SliDL: A toolbox for processing whole-slide images in deep learning.

The inspection of stained tissue slides by pathologists is essential for the early detection, diagnosis and monitoring of disease. Recently, deep learning methods for the analysis of whole-slide images (WSIs) have shown excellent performance on these tasks, and have the potential to substantially reduce the workload of pathologists. However, WSIs present a number of unique challenges for analysis, requiring special consideration of image annotations, slide and image artefacts, and evaluation of WSI-trained model performance. Here we introduce SliDL, a Python library for performing pre- and post-processing of WSIs. SliDL makes WSI data handling easy, allowing users to perform essential processing tasks in a few simple lines of code, bridging the gap between standard image analysis and WSI analysis. We introduce each of the main functionalities within SliDL: from annotation and tile extraction to tissue detection and model evaluation. We also provide 'code snippets' to guide the user in running SliDL. SliDL has been designed to interact with PyTorch, one of the most widely used deep learning libraries, allowing seamless integration into deep learning workflows. By providing a framework in which deep learning methods for WSI analysis can be developed and applied, SliDL aims to increase the accessibility of an important application of deep learning.

Quantification of TFF3 expression from a non-endoscopic device predicts clinically relevant Barrett's oesophagus by machine learning.

BACKGROUND: Intestinal metaplasia (IM) is pre-neoplastic with variable cancer risk. Cytosponge-TFF3 test can detect IM. We aimed to 1) assess whether quantitative TFF3 scores can distinguish clinically relevant Barrett's oesophagus (BO) (C≥1 or M≥3) from focal IM pathologies (C<1, M<3 or IM of gastro-oesophageal junction); 2) whether TFF3 counts can be automated to inform clinical practice. METHODS: Patients from the Barett's oEsophagus Screening Trial 2 (BEST2) case-control and BEST3 randomised trials were used. For aim 1, TFF3-positive glands were scored manually and correlated with clinical diagnosis. For aim 2, machine learning approach was used to obtain TFF3 count and logistic regression with cross-validation was trained on the BEST2 dataset (n = 529) and tested in the BEST3 dataset (n = 158). FINDINGS: Patients with clinically relevant BO had higher mean TFF3 gland count compared to focal IM pathologies (mean difference 4.14; 95% confidence interval, CI 2.76-5.52, p < 0.001). The mean class-balanced validation accuracy was 0.84 (95% CI 0.77-0.90), and precision of 0.95 (95% CI 0.87-1.00) for detecting clinically relevant BO. Applying this model on BEST3 showed precision of 0.91 (95% CI 0.85-0.97) for focal IM pathologies with a class-balanced accuracy of 0.77 (95% CI 0.69-0.84). Using this model, 55% of patients (87/158) in BEST3 would fall below the threshold for clinically relevant BO and could avoid gastroscopy, while only missing 5.1% of patients (8/158). INTERPRETATION: Automated Cytosponge-TFF3 gland quantification may enable thresholds to be set to trigger confirmatory gastroscopy to minimize overdiagnosis of focal IM pathologies with very low cancer-associated risk. FUNDING: Cancer Research UK (12088/16893 and C14478/A21047).

Triage-driven diagnosis of Barrett's esophagus for early detection of esophageal adenocarcinoma using deep learning.

Deep learning methods have been shown to achieve excellent performance on diagnostic tasks, but how to optimally combine them with expert knowledge and existing clinical decision pathways is still an open challenge. This question is particularly important for the early detection of cancer, where high-volume workflows may benefit from (semi-)automated analysis. Here we present a deep learning framework to analyze samples of the Cytosponge-TFF3 test, a minimally invasive alternative to endoscopy, for detecting Barrett's esophagus, which is the main precursor of esophageal adenocarcinoma. We trained and independently validated the framework on data from two clinical trials, analyzing a combined total of 4,662 pathology slides from 2,331 patients. Our approach exploits decision patterns of gastrointestinal pathologists to define eight triage classes of varying priority for manual expert review. By substituting manual review with automated review in low-priority classes, we can reduce pathologist workload by 57% while matching the diagnostic performance of experienced pathologists.