Poster L29 - Advanced Computational Approaches in High-Density Cell Segmentation and Tracking for Cancer Metastasis Research

Introduction: Cancer metastasis, the primary cause of mortality, involves complex stages from invasion to the development of secondary tumors. A crucial aspect, interfacial invasion, entails interactions between cancer cells and the extracellular matrix (ECM). Understanding these interactions is vital for developing early therapeutic strategies. The spheroid interfacial invasion model is instrumental for simulating tumor microenvironments to study these mechanisms. However, while differential interference contrast (DIC) microscopy is commonly employed for analyzing invasive cells, tracking these cells in DIC images is challenging. This difficulty stems from the labor-intensive nature of handling large cell quantities.

Cell tracking typically consists of two tasks: single cell segmentation and subsequent tracking. Numerous efficient models and algorithms have been developed for these tasks. For instance, nnU-Net and the recently introduced U-Mamba excel in medical image segmentation, while cell tracking often utilizes algorithms based on the Hungarian method and clustering techniques. However, these models and algorithms are primarily effective in simple scenarios with widely spaced, few cells, and rare cell division behaviors. Their performance significantly diminishes in complex scenarios characterized by dense cell populations and frequent cell divisions.

In response to these challenges, we propose a new pipeline designed to handle high-density cell segmentation and tracking. This pipeline accurately segments cells in densely populated images and tracks them in dense segmentation results, identifying potential splitting behaviors. By seamlessly integrating these components, our pipeline automates the processing of cell images captured over time intervals, significantly reducing the need for manual tracking and workload.

Materials and

Methods: * Data Acquisition

We used DIC microscopy to observe the invasion of MDA-MB-231 tumor spheroids. Cultured for four days in microwells, they were embedded in type I collagen scaffolds from rat tails with varied interfaces. Over 24 hours, we recorded spheroid invasion at ten-minute intervals. Post-imaging, the cells were manually annotated with polygonal shapes.

* Model Development and Evaluation

Our pipeline, shown in Figure 1, consists of two main components: single cell segmentation module and cell tracking module. It starts with segmenting cells from images to pinpoint single cell positions via the first module, followed by tracking these cells across sequential images using the tracking module.

** Single Cell Segmentation Module

We employed different configurations of the YOLO v8, with increasing parameter sizes: v8n, v8s, v8m, v8l, and v8x. The dataset includes 194 training images and 10 validation images. Given the dataset's limited size, we applied data augmentation techniques like HSV adjustment and geometric transformations. We utilized Dice loss for segmentation, complemented by Distribution Focal Loss for bounding boxes, and implemented early-stopping based on the highest validation precision achieved.

** Cell Tracking Module

The tracking algorithm uses Intersection of Union (IoU) to compute overlaps between cells at consecutive time points. To improve tracking of closely packed or irregularly shaped cells, we introduced the Minimum Area Coverage Ratio (MACR), which better detects and tracks dividing cells by analyzing the overlap relative to the smallest object area. Additionally, morphological dilation was found to enhance tracking accuracy, with results displayed in Figure 4.

Results, Conclusions, and Discussions: * Results

Figure 2 presents the comparison results of various YOLO v8 models for single cell segmentation. Notably, the YOLO v8n model is among the best, achieving the highest precision at 0.80 and mAP50 at 0.67. The visualization of segmentation results by YOLO v8n is illustrated in Figure 3.

Additionally, the integration of IoU with our proposed MACR and morphological dilation techniques significantly enhanced tracking precision. This was particularly effective for cells with complex or atypical morphologies, as demonstrated in Figure 4. Expert evaluations confirmed the high accuracy of our tracking method, noting a strong correspondence between predicted and ground truth data over various time points.

* Discussions

In this study, we employed the YOLO v8 model for single-cell segmentation. Remarkably, the YOLO v8n variant, which has the least number of parameters, delivered the highest precision and mAP50. Contrary to expectations in deep learning, where a greater number of parameters typically suggests enhanced capability to model complex functions and capture intricate data features, the smaller model outperformed its larger counterparts. This outcome highlights potential areas for improvement in larger models, possibly through further dataset expansion or better adaptation to our experimental conditions.

In the Tracking Module, the application of MACR and morphological dilation was crucial for accurately tracking cells with unique morphological characteristics. As clearly shown in Figure 4, the tracking algorithm exhibited various errors when either MACR or morphological dilation was not applied. However, when both operations were performed simultaneously, all errors were eliminated.

* Conclusions

Our models have demonstrated strong performance in single-cell segmentation and tracking tasks. There is potential to further boost model performance by optimizing training processes and enlarging the dataset. Such refinements could enable more precise analyses of cellular behavior, contributing deeper insights into cancer biology. This study not only advances the field of cancer research but also establishes a foundation for future investigations that could extend these findings, enhancing what is achievable in computational biology and cellular imaging.

Acknowledgements (Optional):