2.1 Search Strategy & Inclusion Exclusion Criteria

We extracted only the papers which are relevant to this research study. Therefore, we crafted our searches carefully. Rather than just typing in primary keywords, we mixed and matched terms that fit exactly what we were looking for. For example, instead of just looking up "computer vision in healthcare" we dialed in our search to be more detailed. The following search string enabled the search engines of the databases to locate and display the articles that align with our intended ones. The general search string for all the databases was:

[("Computer Vision" AND "Healthcare") AND ("Drug Discovery" OR "Disease Detection" OR "Remote Patient Monitoring" OR "Cancer Detection" OR "Cell Identification" OR "Surgery Assistance" OR "Human Identity Decoding") NOT ("Surveillance" OR "Robotics" OR "Manufacturing" OR "Agriculture" OR "Transportation")].

Prior to constructing our search string, we identified five domains from previous literature ^{[1,2,5,6, 11-15]} as significant areas where CV techniques have shown potential for transformative impact in healthcare. Our goal was to focus our review on domains where CV techniques could contribute most significantly to advancements in healthcare outcomes. By predefining these domains in our search string, we aimed to ensure a targeted and relevant exploration of the literature. The exclusion of domains such as surveillance, robotics, manufacturing, agriculture, and transportation from our search string is determined by our preliminary literature study. The excluded domains relate to industrial, agricultural, or security applications of CV technologies which differ from healthcare domains. Our review aimed to concentrate on areas where CV directly influences healthcare such as patient care, medical diagnostics, and therapeutic interventions. Therefore, we excluded these domains. The construction of our search string and the selection of included and excluded domains were the result of careful consideration aimed at maximizing the relevance of our systematic review.

The inclusion criteria refer to the specific conditions that a study needs to meet to be part of the review. On the other hand, the exclusion criteria list the conditions that disqualify a study from the review ^[17]. We have followed specific rules regarding the studies included in our review. Table 1 presents inclusion and exclusion criteria for this study.

2.2 Data Extraction and Analysis

Data collection is a systematic process of gathering and measuring information on variables for research or analysis. It ensures accurate and relevant information for decision- making and evaluation ^[26]. For our review, we carefully examined articles in each stages and selected relevant studies. As summarized in Table 2, this careful curation from multiple sources has assisted us in identifying gaps in knowledge and important contributions in the field. We started with 125 papers from our search strategy. It covers various aspects of healthcare.

Each paper was examined focusing on titles, abstracts, and key findings. Based on the relevance, and methodological soundness we narrowed our list to 20 papers for a deeper analysis. Fig. 3 depicts the number of studies selected for this study where CV techniques are successfully applied in healthcare domains. After analyzing the data, we have found five major domains such as drug discovery, disease detection, remote patient monitoring, surgical assistance, and human identity identification. Most studies utilized CNN for their investigations and surgical assistance was found as the predominant area where CV techniques were applied.

After data analysis, we have found eight particular techniques have gained attention in the medical field because of their effectiveness in handling different medical imaging problems. These techniques include CNN, AdaBoost, Convolutional LSTM (ConvLSTM), EfficientNetB0, LSTM, Mask R-CNN, Lucas-Kanade, and VGG ^[18-25]. Fig. 4 shows number of articles reviewed in this study by publication year.

The horizontal bar diagram in Fig. 4 shows the number of articles published from 2020 to 2023 selected for this study. It shows that 2020 had the highest publication with 6 articles, followed by 5 articles each in 2021 and 2022. 4 articles have been selected from 2023 for this study.

Fig. 3. CV techniques applied in healthcare domains.

Fig. 4. Number of articles by publication year.

