Lymph node metastases detection in Whole Slide Images using prototypical patterns and transformer-guided multiple instance learning

Abstract Background: The examination of lymph nodes (LNs) regarding metastases is vital for the staging cancer patients, which necessary diagnosis and adequate treatment selection. Advancements in digital pathology, utilizing Whole-Slide Images (WSIs) convolutional neural networks (CNNs), pose new opportunities to automate this procedure, thus reducing pathologists’ workload while simultaneously increasing accuracy detection. Objective: To address task LN-metastases detection, use weakly supervised transformers are applied analysis WSIs. Methods & Materials: As WSIs too large be processed as a whole, they divided into non-overlapping patches, converted feature vectors using CNN network, pre-trained on HE-stained colon resections. A subset these patches serves input transformer predict if LN contains metastasis. Hence, selecting representative an important part pipeline. Hereby, prototype based clustering employed different sampling strategies tested. Finally, chosen fed transformer-based multiple instance learning (MIL) architecture, classifying LNs healthy/negative (that is, containing no metastases), or metastatic/positive metastases). proposed model trained only Camelyon16 training data (LNs from breast patients), evaluated test set. Results: achieves accuracies up 92.3% (from LNs). While struggles with smaller metastases, high specificities 96.9% can accomplished. Additionally, primary tumor (colon), where between 62.3% 95.9% could obtained. Conclusion: investigated transformer-model performs very good public data, but domain transfer needs more research.