(J-392) Computer Vision to Predict Cell Seeding Coverage in Re-endothelialized Mouse Lungs

Introduction:: Transplantation of donor grafts recellularized with recipient-derived cells or non-immunogenic universal cell lines is a potential means of reducing the graft rejection and post-transplant complications in lung transplantation. Achieving a fully recellularized lung, however, remains a major limitation. A key parameter of evaluating recellularization efficacy in the field of tissue engineering is the cell seeding coverage (CSC), which is the ratio of seeded cell area to the total area of the lung scaffold. To calculate the CSC from a histological image, the lung scaffold and the seeded cell areas must be quantified. Semantic segmentation is a possible means of providing an end-to-end pixel-wise analysis of these parameters. In this work, the ability of semantic segmentation to accurately automate the work-intensive, time-consuming, and error-prone analysis of histological images is investigated. Specifically, the U-Net and LinkNet semantic segmentation models are applied to re-endothelialized mouse lung images, generating pixel-wise classifications of the lung scaffold and the seeded cell areas to predict the CSC. The performance of these models when trained on complete images and on image patches are compared.

Materials and Methods:: Raw histological images of mouse lungs seeded under varying boundary conditions were obtained, preprocessed, and annotated. The raw histological images contained noise in the form of dark regions, dust, and/or artifacts, as observed in Fig. 1a. Denoising operations were performed as required to remove these artifacts and markings, resulting in Fig. 1b. The denoised and enhanced image was then subjected to Otsu thresholding to separate high intensity pixels from low intensity pixels under the assumption that high intensity pixels were seeded cells and that low intensity pixels were tissue. However, heart tissue was also grouped in with the high intensity pixels, as shown in Fig. 1c. To correct this, a mask was manually drawn using the APEER^TM software by ZEISS to limit the area of analysis to lungs alone, as observed in Fig. 1d. The area of analysis mask and the Otsu-thresholded mask were then multiplied, generating the product mask observed in Fig. 1e.

Semantic segmentation was performed using the U-Net and LinkNet models. To train and test these models, k-folds cross validation was employed. Techniques including patch-based learning, data augmentation, and transfer learning were utilized to compensate for the low image quantity. Vanilla and patch-based models were evaluated and compared against one another in terms of their ability to classify lung scaffold and seeded cell pixels and in the accuracy of their CSC predictions.


Results, Conclusions, and Discussions:: The precision, recall, and intersection over union (IOU) of the vanilla and patch-based semantic segmentation models averaged across the four cross-validation iterations are summarized in Fig. 2. From the results reported in Table I, the patch-based U-Net model returns the lowest root mean square error (RMSE) of 1.57% ± 0.36% for cell seeding coverage, which is noteworthy because the vanilla U-Net returns the worst RMSE of 3.66% ± 4.33%. This improvement demonstrates the advantages provided by a patch-based learning approach when training on limited data. The patch-based LinkNet also shows an improvement over the vanilla LinkNet, reporting a 1.76% ± 0.31% RMSE over the 2.20% ± 0.36% RMSE.
In this work, lung scaffold and seeded cell classes were accurately segmented, and the CSC was accurately predicted in histological images of re-endothelialized mouse lungs via semantic segmentation. Though each of the semantic segmentation models performed satisfactorily, the accuracy of the patch-based models, particularly the patch-based U-Net model, outperformed the vanilla models in the pixel-wise classification of lung scaffold area and seeded cells and in the prediction of CSC. These results demonstrate a captivating proof-of-concept for semantic segmentation as a viable means of automating the pixel-wise analysis of re-endothelialized images, potentially saving the time and decreasing the workload of those working in this field while also avoiding potential errors that may arise from subjective user-specific manual input.


Acknowledgements (Optional): :

References (Optional): :