3.1 RQ1: What Types of CV Techniques have been Investigated for Various Healthcare Applications?

DL techniques specifically CNNs have emerged as effective solutions for diverse healthcare applications. They increase the accuracy of diagnosis and treatment planning across various healthcare applications. For instance, the EfficientNet-B0 Algorithm is a type of CNN that effectively detects skin cancer with an Area Under the Curve (AUC) of 0.909 ^[27] while CNNs achieve an accuracy of 83.2\% in similar tasks ^[28]. Among twenty studies reviewed, two specifically experimented with skin cancer detection using DL techniques though they measured outcomes in two different ways. One study highlighted the enhanced accuracy in breast cancer detection using the AdaBoost algorithm ^[29]. Apart from skin cancer detection, CNNs have been applied to different types of healthcare applications such as predicting Lyme disease risk with 92\% accuracy ^[31], developing assistive tools for diagnosing brain tumors with 90\% accuracy ^[32], assisting in surgery ^[41,46], and enabling remote patient monitoring ^[37]. Beyond CNNs, there are other techniques being utilized in various healthcare applications such as LSTM, AdaBoost, VGG, and the Stripenn algorithm. One study reported that the convLSTM model effectively classified suturing gestures with an AUC of 0.88 ^[44] while VGG models identified faces with 97.35\% accuracy ^[36]. The effective application of DL involves systematic steps. Comprehensive datasets contributed to the success of drug identification and face retrieval ^[35]. Transfer learning ^[41,44] optimized scenarios with limited data, while regularization countered overfitting ^[38]. Model interpretability was advanced ^[42] to develop clinical trust and collaboration with experts ^[42,46] to ensure clinical alignment. Although the impact of DL techniques on various healthcare applications is proven through systematic strategies mentioned across studies but challenges persist in CV systems’ seamless integration with existing healthcare systems.

3.2 What Techniques have been Developed to Enhance the Interpretability of CV Models in Healthcare Applications?

Enhancing the interpretability of CV models in healthcare applications is significant for clinician trust and model adoption. Various techniques have been developed to address this issue which focuses on both global and local interpretability. Self-attention mechanisms ^[37] enable CV models to highlight significant image regions for better understanding. Additionally, model-specific techniques interpret predictions generated by a specific model such as a neural network. Visualization of feature activations ^[44] demonstrates learned representations and links them to medical characteristics. Another category is model-agnostic techniques which provide easy-to-understand explanations of predictions. Integrating domain knowledge ^[37] augments interpretability by aligning predictions with medical insights. Techniques like Grad-CAM and SHAP ^[27,38] offer visual explanations that clarify model decision rationales. These strategies facilitate transparent and clinically relevant CV applications. Although integrating interpretable models and self-attention mechanisms has led to higher accuracy but there is a need for models that can generalize well across various medical conditions to ensure consistent interpretability which still remains a challenge. Moreover, the ability to operate in real-time and at scale is also a challenge.

3.3 RQ3: In what Ways is the Integration of CV and Sensors Optimized to Deliver Real-time Feedback Across Various Healthcare Applications?

The integration of CV and sensors in healthcare applications has been optimized to deliver real-time feedback across various domains such as patient monitoring, drug discovery, and surgical assistance. One study showed that integrating SAS software with fiber optic strain sensors enables immediate evaluation of suturing skills ^[42]. It enhances precision and provides actionable insights for skill refinement. This integration offers real-time assessment of surgical techniques and contributes to quantifiable proficiency metrics. Also, it contributes to minimally invasive surgery scenarios by leveraging ML algorithms to analyze sensor data and visual inputs. The integrated approach could enable prompt recognition of effective surgical maneuvers which promote better decision-making. The evolution of algorithms and synchronization mechanisms has the potential to revolutionize surgical training which increases patient safety ^[44]. For remote patient monitoring, incorporating both CV and physiological sensors has led to continuous health monitoring and real-time feedback. For example, a wearable sensor-integrated Internet of Things (IoT)-based system has shown notable success in effectively monitoring patients remotely ^[40]. This real-time feedback has improved the effectiveness of interventions through immediate adjustments based on sensor data. In drug discovery, the Deep Learning Drug Identification (DLDI) model identified prescribed drugs with more than 90\% accuracy from blister packaging ^[35] which shows the impact of CV and sensor integration on real-time feedback. Despite these advancements, the collection and transmission of patient health data have faced privacy and security challenges.

