(B-48) Morphological similarity measure for better tumor tissue sampling

Introduction:: While most cancers are of monoclonal origin, the acquisition of mutations by different cells in an expanding population leads to diversification into subclones and dramatically increases genetic tumor heterogeneity [1]. This heterogeneity makes it difficult to effectively characterize tumors or predict their future behavior. Currently, the multiregional sequencing experiments are the gold standard for understanding intra-tumor heterogeneity. However, such experiments are expensive, and when sampling randomly, a large number of areas must be sampled to capture the full range of genetic states present [2]. Here, we hypothesize that tumor tissue morphology is a reflection of an underlying genetic state, and by applying deep learning to hematoxylin and eosin (H&E) stained tissue slides, we can more efficiently sample intra-tumor heterogeneity. To this end, we present a deep learning model that we have developed to measure morphological similarity of histopathological images. We posit that by preferentially sampling morphologically distinct areas, we will be able to capture the driver mutations in a patient with fewer areas needing to be sampled. Moreover, the model allows for an easy and fast description of cancer tissue morphology, and therefore enables the exploration genetic and morphologic heterogeneity.


Materials and Methods:: We chose clear cell renal cell carcinoma (ccRCC) as our model system, as it is the paradigm for intra-tumor heterogeneity. To develop a measure of morphologic similarity, we sought to train deep learning models that take as input H&E stained image patches and generate vector profiles such that profiles of patches from spatially adjacent tumor areas are similar. We trained the model using the BYOL deep learning framework, which, unlike more computationally heavy contrastive approaches that need large banks of dissimilar patches, only needs examples of similar image patches (Fig. 1A) [3]. We considered a range of model architectures and input types. For encoders, we tested both Convolutional Neural Network (ResNet50), as well as Vision Transformer (ViTb). The input consisted of a pair of patches that were: augmentation of the same patch (same), augmentation of patches that were close spatially on a slide (close), or a combination of the two (combined). We benchmarked these approaches based on how frequently the neighbors of a given patch came from a spatially adjacent area (Fig. 1B). Next, we applied our winning model to profile data from a multiregional sequencing experiment, where we have H&E images and genetics from multiple areas per patient. Now, based on the morphological similarity of the samples, we are in the process of devising a sampling strategy that captures the largest number of driver mutation in a patient with the smallest number of areas sampled.


Results, Conclusions, and Discussions:: We found that we achieved the best performance with a Vision Transformer (ViTb) backbone, where BYOL training was performed with similar patches being spatially adjacent (either combined with the augmented versions of the same patch or not). This model leads to a major improvement over BYOL training with just augmentations of the same patch, or over using pretrained ViTb and ResNet50 models, which are the more traditional approaches. Tissue morphology is complex, and manually quantifying similarity is both challenging and subjective. Because of that, the ability of our methodology to retrieve similar patches is highly encouraging. Based on this performance, when applied to whole slide images, we expect the model to be able to infer which parts of tumor tissue are most similar morphologically, which we hypothesize also reflects genetic similarity. Therefore, as an application of the similarity measure, we are developing a tumor sampling methodology that identifies the optimal strategy to capture heterogeneity of tumor tissue with a minimum number of samples. The basic idea of this strategy is to pick samples to cover the broadest range of morphologies. For example, in Figure 2A, we plot the relationship between morphological profiles of six samples captured from a single patient. If we had to pick three samples, we would thus pick one in each of the distinct morphological areas as illustrated here. Our current work is focused on developing a more general sampling strategy for any number of areas sampled, and we hypothesize that by doing so, we will capture important driver mutations better than when sampling randomly (Fig. 2B). If successful, this approach could leverage morphology to increase the efficiency of multiregion sequencing experiments. More broadly, as we recognize that there is a need for further analysis of how morphology and genetics correlate with each other, we propose to utilize our deep learning model to further explore links between changes in morphology and the mutational landscapes.


Acknowledgements (Optional): : This work was supported by The Cancer Prevention and Research Institute of Texas (CPRIT): RP220294, RP210041.

References (Optional): : [1] M. Gerlinger and C. Swanton, "How Darwinian models inform therapeutic failure initiated by clonal heterogeneity in cancer medicine," British Journal of Cancer, vol. 103, no. 8, pp. 1139-1143, 2010.
[2] S. Turajlic, H. Xu, K. Litchfield, A. Rowan, T. Chambers, J. Lopez, D. Nicol, T. O'Brien, L. J, S. Horswell, M. Stares, L. Au, M. Jamal-Hanjani, B. Challacombe, A. Chandra, S. Hazell, C. Eichler-Jonsson, A. Soultati and S. Chowdhury, "Deterministic Evolutionary Trajectories Influence Primary Tumor Growth: TRACERx Renal," Cell, vol. 173, no. 3, pp. 581-594, 2018.
[3] J. Grill, F. Strub, F. Altché, C. Tallec, P. H. Richemond, E. Buchatskaya, C. Doersch, B. A. Pires, Z. Daniel Guo, M. G. Azar, B. Piot, K. Kavukcuoglu, R. Munos and M. Valko. “Bootstrap your own latent: A new approach to self-supervised Learning,” arXiv, 2020