Radboud University Medical Center Nijmegen
EDL P16-25 P5: Deep Learning for Human and Animal Health (HAAH)
Automated scoring of nuclear pleomorphism scoring as a spectrum in whole-slide breast images.
In routine clinical practice, pathologists score nuclear pleomorphism as 1 (no change), 2 (moderate) or 3 (marked) as per the breast cancer grading guidelines in an effort to quantify the qualitative analysis of the extent of abnormalities in the appearance of tumor nuclei. Moreover, this classification rule enforces a single numerical pleomorphism score for the entirety of a whole-slide image which may contain multiple tumor regions with varying tumor morphology. As a consequence, pathologist agreement is poor, and nuclear pleomorphism is the most subjective component of the breast cancer grading criteria.
In this project, we prove that this inherently continuous nature of nuclear pleomorphism can fully be realized through training deep learning models without time-consuming nuclei-level annotations by pathologists which is one of the key limiting factors in computational pathology. Our approach is capable of scoring a continuous spectrum of nuclear pleomorphism on whole-slide images, enabling an unprecedented way of analyzing the pleomorphism continuum in whole-slide images without observer subjectivity as well as guiding sampling for molecular testing by informing the evolution of tumor biology. Through multiple experiments, we also demonstrate that our deep learning model performs as well as the top-level performing pathologists when the continuous scores are translated into the traditional three-category classification.
We believe that our work is going to influence many works to come to systematically analyze and re-evaluate the guidelines which were established with the limitations of the technology and the understanding of the subjects at the time, using the much powerful Artificial Intelligence tools of today