Detection of Decidual Vasculopathy using Multiresolution Hierarchical Neural Network

Introduction:: Microscopic analysis of the human placenta has the potential to be very useful to determine the anatomic basis of pregnancy-related complications.¹ Preeclampsia affects 2% to 8% of pregnancies and is the leading cause of preterm birth and related neonatal morbidity in the developed world.^2,3 In particular, the presence of a placental lesion called decidual vasculopathy (DV) is often found in cases of preeclampsia, and when detected in an uncomplicated pregnancy, DV has been correlated to the occurrence of preeclampsia and other adverse outcomes in subsequent pregnancies.^4,5 Due to resource constraints, most placentas from uncomplicated pregnancies are discarded with no microscopic inspection being performed.⁶ This work seeks to locate blood vessels, identifying diseased blood vessels on Whole Slide Images(WSIs) and predicting the overall preeclampsia risk for the placenta.⁷ Our algorithm also incorporates patient metadata which makes the algorithm more accurate for the predictions. This study builds off previous work through examining the ability to implement different scanner images that are found across the nation like Hologic Genius(.tif) and Hamamatsu NanoZoomer(.ndpi), along with different image resolutions and the robustness to image perturbations.

Materials and Methods:: RetinaNet localization framework, with ResNet backbone feature extractor ,initialized with pre-trained weights from MS-COCO Dataset is used to build the object detection neural network which detects blood vessels on stain normalized images(Macenko et.al⁸). XCeption CNN backbone ,initialized with ImageNet data and optimized with Adam Optimizer⁹, dropout¹⁰ is used to build classification neural network which classifies the detected blood vessels into “Diseased” and “Healthy”. The pipelines used in this study were trained, validated and tested on 181 archival placenta cases, using whole slide image(WSIs) that were scanned glass slides with hematoxylin and eosin stained tissue sections. The slides were cut at 4 to 5 mm thickness and scanned at x20 magnification on an Aperio AT whole slide scanner (Leica Biosystems, Wetzlar, Germany), resulting in digital slides with a resolution of 0.50 mm/pixel, using bright-field microscopy. WSIs from Hologic Genius, Hamamatsu NanoZoomer, AT2(.svs) Image Scanners at University of Michigan , using the same glass slide procedure, were used to test the generalizability of the model. Experiments involving tiling the images at different resolution levels (Level2 and Level1, with 15% overlap) of WSI Images, perturbing hyper-parameters (i.e., confidence thresholds) for neural networks, using Blur images (boxblur: radius of 4) and eliminated blood vessel images were conducted. The WSI slides were coupled with Neural Network built on python using OpenSlide Library.

Results, Conclusions, and Discussions:: Results and Discussion:The overall aim of this work was to investigate generalizability of the prior deep learning pipeline on multi-cellular placenta images through experiments involving multiscanner images, multi-resolution images, varying confidence thresholds, and blood-vessel eliminated images. The model’s blood vessel detections are very accurate (Fig1a), but despite having similar resolution at same level, the model predicted twice as many predictions for .tif than .ndpi, while the least total predictions were obtained for .svs images at Level2 (Fig1b). The higher resolution tiled images resulted in the higher Healthy Vessel Ratio (HVR) over Diseased Vessel Ratio (DVR), but the number of diseased predictions were reduced (Fig1b). While for .svs images the number of diseased predictions increased (Fig1b). A higher confidence threshold enabled more accurate predictions, while the lower threshold resulted in a higher number of inaccurate predictions. Also, when the input images were perturbed with few blood vessel eliminations, it gave inaccurate predictions but the number of total predictions substantially decreased than before, as expected.

Conclusion: This study shows that image perturbations linked to changing resolution, hyper-parameters, and removing relevant features affect the prediction outcomes of blood vessels and its classification. The observed difference in reduction of .tif and .ndpi diseased vessel predictions, and increase of .svs diseased predictions with resolution indicate that model’s weights are more biased towards images on which it was trained. The model produces more predictions for .tif and .ndpi files than for .svs files, as higher resolution tiles are fed into the model for .tif and .ndpi files, resulting in more detailed images and more predictions. Among .tif and .ndpi images, .tif are more likely to generate more predictions due to their slightly higher pixel resolution at the same magnification level. Collectively, these findings suggest that model’s trained weights are highly sensitive to resolution of image, while image features and pixel intensities play a critical role in performing blood vessel detections and classification processes.

Acknowledgements (Optional): :

References (Optional): : 1. Salafia, CM., Vintzileos, AM.: Why all placentas should be examined by a pathologist. Am J Obstet Gynecol 1990, 163:1282e1293.

2. ACOG practice bulletin no. 202: gestational hypertension and preeclampsia. Obs Gynecol 2019, 133:e1ee25.

3. Roberts, JM., August, PA., Bakris, G., Barton, JR., Bernstein, IM., Druzin, M., Gaiser, RR., Granger, JP., Jeyabalan, A., Johnson, DD.: Hypertension in pregnancy: executive summary. Obstet Gynecol 2013, 122:1122e1131.

4. Stevens, DU., Al-Nasiry, S., Bulten, J., Spaanderman, MEA.: Decidual vasculopathy and adverse perinatal outcome in preeclamptic pregnancy. Placenta 2012, 33:630e633.

5. Hauspurg, A., Redman, EK., Assibey-Mensah, V., Tony Parks W., Jeyabalan, A., Roberts, JM., Catov, JM.: Placental findings in non-hypertensive term pregnancies and association with future adverse pregnancy outcomes: a cohort study. Placenta 2018, 74:14e19.

6. Hecht, JL., Zsengeller, ZK., Spiel, M., Karumanchi, SA., Rosen, S.: Revisiting decidual vasculopathy. Placenta 2016, 42:37e43.

7. Clymer, D., Kostadinov, S., Catov, J., Skvarca, L., Pantanowitz, L., Cagan, J., and Leduc, P. Decidual Vasculopathy Identification in Whole Slide Images Using Multiresolution Hierarchical Convolutional Neural Networks. The American Journal of Pathology 190(10), 2111–2122 (2020).

8. Macenko, M., Niethammer, M., Marron, JS., Borland, D., Woosley, JT., Guan, X., Schmitt, C., Thomas, NE.: A method for normalizing histology slides for quantitative analysis. Boston, MA: IEEE International Symposium on Biomedical Imaging: From Nano to Macro, 2009. pp. 1107e1110.

9. Kingma, DP., Ba, J.: Adam: a method for stochastic optimization. arXiv 2015, arXiv:1412.6980v9.

10. Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., Salakhutdinov, R.: Dropout: a simple way to prevent neural networks from overfitting. J Mach Learn Res 2014, 15:1929e1958.