Download This Paper
Preprints with The Lancet is a collaboration between The Lancet Group of journals and SSRN to facilitate the open sharing of preprints for early engagement, community comment, and collaboration. Preprints available here are not Lancet publications or necessarily under review with a Lancet journal. These preprints are early-stage research papers that have not been peer-reviewed. The usual SSRN checks and a Lancet-specific check for appropriateness and transparency have been applied. The findings should not be used for clinical or public health decision-making or presented without highlighting these facts. For more information, please see the FAQs.
Comprehensive AI Model Development for Gleason Grading: From Scanning, Cloud-Based Annotation to Pathologist-AI Interaction
33 Pages Posted: 25 Jul 2022
More...Abstract
Background: AI-based solutions for automated Gleason grading have been developed to assist pathologists to make rapid and quantitative assessments, but the generalization across various scanners and updating AI models continuously using new annotated data from end users remains a key bottleneck in the field.
Methods: We proposed an comprehensive digital pathology workflow for AI-assisted Gleason grading, incorporating an image quality check software A!magQC, a cloud-based annotation platform A!HistoNotes and Pathologist-AI Interaction (PAI) strategy. In this study, 187 prostatectomy and 156 biopsy specimens (24,371 mm2 of annotations) were collected from National University Hospital Singapore. Images were first acquired from Akoya Biosciences scanner. In addition, a subset of prostatectomy specimens was scanned by 4 other scanners. Image normalization and color augmentation were applied to increase the generalizability of the model. Lastly, clinical experiments were conducted with 5 pathologists from Singapore and China to investigate the accuracy, efficiency, and consistency of Gleason grading assisted by our model. Furthermore, we performed semi-automatic annotation and attempted to update the existing model using new data, enabling the PAI.
Finding: For images scanned by Akoya Biosciences scanner, on average, our AI model achieves a sensitivity of 85%, specificity of 96% and F1 score of 78% on Gleason grading for prostatectomy specimens, and 96% sensitivity on tumor detection for biopsy specimens. For images scanned by other 4 scanners, the average F1 score increased from 67% to 75% after adopting our generalization solution. In the clinical experiments, our model was integrated into A!HistoNotes and accelerated the Gleason scoring by 43%. Furthermore, it increased the annotation efficiency by 60% and was improved by incorporating semi-automatic annotations and incremental learning.
Interpretation: The proposed workflow helps to develop a generalised AI model for Gleason grading to assist pathologists in routine work. It can be applied to images scanned by a wide range of scanners. In addition to AI-assisted diagnosis, the model can be updated through the PAI efficiently.
Funding This project is funded by Exploit Technologies Gap-funded Project: “AI based H&E image analysis for prostate cancer staging” (EPTL/18-Gap029-R20H).
Declaration of Interest: All authors declare no competing interests.
Ethical Approval: Prostatectomy and biopsy formalin-fixed paraffin-embedded (FFPE) tissue specimens were collected from the Department of Pathology, National University of Singapore Hospital (NUHS) approved by the NHG Institutional Review Board (DSRB study reference number: 2018/01186).
Keywords: Digital Pathology, artificial intelligence, deep learning, Prostate Cancer, Pathologist-AI Interaction, Incremental Learning, Cloud Annotation, WSI Image Quality Control
Suggested Citation: Suggested Citation