Results from vitro/vivo

Throughout the development of the code, different configurations of the neural network architecture for training based on the melanoma databases have been tested to optimise the results. This architecture is made up of multiple convolutional layers and dense neurons that extract the main features of each image. Below is a visual representation of the modifications that the convolutional layers make on the moles images: 

It is important in diagnostics to ensure that the results are accurate at least as well as or better than human results. Generally speaking, the accuracy of a system translates into a higher ratio of true positives and negatives versus false positives and negatives. However, in the application to medical diagnostics, the statistical analysis of true positives versus false negatives is very important. Therefore, different metrics will have to be addressed to be able to analyse the results globally, some of the most commonly used are:

  • Accuracy: for overall performance measurement. 
  • Sensitivity: measures the rate of true positives (TP) against the total number of existing positives.
  • Specificity: it measures the rate of true negatives (TN) against the total number of existing negatives

ROC curve has also been used, whose area reflects the predisposition of the model to classify more elements as positive than negative (AUC>50%) and vice versa. Here it is emphasised that in a cancer diagnosis a false positive is better than a false negative. This metric is commonly used in clinical investigation.

The final results of the model are shown below:

True labels vs predicted labels obtained with our model are shown below, based on the so called confusion matrix: