Whole slide image representation in bone marrow cytology

Abstract

One of the goals of AI-based computational pathology is to generate compact WSI representations, identifying the essential information required for diagnosis. While such approaches have been applied to histopathology, few applications have been reported in cytology. Bone marrow aspirate cytology is the basis for key clinical decisions in hematology. However, visual inspection of aspirate specimens is a tedious and complex process subject to variation in interpretation, and hematopathology expertise is scarce. The ability to generate a compact representation of an aspirate specimen may form the basis for clinical decision support tools in hematology. We have previously published an end-to-end AI-based system for counting and classifying cells from bone marrow aspirate WSI. Using deep embeddings from this model, we construct bags of individual cell features from each WSI, and apply multiple instance learning to extract vector representations for each WSI. Using these representations in vector search, we achieved 0.58 ± 0.02 mAP@10 in WSI-level image retrieval, which outperforms the Random baseline (0.39 ± 0.1). Using a weighted k-nearest-neighbours (k-NN) model on these slide vectors, we predict five broad diagnostic labels on individual aspirate WSI with a weighted-macro-average F1 score of 0.57 ± 0.03 on the test set of 278 randomly sampled WSIs, which outperforms a classifier using empirical class prior probabilities (0.26 ± 0.02). We present the first example of exploring trainable mechanisms to generate compact, slide-level representations in bone marrow cytology with deep learning. This method has the potential to summarize complex semantic information in WSIs toward improved diagnostics in hematology, and may eventually support AI-assisted computational pathology approaches.

Applied machine learning in hematopathology

“…48 These semantic labels and bags of individual cell features were used to train a multiple instance learning (MIL) model, where attention-pooling was applied to generate WSI-level representations. 49 Such approaches could potentially be used for simple slide-level diagnosis or patient matching in hematopathology.…”

“…To overcome this issue, deep learning natural language processing (NLP) models for processing BMA synopsis have been proposed to report to generate vector‐based slide‐level “semantic labels” 48 . These semantic labels and bags of individual cell features were used to train a multiple instance learning (MIL) model, where attention‐pooling was applied to generate WSI‐level representations 49 . Such approaches could potentially be used for simple slide‐level diagnosis or patient matching in hematopathology.…”