Page 160 - AI for Good Innovate for Impact

AI for Good Innovate for Impact

(continued)

Item Detail
Key aspects of the The proposed solution to these major problems is EquiDermAI, an
solution intelligent framework for overcoming racial and digital barriers in AI
diagnosis of any skin-related lesion/disease.
There are 2 main components that address the two major problems
previously explained:
1. Deep-Generative Style Transfer for Skin Tone Data Augmentation:
This component addresses the racial divide in intelligent dermato-
logical diagnosis. EquiDermAI uses an image-to-image translation
method based on CycleGAN (Zhu et al., 2017) to perform seman-
tic-preserving transformation of skin lesion images. Specifically, it
adapts lesion images from overrepresented lighter skin tones to
reflect underrepresented darker skin tones while preserving lesion
morphology and clinical features. Unlike classical neural style transfer,
which may distort structural pathology, CycleGAN operates without
requiring paired training data and focuses on domain adaptation,
enabling realistic generation of lesions on diverse skin tones without
altering diagnostic-relevant morphology. By augmenting datasets
with these clinically faithful synthetic images, EquiDermAI improves
diagnostic model generalisation across skin types and helps reduce
misdiagnosis rates in underrepresented populations.
2. Quantized TinyML Development for Resource-Constrained Areas:
This component addresses the digital divide. With many regions in
the world lacking access to powerful technological tools, it is import-
ant to keep diagnostic resources very efficient. EquiDermAI further
optimizes AI models for low-power devices, enabling accurate diag-
nosis in resource-constrained areas with limited computing power
or internet connectivity (Warden & Situnayake, 2019).
These two components are intricately designed to contribute to the
overall goal of creating an inclusive and accessible skin disease diagnosis
system. By combining advanced deep learning techniques with efficient
deployment strategies, EquiDermAI aims to overcome both racial and
regional barriers in dermatological care, democratizing accurate skin
disease diagnosis.
CONSIDERATIONS: There are legitimate concerns regarding whether
generated images retain diagnostic fidelity, particularly when transform-
ing skin tone while preserving the clinical morphology of lesions. To
address this, EquiDermAI will involve expert validation from board-cer-
tified dermatologists and skin-of-colour specialists in future iterations.
These experts will assess the realism and clinical accuracy of the synthetic
images, and validation may incorporate structured annotation protocols
with inter-rater reliability metrics (e.g., Cohen’s kappa) to quantify agree-
ment. Alternatively, or in parallel, automated similarity assessments may
be employed by comparing the learned feature embeddings of real
and generated lesions using pre-trained dermatological classification
networks. Techniques such as perceptual loss or cosine similarity in latent
space can serve as quantitative proxies for evaluating whether lesion
morphology remains preserved during style transformation. These
safeguards aim to ensure that synthetic data used for training maintains
clinical utility and does not introduce misleading artifacts.

Technology keywords Deep Learning, Generative Adversarial Network (GAN), Data Augmen-
tation, tinyML, TensorFlow Lite, Computer Vision, Swift/CoreML

Data availability Private

124