Heeyoung Ahn

Nuclear detection, segmentation and morphometric profiling are essential in helping us further understand the relationship between histology and patient outcome. To drive innovation in this area, we setup a community-wide challenge using the largest available dataset of its kind to assess nuclear segmentation and cellular composition. Our challenge, named CoNIC, stimulated the development of reproducible algorithms for cellular recognition with real-time result inspection on public leaderboards. We conducted an extensive post-challenge analysis based on the top-performing models using 1,658 whole-slide images of colon tissue. With around 700 million detected nuclei per model, associated features were used for dysplasia grading and survival analysis, where we demonstrated that the challenge's improvement over the previous state-of-the-art led to significant boosts in downstream performance. Our findings also suggest that eosinophils and neutrophils play an important role in the tumour microevironment. We release challenge models and WSI-level results to foster the development of further methods for biomarker discovery.

Rapid and accurate identification of fire fronts is essential for determining the extent and direction of wildfire spread, which is critical for developing effective suppression strategies.We created a comprehensive fire front dataset, containing 7,869 mask-labeled images extracted from helicopter footage captured during actual wildfire suppression operations.This study employed segmentation models based on transformer architectures to analyze helicopter imagery, specifically targeting regions of active fire and ground-level smoke.The Swin Transformer and Mask2Former models were trained, and their performances were compared.We evaluated model performance using both quantitative metrics and qualitative visual assessments.While the mean intersection over union scores achieved by the models were moderate, at 0.621 and 0.641, respectively, subsequent post-processing allowed for the extraction of fire front patterns closely aligned with real-world labeled data.Our future research aims to enhance dataset quality further and improve model accuracy, ultimately contributing to a robust and practical automated system for fire front detection.

Building segmentation models that can deal with rare and small nuclear objects in hematoxylin and eosin (H&E) stained pathologic images is a challenging task in digital pathology. Applying image augmentation can help alleviate this challenge. Hence, we propose new class-controlling copy-paste augmentation using a prepared nuclear objects set. Several image augmentations have been developed in computer vision for improving model performance; however, most of them are general-purpose methods and have not been designed for a specific domain. Our proposed method is appropriate for the pathology domain and provide strong regularization to make the model robust. In addition, it has the advantage of alleviating class imbalance problem, which is very common in histology datasets for nuclear segmentation. In our cross-validation experiments on a multi-tissue histology dataset, our method improves PQ and mPQ+ from 64.31 to 64.52 and 52.3 to 52.9, respectively.