Computer Aided Dysplasia Grading for Barrett’s Oesophagus Virtual Slides

Dysplasia grading in Barrett’s Oesophagus has been an issue among pathologist worldwide. Despite of the increasing number of sufferers every year especially for westerners, dysplasia in Barrett’s Oesophagus can only be graded by a trained pathologist with visual examination.
Therefore, we present our work on extracting textural and spatial features from the tissue regions. Our first approach is to extract only the epithelial layer of the tissue, based on the grading rules by pathologists. This is carried out by extracting sub images of a certain window size along the tissue epithelial layer. The textural features of these sub images were used to grade regions into dysplasia or not-dysplasia and we have achieved 82.5% AP with 0.82 precision and 0.86 recall value. Therefore, we have managed to overcame the ‘boundary-effect’ issues that have usually been avoided by selecting or cropping tissue image without the boundary.
Secondly, the textural and spatial features of the whole tissue in the region were investigated. Experiments were carried out using Grey Level Co-occurrence Matrices at the pixel-level with a brute-force approach experiment, to cluster patches based on its texture similarities.Then, we have developed a texture-mapping technique that translates the spatial arrangement of tissue texture within a tissue region on the patch-level. As a result, three binary decision tree models were developed from the texture-mapping image, to grade each annotated regions into dysplasia Grade 1, Grade 3 and Grade 5 with 87.5%, 75.0% and 81.3% accuracy percentage with kappa score 0.75, 0.5 and 0.63 respectively.
A binary decision tree was then used on the spatial arrangement of the tissue texture types with respect to the epithelial layer to help grade the regions. 75.0%, 68.8% and 68.8% accuracy percentage with kappa value of 0.5, 0.37 and 0.37 were achieved respectively for dysplasia Grade 1, Grade 3 and Grade 5. Based on the result achieved, we can conclude that the spatial information of tissue texture types with regards to the epithelial layer, is not as strong as is on the whole region.
The binary decision tree grading models were applied on the broader tissue area; the whole virtual pathology slides itself. The consensus grading for each tissue is calculated with positivity table and scoring method. Finally, we present our own thresholded frequency method to grade virtual slides based on frequency of grading occurrence; and the result were compared to the pathologist’s grading. High agreement score with 0.80 KV was achieved and this is a massive improvement compared to a simple frequency scoring, which is only 0.47 KV.