Page 64 - AIH-1-2

Artificial Intelligence in Health AI in AD – Diagnosis and monitoring

One major challenge is addressing bias and limitations through collaborations or utilizing telemedicine platforms
in dataset representation, which are evident in the diversity to expand the spectrum of the dataset.
in skin types, conditions, and demographic factors. These To reconcile the disparities between source and target
aspects may not have been adequately emphasized in datasets in AI scenarios for dermatology, domain adaptation
training datasets, leading to biases in model outcomes. techniques have been applied. These techniques aim to
85
90
The complexity adds an extra layer to the dynamics related align the distributions of data and increase adaptability
to clinical practices, impacting data relevance over time and without leading to overfitting incapacitation. By focusing
asking for continuous adaptation of AI models. In addition, on key features and applying an expert-driven targeted
providing access to various datasets is crucial for adequate approach within the limitations posed by limited datasets,
model training, but it poses challenges, and potential effective solutions can be developed to address limited
limitations may restrict effective model generalization. data. Overfitting can be mitigated by incorporating
91
Moreover, ethical considerations and patient privacy issues regularization techniques such as dropout or weight
further complicate matters, especially concerning sensitive decay during training. Ensuring the high quality of the
92
dermatological information. Careful balancing of the use limited dataset is crucial as it significantly influences
of patient information for AI research with individual the proper exploitation and effective performance of AI
privacy protection is essential. These factors make it models. By refining these strategies through the iterative
challenging to predict the performance of AI models in process that is in line with evolving research and clinical
actual clinical environments, which can significantly differ needs, practitioners could effectively overcome the
from controlled research environments due to diverse hurdles imposed by limited datasets in AI applications to
patient populations, variations in clinical workflows, and dermatology.
the dynamic nature of healthcare. Effectively recognizing
and addressing these nuanced limitations in AI applications 11. The acceptance of AI in dermatology:
in dermatology are pivotal for developing models that not Attitude attribute
only demonstrate technical proficiency but also seamlessly
align with the complex yet ever-changing realities of The application of AI to medical image recognition has
clinical practice. 86 garnered substantial attention recently, particularly in the
fields of dermatopathology and dermatology. The growing
Overcoming the challenges of small datasets in AI advancements in AI technology make its use as a decision
applications in dermatology requires an appropriate and support tool for dermatologists – particularly in diagnosis
well-designed strategy. Data augmentation is the key support – increasingly relevant within the current legal
approach that enhances dataset quality by transforming and health-care frameworks. With the growing utilization
images into different forms, thus improving data diversity of AI by both patients and medical professionals,
and allowing the model to learn more effective features. numerous regional and international survey studies
87
Transfer learning proves advantageous when models have been conducted to gauge perceptions and attitudes.
initially trained on huge datasets are fine-tuned using Between January and June 2019, a comprehensive online
particular dermatology datasets, allowing them to gain survey was distributed to 1271 participants across 92
general knowledge from a wider context. In addition, countries. The results revealed that respondents identified
88
the incorporation of carefully developed artificial data, dermoscopic images as the most promising application
resembling the characteristics of dermatological conditions, of AI in dermatology. Significantly, 77.3% of participants
serves as an effective pathway toward diversification, expressed approval or strong approval of AI’s potential in
underscoring the importance of adequate representation. improving dermatology, with 79.8% incorporating AI into
Active learning adds a new iterative retraining paradigm their medical education. However, only a minimal 5.5%
where the model selectively prioritizes informative or (70 out of 1,271) agreed or strongly agreed with the notion
challenging samples during every retraining cycle, thereby that AI would replace dermatologists in the near future.
refining its performance. Ensemble models, which A comparable international survey was conducted among
89
employ different architectures and hyperparameters, help dermatopathologists by the same research team, involving
minimize the effect of limited data by mixing predictions. 718 respondents from 91 countries. The findings revealed
Collaboration with other institutions, clinics, or research that 84.1% of respondents thought AI should be included
groups, along with data pooling, ensures the development in medical education, and 72.3% of respondents agreed or
of a broader and more diverse dataset while adhering strongly agreed that AI will improve dermatopathology.
to established rules on privacy and ethics. Active data Only 6.0% of respondents thought AI would eventually
collection remains crucial, requiring regular acquisition replace human pathologists. Interestingly, 79.2% of

Volume 1 Issue 2 (2024) 58 doi: 10.36922/aih.2775