3.4 RQ4: What are the Most Effective Strategies to Establish Standardized Performance Metrics and Evaluation Protocols for CV Models in Healthcare Applications?

Establishing standardized performance metrics and evaluation protocols for CV models in healthcare applications is crucial for achieving consistency and reliability. A structured framework is essential but various studies have utilized different types of metrics for evaluation due to the examination of different datasets. Accuracy and AUC are the two most commonly used performance metrics across the studies. For instance, out of twenty studies reviewed, eight selected accuracy as the performance metric while three went for AUC. Strategies derived from different research studies focus on effective approaches. One study ^[44] shows the significance of disease-specific metrics by achieving notable classification and identification performance using AUC. Another study ^[36] reports the importance of domain-specific evaluation by attaining a 97.35\% accuracy rate in face detection. The integration of explainable AI techniques as demonstrated in these studies ^[37,44] not only enhances transparency but also helps in the evaluation process. Collaborative efforts with clinicians and domain experts are important to ensure clinically meaningful evaluation criteria. Additionally, benchmark datasets as mentioned in one study ^[44], provide a foundation for standardized assessment. The adoption of standardized performance metrics and evaluation protocols will enhance the accuracy of CV models. However, to ensure the robustness of CV models more diverse datasets can be produced and tested by CV models.

3.5 RQ5: What Ethical Guidelines and Data Governance Principles are Essential to Ensure Responsible Deployment of CV Techniques in Healthcare?

The responsible deployment of CV techniques in healthcare requires adherence to ethical guidelines and data governance principles. Ensuring the privacy of medical data is a challenge. During the experiment informed consent is obtained from patients but maintaining privacy becomes complex when various stakeholders are involved during experiments. Ethical guidelines and data governance have been emphasized across different studies. One study highlights privacy-preserving techniques, particularly in sensitive applications like human face recognition ^[36]. Other research has explored blockchain-based patient data control where data anonymization and de-identification have been performed to mitigate risks associated with patient data exposure ^[43]. The application of fairness principles can help to mitigate biases while interdisciplinary collaboration and regulatory involvement are crucial for developing a comprehensive ethical framework ^[45]. Transparency and accountability are important for justifying model decisions and addressing unexpected outcomes. As the AI techniques are evolving rapidly, it is crucial that existing ethical guidelines and governance frameworks be continuously updated to tackle new challenges. Overall, the integration of CV in healthcare requires a balanced approach that prioritizes patient privacy.

3.6 What Approaches have been Explored to Leverage CV for Predictive Analytics and Personalized Treatment Planning?

Leveraging CV for predictive analytics and personalized treatment planning in healthcare includes various methods as highlighted by different studies. In predictive analytics, out of twenty studies reviewed, eight studies have utilized CNN models where the highest testing accuracy was 92\% in predicting Lyme disease risk ^[31]. Other approaches, such as convLSTM and the AdaBoost algorithm have also been explored for predictive analytics. The convLSTM model effectively classified suturing gestures with an AUC of 0.88 ^[44] while AdaBoost showed enhanced accuracy in breast cancer prediction. The integration of CV with predictive analytics enables more precise treatment planning, especially in complex cases like cardiac care which is evidenced by a study that reported monitoring cardiac patients using a smartphone camera ^[38]. Various CV-based works have been reported for personalized treatment including the development of a CV-based system for Parkinson’s patients ^[39], and CNN and SAS aided systems to improve personalized surgery treatment ^[41,42]. Motion analysis has increased the accuracy of these aided systems. The integration of visual and sensor-generated data offers opportunities for dynamic personalized interventions ^[43]. DL models enhance breast cancer detection when combined with other patient data such as radiomics and genomics which offers a holistic approach to personalized treatment planning ^[29]. Collaborative interdisciplinary efforts ensure comprehensive personalized approaches although challenges remain in seamlessly integrating these advanced CV techniques into existing clinical workflows. Overall, these findings highlight the significant potential of leveraging CV for predictive analytics and personalized treatment planning in healthcare