Semi-automated Ki67 Index Label Estimation For HE Images Classification

The quantification of Ki67 proliferation index conducted through immunohistochemistry staining holds significant importance in histopathology. However, this process has several limitations, including variability and subjectivity in interpretation, along with challenges in quantification, rendering the process time-consuming and costly. Consequently, the utilization of neural network models presents a promising avenue for enhancing this domain. Yet, the creation of a sizable, high-quality annotated dataset remains a laborious task for experts. In this paper, we propose a validation and subsequent improvements of previously suggested semi-automated approach for generating Ki67 scores from pairs of hematoxylin and eosin (HE) and immunohistochemical slides, aiming to reduce the reliance on expert intervention. This approach integrates image analysis techniques such as clustering and optimization for tissue registration. Through the proposed approach and modification of worklfow, aiming to reduce the variability of the quantification error on different whole slide images, we annotated HE patches and conducted multiple experiments to fine-tune ResNet models for predicting Ki67 scores from HE images.