USCAP 110th Annual Meeting
March 13 – 18, 2021
Poster II – Monday PM – Informatics
Background:
Colorectal cancer is one of the top three cancers afflicting men and women worldwide, with an estimated 1.8 million people diagnosed in 2018. As colonoscopies become more accessible and affordable, colorectal biopsies account for a great part of the histopathology laboratory workload. Artificial Intelligence (AI) promises to streamline workflow, and improve cancer detection and classification; enabling busy pathologists to focus on higher order decision-making tasks. We trained and validated a unique AI deep learning model as an assistive tool to screen for colonic malignancies in colorectal specimens.
Design:
The study cohort consists of Whole Slide Images (WSI) obtained from 294 colorectal specimens. Utilising Qritive’s unique deep learning model based on a region-based convolutional network (Mask-RCNN) architecture with a ResNet-50 feature extraction backbone that provides glandular segmentation and a classical machine learning classifier, we first trained our glandular segmentation deep learning model using pathologist’s annotations on a training cohort of 66191 image tiles extracted from 39 WSIs (8 biopsies and 31 resections). We then applied a classical machine learning-based slide classifier that sorted the WSIs into low-risk (benign, inflammation) and high-risk (dysplasia, malignancy) categories. We then further trained the composite AI-model’s performance on a larger cohort of 105 resections WSIs and validated our findings on a cohort of 150 biopsies WSIs, against the classifications of two independently blinded pathologists. We evaluated the area under the receiver-operator characteristic curve (AUC) and other performance metrics.
Results:
The AI-model achieved an AUC of 0.917 in biopsies WSIs. Notably the AI-model achieved excellent sensitivity (97.4%) in detection of high-risk features of dysplasia and malignancy, but demonstrated lesser specificity (60.3%). Our ongoing review suggests that our composite AI-model is imperfect in situations of abundant mucin and low grade dysplastic glands, resulting in the higher than expected false positive rates.
Conclusion:
We demonstrate that our unique composite AI-model incorporating both a glandular segmentation deep learning model and a classical machine learning classifier has promising ability in picking up high risk colorectal features (Figure 1). The concomitant high sensitivity highlights its role as a potential screening tool in assisting busy pathologists by outlining the dysplastic and malignant glands. Ongoing calibration and training of our composite AI-model will improve its accuracy in risk classification of colorectal specimens.