Cell Detection with Star-convex Polygons (1806.03535v2)

Abstract: Automatic detection and segmentation of cells and nuclei in microscopy images is important for many biological applications. Recent successful learning-based approaches include per-pixel cell segmentation with subsequent pixel grouping, or localization of bounding boxes with subsequent shape refinement. In situations of crowded cells, these can be prone to segmentation errors, such as falsely merging bordering cells or suppressing valid cell instances due to the poor approximation with bounding boxes. To overcome these issues, we propose to localize cell nuclei via star-convex polygons, which are a much better shape representation as compared to bounding boxes and thus do not need shape refinement. To that end, we train a convolutional neural network that predicts for every pixel a polygon for the cell instance at that position. We demonstrate the merits of our approach on two synthetic datasets and one challenging dataset of diverse fluorescence microscopy images.

Cell Detection with Star-convex Polygons

The paper "Cell Detection with Star-convex Polygons" presents a novel approach to the task of detecting and segmenting cells and nuclei in microscopy images. The authors propose a method that leverages star-convex polygons as the shape representation, hypothesizing that this approach will perform better in conditions where cells are densely packed or have irregular shapes.

Methodology

The core innovation is the use of star-convex polygons, which offer a more flexible and accurate shape representation compared to traditional bounding boxes. The authors train a convolutional neural network (CNN) to predict a polygon for each pixel, effectively transforming each pixel into a potential cell nucleus candidate. Specifically, a U-Net architecture is employed to densely predict object probabilities ($d_{i,j}$) and the radial distances ($r^k_{i,j}$) that parameterize the star-convex polygons.

For the training phase, a binary cross-entropy loss is used for object probability predictions, and a mean absolute error loss is used for the polygon distances, weighted by the ground truth object probabilities. This weighting ensures that the center pixels, which are more likely to be accurately represented, contribute more significantly to the loss.

Evaluation and Results

The approach is evaluated on two synthetic datasets and one challenging dataset of diverse fluorescence microscopy images. The authors compare their method against popular segmentation techniques, including U-Net with two and three output classes and Mask R-CNN, a state-of-the-art instance segmentation method.

Numeric Performance

The results are quantified using the average precision (AP) across several intersection over union (IoU) thresholds:

• Dataset A (Synthetic, touching ellipses): For moderate IoU thresholds ($\tau < 0.7$), the proposed star-convex polygon method achieves perfect or near-perfect results. Mask R-CNN underperforms due to overlapping instances, which highlights a fundamental limitation of bounding-box based methods in certain conditions.
• Dataset B (Synthetic, evolving cell population): The star-convex polygon method demonstrates almost perfect accuracy for many IoU thresholds, showcasing its robustness in handling dense and crowded cell configurations.
• Dataset C (Real fluorescence microscopy images): The proposed method surpasses other methods for IoU thresholds of $\tau < 0.75$, and maintains reasonable performance even for higher thresholds. The 2-class U-Net baseline shows limitations when dealing with touching cells, often merging them erroneously.

Implications and Future Directions

The use of star-convex polygons introduces a significant improvement in the detection and segmentation of cell nuclei, particularly in complicated scenarios with high cell density. This method avoids common pitfalls associated with bounding-box based approaches, such as erroneously merging cells or failing to accurately approximate cell shapes without extensive post-processing.

Theoretical Implications: This paper's results suggest that direct shape prediction models can significantly enhance detection accuracy, particularly in biological imaging applications. It also opens avenues for exploring other shape representations that may offer similar or better performance in different contexts.

Practical Implications: For practitioners, the proposed method offers a simpler yet effective alternative to more cumbersome approaches like Mask R-CNN. Its fewer parameters and ease of training make it appealing for rapid deployment in various biological studies, including high-content screening and developmental biology.

Future Developments: Future research could focus on extending this method to multi-class detection scenarios and integrating it into cell tracking systems. The ability of the method to predict plausible shapes even for partially visible objects at image boundaries presents exciting opportunities for improving cell tracking algorithms, particularly in time-lapse microscopy where cells move in and out of the field of view.

Conclusion

This paper presents a compelling case for the use of star-convex polygons in cell detection and segmentation tasks. Through rigorous evaluation, the authors demonstrate that their proposed method offers superior performance in complex and crowded environments, making it a valuable tool for the biological imaging community. The method's simplicity, accuracy, and robustness underscore its potential for broad application and future enhancements in the domain of cell imaging.