A deep learning model to predict Ki-67 positivity in oral squamous cell carcinoma

Abstract

Anatomical pathology is undergoing its third revolution, transitioning from analogical to digital pathology and incorporating new artificial intelligence technologies into clinical practice. Aside from classification, detection, and segmentation models, predictive models are gaining traction since they can impact diagnostic processes and laboratory activity, lowering consumable usage and turnaround time. Our research aimed to create a deep-learning model to generate synthetic Ki-67 immunohistochemistry from Haematoxylin and Eosin (H&E) stained images. We used 175 oral squamous cell carcinoma (OSCC) from the University Federico II’s Pathology Unit’s archives to train our model to generate 4 Tissue Micro Arrays (TMAs). We sectioned one slide from each TMA, first stained with H&E and then re-stained with anti-Ki-67 immunohistochemistry (IHC). In digitised slides, cores were disarrayed, and the matching cores of the 2 stained were aligned to construct a dataset to train a Pix2Pix algorithm to convert H&E images to IHC. Pathologists could recognise the synthetic images in only half of the cases in a specially designed likelihood test. Hence, our model produced realistic synthetic images. We next used QuPath to quantify IHC positivity, achieving remarkable levels of agreement between genuine and synthetic IHC.

Furthermore, a categorical analysis employing 3 Ki-67 positivity cut-offs (5%, 10%, and 15%) revealed high positive-predictive values. Our model is a promising tool for collecting Ki-67 positivity information directly on H&E slides, reducing laboratory demand and improving patient management. It is also a valuable option for smaller laboratories to easily and quickly screen bioptic samples and prioritise them in a digital pathology workflow.