Abstract
Artificial intelligence in medicine is frequently evaluated according to improvements in
predictive accuracy, yet its broader significance lies in how technological systems are
constructed and situated within healthcare institutions. My technical capstone project, Machine
Learning: Using ML to Determine Malignancy of Skin Cancer Lesions, involved developing a
convolutional neural network capable of classifying dermoscopic images as benign or malignant.
I undertook this project to examine how machine learning might support earlier detection of skin
cancer and alleviate diagnostic strain in high-volume clinical environments. In parallel, my STS
research paper, Exploring Healthcare Equity in the Adoption of AI and Associated Technologies
for Cancer Treatments, investigated how the integration of AI into cancer care shapes healthcare
equity across institutional and demographic contexts. I pursued this inquiry to assess whether
technological advances in diagnostic capability translate into equitable access and outcomes.
Considered together, these projects reflect a unified concern: the same decisions about data,
infrastructure, and deployment that influence model performance also structure the distribution
of medical benefit.
The capstone project addressed the growing burden of skin cancer by designing and
training a CNN-based diagnostic model using over 25,000 dermoscopic images from the
International Skin Imaging Collaboration (ISIC) archive. Early detection remains critical for
improving survival outcomes, yet distinguishing between benign and malignant lesions can be
clinically complex and time-intensive. To address this challenge, I implemented image
preprocessing and augmentation techniques, applied class-weight adjustments to mitigate
imbalance, and incorporated regularization strategies to reduce overfitting. Model performance
was evaluated using accuracy, precision, and recall metrics to assess its capacity to generalize
beyond the training dataset.
The model achieved approximately 70% overall accuracy, with stronger performance in
detecting basal cell and squamous cell carcinomas and comparatively lower recall for melanoma
cases. These results demonstrate the feasibility of machine learning as a clinical support tool
while also revealing persistent technical constraints related to dataset composition and
representational diversity. The project ultimately underscores that algorithmic effectiveness
depends not only on architectural sophistication, but on the quality, balance, and contextual
relevance of the data that inform it. In this respect, the technical work illuminated the conditions
under which AI systems can assist clinical decision-making, as well as the limitations that shape
their real-world applicability.
My STS research asked: How does the adoption of AI in cancer diagnosis and treatment
shape healthcare equity, particularly for underserved populations? While clinical literature
frequently emphasizes improvements in efficiency and diagnostic precision, I examined how
institutional capacity, data infrastructures, and governance frameworks influence who gains
access to these innovations. Grounded in the Social Construction of Technology (SCOT)
framework, my methodology combined qualitative case analysis, document analysis of policy
and regulatory materials, and ethical assessment. This approach allowed me to analyze AI
systems as sociotechnical artifacts shaped through negotiation among developers, healthcare
institutions, clinicians, and policymakers rather than as neutral instruments of progress.
The findings demonstrate that equity outcomes emerge through three interrelated
mechanisms: stratified institutional implementation, performance disparities rooted in
historically uneven datasets, and governance processes that stabilize particular interpretations of
technological value. Resource-rich institutions are more likely to shape standards of AI-enabled
care, while under-resourced settings face structural barriers to adoption. At the level of model
construction, aggregate performance metrics can obscure demographic variation, embedding
historical inequities within algorithmic systems. Regulatory and institutional frameworks often
prioritize efficiency and validation thresholds over distributive considerations, allowing these
disparities to become durable. Together, the technical and STS components of my capstone
illustrate that advancing medical AI requires not only improved algorithms, but sustained
attention to the structural conditions that determine how innovation reshapes access, authority,
and care.
