need of uniform adoption of standard assessment criteria for explainability across the research community [6].

3.7. More Doctors Onboard, Please!

When it comes to XAI powered CAD research, from data annotation, to training, to providing ground truth explanations, to qualitatively analyzing the results of such tools, nothing is materializable without medical experts’ involvement and contribution. Being the end-users of such tools, their requirements need to be prioritized above all. The idea of human (doctor)-in-the-loop has been observed frequently in the literature [10], and has been termed as a prerequisite for design, development, and use of XAI based CAD applications [20].

3.8. User Awareness

It is important to thoroughly explain the capabilities, advantages, and limitations of XAI in CAD research to users (patients and doctors) [20].

3.9. Security, Safety, Legal, and Ethical Challenges

AI along with all the explainability and interpretability associated with it, specially in the field of medicine would face safety, legal, and ethical challenges [11]. Because of this field still being in its infancy [17], the regulations will also take time to mature.

Data security and privacy are pivotal in healthcare and diagnostic medicine as patients’ personal information and medical records are involved [109]. With the increasing adaptation of digital healthcare solutions, concerns about the security measures have also been increasing [110]. Various healthcare data security breaches have been reported in the literature. In 2020, an astounding 642 such cases were reported in the United States, with an unauthorized exposure of over 30 million healthcare records [111].

Digital health data commonly referred to as the Electronic Health Records (EHRs) play an important role in centralizing healthcare system. Using this data, medical experts can have access to the entire patient medical history, ailments, and treatment regimens followed in the past. This information is vital to fast, accurate, and safe diagnosis, prognosis, and treatment [112]. These records are, however, highly confidential since they contain the most private information about the patients [113]. Technologies such as cloud computing and remote access, make this data vulnerable to cyber-attacks [114]. In case of AI-based CAD regimens, the training data also comprises of such medical information, including imaging data, patient history, and blood work. Among the many forms of cyber attacks possible on AI powered CAD systems, ‘Data Poisoning’ is the most significant one [115]. Since an AI model makes decisions based on the data fed to it during training (at least in case of supervised learning), any change in that data can result in abnormal inferences. Optimistically thinking, this cannot cause physical harm to the patient, but in worst case scenario, serious damage can be caused in case of choosing a wrong treatment path due to inaccurate diagnosis. XAI can help curb such issues, where a doctor can identify the invalid explanations generated by the CAD tool, reject the diagnosis, and request a technical inspection and reevaluation of the underlying AI regimen. This would not be possible in old-school ‘black box’ AI, where an inference is supposed to be blindly trusted. On the other hand, in the case of smart medical devices, cyber attacks can result in catastrophic consequences, for example, imagine an insulin pump under such attack [115]. Given the increasing severity of such cyber threats with the advancements in technology [113], stern security measures including encryption protocols, patient data anonymization, access control protocols, and XAI are mandatory along with periodic security audits [116].

3.10. Let There Be Symbiosis!

It is important for doctors and data scientists to work together in this direction. The expertise of radiologists in identifying abnormalities from medical imaging data, and that of engineers in developing software are equally important to accomplish this task, and hence a symbiosis seems to be the only way forward. In addition, specialized trainings combining mathematics, data science, and medicine can be imparted to expedite the research in this direction [6]. Nevertheless, it seems pretty premature to comment on how long it would take to reduce the gap between medical and AI domains to zero.

4. Conclusions

Interpretability and explainability, although an absolute necessity for AI, specially in diagnostic medicine, still has a long way to go to achieve the mandatory levels of maturity for integration in regular medical practice. In addition to the lack of generalizability in AI powered CAD tools, scarcity of histopathologically proven and labelled datasets is also one of major shortcomings being faced by the currently available CAD tools. The key takeaways from this study can be summarized as follows:

• The integration of explainability in such CAD tools will surely increase the confidence of medical experts, but the current modes of explanations might not be enough. More thorough and human professor-like explanations are what the healthcare professionals are looking for.
• The quantitative and qualitative evaluation of such XAI schemes requires a lot of attention. The absence of ground truth data for explainability is also one of the major concerns at the moment which needs special attention, along with legal, ethical, safety, and security issues.
• Nothing of this can be materialized without getting both medical professionals and scientists onboard in an absolute symbiosis for working towards this cause.

It would be an ultimate waste of such tremendous advances in technology, computational resources and AI, that we have witnessed recently, if all this potential is not harnessed and channeled into healthcare to improve the quality of lives suffering from neurodegenerative disorders. It is extremely comforting to imagine a world with reduced mortality due to the fast, accurate, and reliable second opinions from XAI powered CAD tools embedded in healthcare system for doctors, to avoid delays and errors in differential diagnoses, resulting in saving precious lives.