Deep representation clustering for histology annotation using few-shot learning and semi-supervised learning

Prostate cancer is one of the most common malignancies affecting male health all around the world, and the precise measurement of prostate-specific antigen (PSA) is important for prostate cancer diagnosis and prognosis. PSA levels can be measured directly from the urine. It is also possible to visualize the distribution of PSA expression in tissue samples. The staining pattern is believed to be relevant for the prognosis of the development of cancer. However, visual analysis of the staining pattern of PSA in the pathology image is difficult and tedious. In this research, we propose a few-shot learning and semi-supervised model for prostate cancer pathology images to improve the accuracy of PSA measurement with a few labeled data. Each sample has four equally large regions labeled with PSA scores, and is split into patches to fine-tune a pretrained deep learning model. Then we generate an embedding for those patches. We thereafter reduce the dimensionality of the embedding using UMAP. The UMAP-generated 3-dimension feature map shows the distribution of patch characteristics, where each point's color corresponds to a patch's feature distribution. We project the points' colors from the feature map back to the original pathology images, and the PSA distribution across the prostate sample is shown. Next, we perform clustering in the UMAP space, which is calculated on the 3D feature map to judge between positive and negative regions to calculate the percentage of positive area in each quarter. This research provides novel methods for calculating PSA scores on pathology images, improving the accuracy of PSA-based diagnosis and aiding in the identification of clinically significant prostate cancer regions.