Depth data improves non-melanoma skin lesion segmentation and diagnosis

Abstract

Examining surface shape appearance by touching and observing a lesion from different points of view is a part of the clinical process for skin lesion diagnosis. Motivated by this, we hypothesise that surface shape embodies important information that serves to represent lesion identity and status. A new sensor, Dense Stereo Imaging System (DSIS) allows us to capture 1:1 aligned 3D surface data and 2D colour images simultaneously. This thesis investigates whether the extra surface shape appearance information, represented by features derived from the captured 3D data benefits skin lesion analysis, particularly on the tasks of segmentation and classification. In order to validate the contribution of 3D data to lesion identification, we compare the segmentations resulting from various combinations of images cues (e.g., colour, depth and texture) embedded in a region-based level set segmentation method. The experiments indicate that depth is complementary to colour. Adding the 3D information reduces the error rate from 7:8% to 6:6%. For the purpose of evaluating the segmentation results, we propose a novel ground truth estimation approach that incorporates a prior pattern analysis of a set of manual segmentations. The experiments on both synthetic and real data show that this method performs favourably compared to the state of the art approach STAPLE [1] on ground truth estimation. Finally, we explore the usefulness of 3D information to non-melanoma lesion diagnosis by tests on both human and computer based classifications of five lesion types. The results provide evidence for the benefit of the additional 3D information, i.e., adding the 3D-based features gives a significantly improved classification rate of 80:7% compared to only using colour features (75:3%). The three main contributions of the thesis are improved methods for lesion segmentation, non-melanoma lesion classification and lesion boundary ground-truth estimation.