Histopathological Evaluation of Abdominal Aortic Aneurysms with Deep Learning: The HistAAA Dataset

Creators

Kolbinger, Fiona R. (Researcher)
El Nahhas, Omar S. M. (Researcher)
Nackenhorst, Maja Carina (Researcher)
Christine, Brostjan (Researcher)
Wolf, Eilenberg (Researcher)
Albert, Busch (Researcher)
Kather, Jakob Nikolas (Researcher)

Description

This dataset accompanies our manuscript "Histopathological Evaluation of Abdominal Aortic Aneurysms with Deep Learning" (1) and comprises feature vectors extracted from digital whole-slide images (WSI) of abdominal aortic aneurysm (AAA) wall samples from 369 patients treated at three European centers and corresponding expert pathologist annotations. This dataset is intended to be used for computational modelling tasks including automated prediction of pathology-related variables such as inflammation, degradation of elastic fibers, and fibrosis (1).

Code for preprocessing is available at https://github.com/KatherLab/STAMP. Code for modeling is available at https://github.com/KatherLab/marugoto. Code for spatial heatmaps and top-attention tiles is available at https://github.com/KatherLab/highres-WSI-heatmaps/tree/AAA_heatmaps.

Patient Cohort

Data from a total of 369 patients (84.6% male, mean age 69.1 ± 8.0 years, average maximum diameter 62.9 ± 15.7 mm) undergoing open AAA repair at the Technical University Munich (TUM, n = 287), the University Hospital Würzburg (UHW, n = 36) and the Medical University Vienna (MUV, n = 46), between 2005 and 2019, are included in this dataset.

To generate annotations, aneurysm samples from the left anterior wall were independently analyzed by three pathologists as described previously (2). In brief, the following histopathological parameters were evaluated: Grade of inflammation in tunica media [none, minor, intermediate, major], grade of inflammation in adventitia [none, minor, intermediate, major], type of inflammation in adventitia and tunica media [none, acute, chronic], Histological Inflammation Scale of Aneurysms (HISA) grade (3) [0, 1, 2, 3, 4], angiogenesis in tunica media [present, not present], calcification in tunica media [present, not present], grade of fibrosis in adventitia [minor, intermediate, major], remaining elastic fibers in tunica media [< 25%, > 25%]. In case of disagreement, consensus was reached through discussion. Hematoxylin and Eosin (HE)- and Elastica van Gieson (EvG)-stained slides were digitized using an Aperio AT2 (Leica, Wetzlar, Germany) slide scanner. Patient sex and smoking history were gathered as binary clinical parameters.

Preprocessing and Feature Extraction

The STAMP protocol was utilized to process the WSIs (4). In brief, WSI were preprocessed by tessellation into 224 x 224 pixel patches at a magnification of 256 µm per pixel, followed by computational background rejection, and color normalization of HE-stained slides. EvG-stained slides were not color-normalized as no color normalization protocols exist for this staining. Subsequently, features were extracted using CTranspath (5), a pre-trained histology image encoder.

References

1. Kolbinger, F. R. et al. Histopathological Evaluation of Abdominal Aortic Aneurysms with Deep Learning. medRxiv 2024.04.23.24306178 (2024) doi:10.1101/2024.04.23.24306178.

2. Nackenhorst, M. C. et al. Abdominal aortic aneurysms harbor different histomorphology not associated with classic risk factors – the HistAAA study. medRxiv 2024.04.16.24305904 (2024) doi:10.1101/2024.04.16.24305904.

3. Rijbroek, A., Moll, F. L., von Dijk, H. A., Meijer, R. & Jansen, J. W. Inflammation of the abdominal aortic aneurysm wall. Eur. J. Vasc. Surg. 8, 41–46 (1994).

4. El Nahhas, O. S. M. et al. From Whole-slide Image to Biomarker Prediction: A Protocol for End-to-End Deep Learning in Computational Pathology. arXiv [cs.CV] (2023).

5. Wang, X. et al. Transformer-based unsupervised contrastive learning for histopathological image classification. Med. Image Anal. 81, 102559 (2022).