Generative Modeling for Melanoma Detection
Typ
Examensarbete för masterexamen
Program
Complex adaptive systems (MPCAS), MSc
Publicerad
2022
Författare
Rosén, Anna
Zoubi, Mohamad Khir
Modellbyggare
Tidskriftstitel
ISSN
Volymtitel
Utgivare
Sammanfattning
Early detection significantly reduces deaths associated with melanoma, a skin cancer.
Despite this information, 80% of skin cancer related deaths are attributed
to malignant melanoma. Melanomas are difficult to diagnose by a dermatologist
(skin doctor) therefore many patients undergo unnecessary surgeries to get a biopsy
that can confirm the disease. Minimizing unnecessary surgeries would leave more resources
that in turn could lead to a higher frequency of earlier diagnosed melanomas.
Machine learning algorithms have shown a great potential in the field of medicine
and could be deployed to help doctors diagnose melanomas. To obtain a high performing
model it is crucial to have a large and balanced dataset. The scarcity of
labeled publicly available medical images makes applying machine learning an obstacle
in this field, thus hindering development. A solution to this problem could
be to synthesize realistic looking images using deep neural networks. One such network
is Generative Adversarial Network (GAN), which has been shown successful
in producing images in the field of medicine. This thesis explores the generation of
synthetic image data for medical purposes and how such data can be evaluated. We
utilize StyleGAN2-ADA to generate synthetic images of melanoma lesions that we
evaluate using both qualitative and quantitative measures. A survey was made to
establish if experts can identify generated images in a mixed dataset. The expert
dermatologists found the images difficult to distinguish from real ones, accordingly
proving that we can synthesize realistically looking images of melanomas. Using
a classifier trained on synthetic melanoma and non-melanoma images we are also
able to reach a high accuracy when validating against real data. Our results show
that synthetic images are verifiably realistic looking. From our research we are
able to conclude that synthetic data can be the answer to further development of
classification algorithms in a clinical setting.