Human-in-the-Loop (HITL) Learning for Identifying Glioblastoma Hallmarks on H&E Slides

Glioblastoma (GB), an aggressive brain tumor with abysmal survival rates, is diagnosed through histopathological examination of hematoxylin and eosin (H&E) stained images. Preservation methods post-resection for H&E slides involve frozen or paraffin-embedded sections, each with trade-offs in diagnosis time and image quality. Regardless of the type of sectioning, visual inspection of H&E-stained whole slide images (WSI) is subjective and time-consuming. We present a deep learning (DL) approach to classify two primary GB hallmarks, cellular tumor from necrotic regions, utilizing a human-in-the-loop (HITL) strategy to address the challenges with variations in WSI across different tissue sections. Specifically, our HITL approach involves actively selecting patches that are identified as most informative by the DL model using uncertainty measurements, while the ones identified to have low certainty, undergo an expert pathologist evaluation. Our HITL model using an EfficientNetB0 architecture, when trained with frozen-embedded tissue samples and tested on paraffin-embedded samples, achieved an accuracy of 0.901 and AUC of 0.944 in distinguishing necrosis from cellular tumor regions. However, by enriching the training set with both frozen- and paraffin-embedded samples along with HITL, the model’s performance demonstrated promising improvements yielding an accuracy of 0.966 and an AUC of 0.995. Our results suggest that training sets with the inclusion of samples with different tissue sectioning may yield more robust DL models for pathology classification.