.png)
The advent of digitalisation of pathological laboratories and recent progress in artificial intelligence-driven computer vision emphasized the development of computer‐assisted quantification, assessment and analysis of whole-slide images (WSI). In this work, we shed light on the nuclear cell proliferation protein Ki‐67, along with the ER (Estrogen Receptor) and PR (Progesterone Receptor), biomarkers being important in prognosis and predictive setting in breast carcinomas. The main objective of this study is to propose an end-to-end pipeline based on a deep neural network for :
① hotspot detection on Ki-67 stained WSI
② proliferation rate quantification for Ki-67
③ ER and PR quantification based on transfer learning from our pre-trained network
Train and test data consists of 3500 Ki-67’s images labelled by three expert pathologists. Annotations were available in three different classes: immunopositive tumour cells (TC), immunonegative TC, and tumor infiltrating lymphocytes (TIL). For training, images and locations of target cells are fed into a segmentation learning pipeline: a U-Net-based architecture. Each tile undergoes preprocessing steps and image augmentation.
Two external and unseen datasets (IHP and AIDPATH) of Ki-67, ER and PR stained WSIs are used to provide an unbiased assessment of the pipeline’s performance. Experienced pathologists individually examined these slides and the proliferation index (PI) was provided. IHP WSI samples were scanned using two scanners : Aperio GT450X (Leica) and Ventana iSan HT (Roche). AIDPATH used the Aperio Scanscope CS (Leica) scanner.
The model's performance was evaluated using the external datasets. The Ki-67 PI is computed over three hotspots with more than 500 marked cells. We report the concordance between our automatic quantification and pathologist's manual assessment over pre-defined cut-offs. For the IHP dataset, our model achieves an accuracy of 97.12% (Ki-67), 96.48% (ER) and 97.05% (PR) respectively. It yields an overall concordance of 98.33% (Ki-67), 98% (ER) and 98.75 (PR) on AIDPATH cohort.
A deep learning architecture trained on annotated Ki-67 WSIs was evaluated on unseen datasets. Furthermore, an assessment was conducted on a transfer learning strategy for quantifying the ER and PR proliferation score. Regarding the classification of patients based on the defined cut-off ranges, our approach achieves an outstanding agreement with pathologists clinical diagnosis. This model has the potential to improve pathologists' efficiency by offering standardized and highly accurate scores within a matter of seconds
C Bossard, Y Salhi, G Quereux, A Khammari, S Salhi, C Jérôme