Automated HER2 Scoring of Breast Cancer Tissue using Upconverting Nanoparticle Images

Sammanfattning

Computer aided pathology is becoming more and more of a requirement within pathology due to increased demand of individualised treatments and personalised medicine. Because of the advance of digital pathology in recent years, where a high resolution camera acquire images of microscope slides, pathologists can now assess tissue samples in digital images. This has enabled automatic assessment of pathological images. A specific area of interest is the quantification of HER2-receptors in breast cancer tissue which decides if targeted therapy can be used or not. The standard staining within the area obstructs cell morphology and the images are difficult to analyse and classify. Existing automated HER2-classification methods in the field rely heavily on colour consistency or are neural networks which are difficult to interpret. Lumito AB has developed a reagent kit that, via laser and upconverting nanoparticles, demonstrates the HER2-expression in separate images that does not interfere with cell-morphology. These images are potentially more suitable for traditional image analysis and could potentially enable the possibility to develop simple, fast and interpretable algorithms that could quantify the HER2-expression and classify tissue samples. In this project, two algorithms were developed for classification of the upconverting nanoparticle based images. They were considered to be simple in the sense that the bases of classifications would be easy to explain to a pathologist due to the fact that they were inspired by the guidelines that pathologists use for HER2-classification. The algorithms performed on par with a pathologist and could be used as a screening tool, reducing the pathologist’s workload. The algorithms were also accurate in classification of the HER2 positive and equivocal tissue samples but fail to classify these unambiguously, and a pathologist would still have to assess these samples manually. It is difficult to say how well the algorithms performed in reality due to the relatively small data-set. This project should be seen as a proof of concept and future work would have to be done to further validate and improve the results even though the start is promising.