The advent of light-sheet fluorescence microscopy (LSFM) allows for the sequential capture of 3D zebrafish cardiac events with a high spatial and temporal resolution. To use these images for quantified structural analysis, the heart tissue must be segmented out from each slice as a binary image. Despite being the gold standard, manual segmentation is profoundly labor-intensive and prone to human error, and therefore, an efficient and robust automatic method is needed for medical image analysis of large volumes of data. While traditional machine learning methods such as convolutional neural networks (CNNs) have the capability to produce precise segmentation results, they require a high volume (>1000) of manually segmented annotated images for training purposes, which is often not feasible for specialized imaging modalities. In this regard, we propose (1) integrating novel machine learning methods with optical techniques for robust data-efficient image segmentation and (2) incorporating the proposed method to assess basic morphological consequences of doxorubicin treatment on zebrafish cardiac development.
3-month old zebrafish (n=7) were injected intraperitoneally with a 20 µg/g dose of doxorubicin and imaged 30 days later using the LSFM alongside a control group (n=6). Our proposed machine learning method is composed of two parts—feature extraction and classification. For feature extraction, we split each image into patches and applied the subspace approximation with augmented kernels (SAAK) transform (Collaboration with Dr. Jay Kuo at USC), a multi-stage mathematical operation based on principle component analysis. Next, the outputs of the SAAK transform were passed to a random forest classifier along with the pixel intensities and locations. We applied the trained algorithm to segment the adult zebrafish heart and validated the results with manually segmented data. In parallel, we integrated a spiral phase filter into our in-house LSFM system to obtain edge enhanced images. We used the segmented images and the edge images to calculate the volume and surface area of heart tissue respectively.
Validation with manually segmented patches yielded an average dice similarity coefficient (DSC) of 0.86 ± 0.03 for the segmentation results of the SAAK transform, compared to 0.82 for a RefineNet implementation. Additionally, the SAAK transform approach was robust against perturbations and translational variance, maintaining 97.37% classification accuracy in the face of adversarial attacks. The mean surface area to volume ratio was 0.14 ± 0.03 µm-1 for the control group and 0.10 ± 0.03 µm-1 for the doxorubicin-injected group, yielding a statistically significant difference (p = 0.029).
The DSC suggests that with only 18 training images, the SAAK transformation segmentation results are accurate and exceed the quality produced by the CNN, which comparatively, requires over 50 times the training data. The decreased surface area to volume ratio observed in the doxorubicin group is a potential marker of reduced trabeculation. Qualitatively, zebrafish hearts are more intricately patterned than human organs, suggesting that this segmentation method holds significant promise for medical imaging applications. Moreover, we demonstrate the application-dependent optionality of the spiral phase filter—(1) integration in the LSFM to reduce noise and enhance details in the segmentation result at the cost of a slightly reduced dice coefficient or (2) application after segmentation to analyze boundary-defined structural parameters like surface area. Together, the SAAK transform and the spiral phase filter extract meaningful features from minimal data, addressing the major imaging hurdle of efficiency.