Unraveling the Effect of Demographic Factors on the Performance of Melanoma Classification

Melanoma remains the most lethal form of skin cancer, responsible
for the majority of skin cancer-related fatalities despite its relatively
low incidence cite{milton2018}. Melanoma can spread to different parts
of the body if it is not identified and treated in its early stages. In
cite{kumar2021, kumarVatsa2022} previous work on melanoma
classification, the VGG16 architecture of the CNN performed better on the
cancer dataset than any other popular network.
%
The early detection of melanoma by analyzing skin lesion images aims to
enhance early diagnosis and accessibility. However, a significant challenge
arises when incorporating demographic factors as biases in training data,
as they can lead to disparities in model performance across different
populations. These biases often stem from the underrepresentation of
certain demographic groups in medical datasets, leading to lower accuracy
for underserved communities. Such disparities can have serious
consequences, including delayed diagnoses and inadequate treatment
recommendations, further increasing healthcare inequities
cite{dietterich2000}. Therefore, this study examines the effects of
demographic factors such as age, gender, and data scalability on the early
detection of melanoma. It also evaluates and compares two distinct deep
learning approaches for classifying melanoma from dermoscopic images and
associated patient metadata. The first experiment establishes a baseline
using a VGG-16 convolutional neural network (CNN) trained via transfer
learning. The second, expanded experiment introduces a novel multimodal
ensemble model that synergistically combines an EfficientNetB0 CNN with a
Multi-Layer Perceptron (MLP) to process both image and tabular data
concurrently.