Digital diagnostics: using machine learning models for skin cancer detection

Machine learning is extending the reach of medical science, allowing it to see further and relieve some of the burden from time-pressured healthcare professionals. Researchers are now exploring its use to transform disease detection. 

With a disease such as skin cancer, where early diagnosis is critical, the efficiencies gained from applied machine learning techniques could be crucial. At Modern College of Business and Science (MCBS) in Oman, Vijaya Padmanabha, associate professor and deputy head in the Department of Mathematics and Computer Science, has developed a model for classifying melanoma types that demonstrates the effectiveness of machine learning in clinical contexts. Padmanabha and her colleagues applied transfer learning to the GoogLeNet convolutional neural network (CNN) to build this model. 

Transfer learning reduces computational costs when constructing machine learning models. It is the digital equivalent of transferable skills – you can take a functionality from the CNN and retrain it to perform similar tasks. Drawing on a data set of over 25,000 images from the International Skin Imaging Collaboration project, Padmanabha and her colleagues designed an image classification model using the white smell agent as an optimisation metaheuristic – an algorithmic framework that tunes the model and guides it towards discoveries.

The team encountered technical hurdles along the way, such as overfitting, which can happen when a model yields accurate results with its training data but is less effective on fresh or unseen data sets. In real-world applications such as melanoma classification, this might make the model overly sensitive to irrelevant data, leading to incorrect conclusions. To make their model less susceptible to overfitting, the team applied max pooling – applying a filter to the images, downsampling them and summarising the most important features. The completed model classified melanoma types with 91 per cent accuracy.

“It is of high accuracy because we used transfer learning,” explains Padmanabha. “Usually, people don’t use transfer learning; they use a CNN model. But we used transfer learning with the fractional white smell, which gives better convergence accuracy.” Padmanabha’s model could be used to screen for breast cancer, brain tumour segmentation and retinal disease detection. “You can apply it in many ways, such as in blood pressure and oxygen saturation monitoring and robotic surgery and medical robotics,” she says. “You can use it anywhere because it is particularly useful for segmentation purposes.”

Similar to the melanoma classification project, the chances are that Padmanabha’s future projects will also have an interdisciplinary component. She says this is one of the benefits of working in computer science at a time when technology is advancing rapidly. People are keen to collaborate with experts like her, as machine learning can be a force multiplier for research impact.

Find out more about Modern College of Business and Science.