Automatic Detection of Tumour Infiltrating Lymphocytes in Breast Cancer Whole Slide Images

Abstract: Cancer is one of the most common diseases this century, with breast cancer being the most common form. Pathological examination is used to detect and quantify Tumour-infiltrating lymphocytes (TILs) in breast cancer Whole Slide Images (WSIs), which can be done manually or automatically. Analysing these Whole Slide Images manually takes a lot of time and requires a lot of concentration from the pathologists. Automating this process takes significantly less time, is more efficient and less errorprone than the manual process. This study aims to provide an object detection model to localise and detect plasma and lymphocyte cells in breast cancer tissue. This research is conducted using the Design Science research strategy. First, data exploration of the TIGER data set took place. The data was pre-processed and split into a training, validation, and test set. Afterwards, the pre-trained YOLOv5 was used to train the object detection model. In the next step, the model is fine-tuned on the validation dataset. Finally, the model was tested and evaluated. The findings from this study show that training the YOLOv5m model without performing augmentation has the best performance. The results of this research are promising: the precision is 0.668, and the recall is 0.621. In conclusion, training an image analysis Deep Learning model to detect and localise tumour-infiltrating lymphocytes on Haematoxylin and Eosin (H&E) breast cancer tissue WSIs is feasible.