Improving Skin Tumour delineation by Optimising Clustering and Training Data Selection

Sammanfattning

Due to the increasing incidence rates of skin cancer, there is an urgent need for novel technologies that can improve diagnostic accuracy and patient outlook. In a previous study, a machine learning model was developed to automatically delineate skin tumours from non-invasive hyperspectral images. However, the reliability of the model was undermined by difficulties with a preprocessing step in which the approximate location of the tumour was found via clustering. Failure to correctly identify the tumour resulted in incorrect annotation of data subsequently used for training the machine learning model. In these cases, the initial error obtained from the clustering propagated throughout the subsequent parts of the pipeline, ultimately causing incorrect final tumour size predictions. This thesis investigates possible improvements of the clustering procedure for approximating the location of the tumour. Specifically, it evaluates three different methods for selecting the correct tumour cluster, examines to what extent the validity of a clustering can be judged by computing silhouette coefficients, investigates and compares the performance of different clustering algorithms, assesses the impact of adding spatial or radial data prior to clustering. The clustering performance improved most significantly when the radial data was incorporated. Despite improvements in the final prediction of the tumour size, the most challenging sample continued to underestimate the tumour size even after correct initial clusters were obtained. This indicates that additional issues likely arise in a subsequent part of the pipeline, possibly in the method that was used to select and annotate the training data from the clustering result.