A BERT model generates diagnostically relevant semantic embeddings from pathology synopses with active learning

Abstract

Background Pathology synopses consist of semi-structured or unstructured text summarizing visual information by observing human tissue. Experts write and interpret these synopses with high domain-specific knowledge to extract tissue semantics and formulate a diagnosis in the context of ancillary testing and clinical information. The limited number of specialists available to interpret pathology synopses restricts the utility of the inherent information. Deep learning offers a tool for information extraction and automatic feature generation from complex datasets. Methods Using an active learning approach, we developed a set of semantic labels for bone marrow aspirate pathology synopses. We then trained a transformer-based deep-learning model to map these synopses to one or more semantic labels, and extracted learned embeddings (i.e., meaningful attributes) from the model’s hidden layer. Results Here we demonstrate that with a small amount of training data, a transformer-based natural language model can extract embeddings from pathology synopses that capture diagnostically relevant information. On average, these embeddings can be used to generate semantic labels mapping patients to probable diagnostic groups with a micro-average F1 score of 0.779 Â ± 0.025. Conclusions We provide a generalizable deep learning model and approach to unlock the semantic information inherent in pathology synopses toward improved diagnostics, biodiscovery and AI-assisted computational pathology.

Whole slide image representation in bone marrow cytology

“…In terms of the weaknesses of our system, WSI labels came from our previously published BERT model's predictions [46] rather than experts, which are not perfectly correct. This was implemented as it is not practically feasible to manually label many hundreds of WSI with keywords from semi-structured diagnostic synopses.…”

“…Labels were created by simplifying the predictions of a fine-tuned BERT model on WSI synopses, as previously reported by our group [46]. This BERT model’s predictions took the form of a multi-label task.…”

“…SupCon loss is also robust to noisy labels and more stable to hyperparameter settings like optimizers and data augmentations [59]. Because our labels come from our previously published BERT model's predictions [46] rather than experts, we must take its prediction errors into consideration. SupCon loss's tolerance to noisy labels became an obvious option for this training.…”

Automatic symptoms identification from a massive volume of unstructured medical consultations using deep neural and BERT models

“…When the data is textual, the extraction of features is a bit different where the aim is to create word or text embeddings. Generally in medical NLP and at the feature extraction level, the BERT (as a state-of-the-art model) was used in several studies as in [14] , [15] , [16] , [17] , whereas, the Word2Vec model was used in [18] , [19] , [20] . In contrast, and at the algorithmic level, various deep learning models were widely used in medical NLP.…”

Automatic Symptoms Identification from Massive Volume of Unstructured Medical Consultations Using Deep Neural and BERT Models

“…When the data is textual, the extraction of features is a bit different where the aim is to create word or text embeddings. Generally in medical NLP and at the feature extraction level, the BERT (as a state-of-the-art model) was used in several studies as in [14,15,16,17], whereas, the Word2Vec model was used in [18,19,20]. In contrast, and at the algorithmic level, various deep learning models were widely used in medical NLP.…”