Page 65 - AIH-1-2

Artificial Intelligence in Health AI in AD – Diagnosis and monitoring

respondents thought that automated suggestions for resources for AI models. 97-99 These datasets often consist of
diagnosing skin tumors had strong or very strong lesions that have been confirmed by pathology, follow-up
potential in terms of diagnostic categorization, whereas examinations, expert consensus, or in vivo confocal
42.6% thought that automated detection of mitosis had the microscopy, which enhances their reliability.
highest potential. 93,94
Prudent consideration is crucial for researchers involved
Patients typically know little about AI than medical in the development of AI programs, especially when facing
professionals. In a qualitative study conducted from May challenges related to training datasets. Estimating the
to July 2019, involving 48 patients and semi-structured optimal number of training images can be challenging, as
interviews for analysis, around 60% of participants having an insufficient dataset may compromise the quality
stated that shorter diagnosis times and easier access to of the program, while an excessively large dataset runs the
healthcare were the two biggest advantages of AI for skin risk of overfitting the ML classifier to the data, limiting its
cancer surveillance. Nonetheless, 40% of participants applicability to external datasets. It is important to note that
expressed concerns about potential dangers, including a advanced mathematical techniques are available to address
rise in patient anxiety. The patients identified the major these challenges, such as dropout, data augmentation,
benefits and drawbacks of AI as the ability to deliver more batch normalization, and others. 92,100,101 These methods
precise diagnoses (33 [69%]) and less precise diagnoses play a key role in preventing overfitting and ensuring
(41 [85%]). Notably, 35 out of 75 patients stated that they the robustness and generalizability of the AI program,
95
would recommend AI to friends and family. In summary, which holds significant clinical relevance from a scientific
pathologists and dermatologists generally hold an perspective. Efficient utilization of the dataset is crucial in
optimistic view of the prospective advantages and effects achieving the desired accuracy for specific classifications.
of AI in the field of dermatology. However, only a minority In addition, the dataset should include a diverse range of
of respondents within the cohort exhibited a good or images from various demographics to ensure that resulting
exceptional comprehension of AI. While most pathologists algorithms have external validity. 102,103 When acquiring
expect AI to be most useful in specific tasks rather than images, it is important to consider potential systematic
offering overall automated diagnostic advice, a majority errors such as variations in lighting, tools, or processes,
of dermatologists believe that AI will improve diagnostic particularly in different clinical settings, to maintain the
capabilities. Overall, only a small percentage (1 – 3%) of research’s validity beyond its original context. Simplifying
pathologists and dermatologists express concern that AI the program’s classifications to those with significant
may soon replace them. As long as AI is used in a way prognostic implications can help reduce the size of the
that maintains the doctor-patient relationship, patients are dataset and the complexity of algorithms. 102,104
amenable to using it to monitor skin conditions.
Randomized clinical trials must be carried out to
12. Perspectives and conclusion evaluate the potential of new computer methods and DL
The potential of AI in the field of AD presents an opportunity in large-scale investigations. Given the limited research
to significantly enhance diagnostic accuracy and provide in this area, these studies are especially important for
personalized healthcare. However, several aspects must gathering data on therapeutic benefits and assisting with
105
be addressed before this innovative approach can be causal inference. Moreover, addressing unmet demands
seamlessly integrated into routine clinical practice. AI is such as cost-effectiveness and safety concerns is critical
gaining recognition at a pace in the field of dermatology, before transferring AI technology from research to clinical
with researchers increasingly focusing on developing AI settings. Robust regulatory procedures are required to
programs that require diverse data sources for training guarantee the safe handling and preservation of private
purposes. These data sources include clinical patient data, information. Another important challenge is ensuring
which encompasses various aspects such as demographics, AI-based healthcare is equitable and inclusive. Healthcare
comorbidities, characteristics of skin lesions, and relevant AI should be trained and validated using population-
laboratory and imaging findings. Furthermore, molecular representative data to achieve generalizable performance
106
profiles obtained from biopsy data, such as proteomic levels. It is crucial to take into account social and health
analysis, provide valuable information. Another avenue inequalities that can exclude kids from particular groups
5,96
involves utilizing data from existing literature. Finally, images who typically have limited access to care. Relying mostly
play a crucial role in the analysis and classification process. on data from majority ethnic groups or patients with
Notably, publicly available benchmarking image datasets high socioeconomic status could introduce bias into
such as the International Skin Imaging Collaboration and AI performance, as the system may pick up diagnostic
PH2 dermoscopic archives serve as instrumental training tendencies from these over-represented groups. 107

Volume 1 Issue 2 (2024) 59 doi: 10.36922/aih.2775