CS22 Model Building and Training of Skin Lesion Analysis
Model Building and Training of Skin Lesion Analysis
Hello, dear friend, you can consult us at any time if you have any questions, add WeChat: THEend8_
Project number: CS22
Maximum number of groups can accept: 2 groups
Project Title: Model Building and Training of Skin Lesion Analysis
Project Description and Scope:
Skin cancer is highly common in Australia and approximately one in five Australians are
exposed to sun in the workplace. Of this over 30% are at a risk of eye damage each year.
Currently, people are exposed to UV rays, even more so in some work environments. UV
radiation has been proved by WHO to be strongly associated with skin lesions and eye
damage. However, traditional diagnostic methods for skin lesion and eye damage require
visual observation by an experienced physician. However, the number of skin lesions in
everyone is large, so this is not an efficient method. On the other hand, the early symptoms
of skin lesion and eye damage are easily ignored by patients, which may lead to worsening
of the disease.
Metasense is committed to providing an imaging and risk profiling system called SafeSpot. It
contains an imaging system consisting of an analyser for classifying lesions as malignant or
benign according to their asymmetry, border irregularities, colour and diameter (the ABCD
rule of dermoscopy). The function of the classification is performed by AI, which intelligently
classifies and determines images of skin lesions and eye damage by means of deep learning
based techniques and gives the corresponding probabilities.
These are preliminary steps and can be used as an initial clinical diagnosis by a
dermatologist or ophthalmologist.
Expected outcomes/deliverables:
·Building on previous work, the deep learning model for skin lesion classification model
needs to have a higher F1_score and more categories, and a higher degree of robustness
and generalizability.
·Additional need for a model that can run on mobile devices.
·Building a method to enable personalised services and update the model with possible new
data in the future.
There might be a potential method such as an Incremental/Continual Lifelong Learning
(Transfer Learning) approach. This approach requires overcoming some of the challenges of
traditional transfer learning such as catastrophic forgetting (interference). (i.e. after the
model has been trained with new data, it still needs to have a high accuracy on the original
data).
Specific required knowledge, skills, and/or technology:
Deep learning/Machine learning; Algorithms; Image processing; Computer Vision; AI; Data
analytics; Experience with AWS/Colab cloud services; Excellent python skills.