Automatic cell classification and quantification with machine learning in immunohistochemistry images.

Abstract

The incidence of lymphoma, a cancer that affects both humans and animals, has witnessed a significant increase. In response, immunohistochemistry (IHC) has become an essential tool for its classification. This prompted us to develop an innovative mathematical methodology for the precise quantification of immunopositive and immunonegative cells, along with their spatial analysis, in CD3-stained lymphoma IHC images. Our approach involves integrating an algorithm based on a mathematical color model for cell differentiation, employing the distinctive morphological erosion, algorithmic transformations, and customized histogram equalization to enhance features. Refined local thresholding enhances classification precision. Additionally, a customized circular Hough transform quantifies cell counts and assesses their spatial data. The algorithms accurately enumerate cell types, reducing human intervention and providing total numbers and spatial information on detected cells within tissue specimens. Evaluation of IHC image samples revealed an overall accuracy of 93.98% for automatic cell counts. The automatic counts and location information were cross-validated by three pathology specialists, highlighting the effectiveness and reliability of our automated approach. Our innovative framework enhances lymphoma cell counting accuracy in IHC images by combining physics-based color understanding with machine learning, thereby improving diagnosis and reducing the risks of human error.