Benchmarking Self-Supervised Contrastive Learning Methods for Image-Based Plant Phenotyping

The rise of self-supervised learning (SSL) methods in recent years presents an opportunity to leverage unlabeled and domain-specific datasets generated by image-based plant phenotyping platforms to accelerate plant breeding programs. Despite the surge of research on SSL, there has been a scarcity of research exploring the applications of SSL to image-based plant phenotyping tasks, particularly detection and counting tasks. We address this gap by benchmarking the performance of 2 SSL methods-momentum contrast (MoCo) v2 and dense contrastive learning (DenseCL)-against the conventional supervised learning method when transferring learned representations to 4 downstream (target) image-based plant phenotyping tasks: wheat head detection, plant instance detection, wheat spikelet counting, and leaf counting. We studied the effects of the domain of the pretraining (source) dataset on the downstream performance and the influence of redundancy in the pretraining dataset on the quality of learned representations. We also analyzed the similarity of the internal representations learned via the different pretraining methods. We find that supervised pretraining generally outperforms self-supervised pretraining and show that MoCo v2 and DenseCL learn different high-level representations compared to the supervised method. We also find that using a diverse source dataset in the same domain as or a similar domain to the target dataset maximizes performance in the downstream task. Finally, our results show that SSL methods may be more sensitive to redundancy in the pretraining dataset than the supervised pretraining method. We hope that this benchmark/evaluation study will guide practitioners in developing better SSL methods for image-based plant phenotyping.

Automatic detection and segmentation of bovine corpora lutea in ultrasonographic ovarian images using genetic programming and rotation invariant local binary patterns

In this study, we propose a fully automatic algorithm to detect and segment corpora lutea (CL) using genetic programming and rotationally invariant local binary patterns. Detection and segmentation experiments were conducted and evaluated on 30 images containing a CL and 30 images with no CL. The detection algorithm correctly determined the presence or absence of a CL in 93.33 % of the images. The segmentation algorithm achieved a mean (±standard deviation) sensitivity and specificity of 0.8693 ± 0.1371 and 0.9136 ± 0.0503, respectively, over the 30 CL images. The mean root mean squared distance of the segmented boundary from the true boundary was 1.12 ± 0.463 mm and the mean maximum deviation (Hausdorff distance) was 3.39 ± 2.00 mm. The success of these algorithms demonstrates that similar algorithms designed for the analysis of in vivo human ovaries are likely viable.

Level set segmentation of bovine corpora lutea in ex situ ovarian ultrasound images

BACKGROUND

The objective of this study was to investigate the viability of level set image segmentation methods for the detection of corpora lutea (corpus luteum, CL) boundaries in ultrasonographic ovarian images. It was hypothesized that bovine CL boundaries could be located within 1-2 mm by a level set image segmentation methodology.

METHODS

Level set methods embed a 2D contour in a 3D surface and evolve that surface over time according to an image-dependent speed function. A speed function suitable for segmentation of CL's in ovarian ultrasound images was developed. An initial contour was manually placed and contour evolution was allowed to proceed until the rate of change of the area was sufficiently small. The method was tested on ovarian ultrasonographic images (n = 8) obtained ex situ. A expert in ovarian ultrasound interpretation delineated CL boundaries manually to serve as a "ground truth". Accuracy of the level set segmentation algorithm was determined by comparing semi-automatically determined contours with ground truth contours using the mean absolute difference (MAD), root mean squared difference (RMSD), Hausdorff distance (HD), sensitivity, and specificity metrics.

RESULTS AND DISCUSSION

The mean MAD was 0.87 mm (sigma = 0.36 mm), RMSD was 1.1 mm (sigma = 0.47 mm), and HD was 3.4 mm (sigma = 2.0 mm) indicating that, on average, boundaries were accurate within 1-2 mm, however, deviations in excess of 3 mm from the ground truth were observed indicating under- or over-expansion of the contour. Mean sensitivity and specificity were 0.814 (sigma = 0.171) and 0.990 (sigma = 0.00786), respectively, indicating that CLs were consistently undersegmented but rarely did the contour interior include pixels that were judged by the human expert not to be part of the CL. It was observed that in localities where gradient magnitudes within the CL were strong due to high contrast speckle, contour expansion stopped too early.

CONCLUSION

The hypothesis that level set segmentation can be accurate to within 1-2 mm on average was supported, although there can be some greater deviation. The method was robust to boundary leakage as evidenced by the high specificity. It was concluded that the technique is promising and that a suitable data set of human ovarian images should be obtained to conduct further studies.

Evaluation of texture features for analysis of ovarian follicular development

We examined the echotexture in ultrasonographic images of the wall of dominant ovulatory follicles in women during natural menstrual cycles and dominant anovulatory follicles which developed in women using oral contraceptives (OC). Ovarian follicles in women are fluid-filled structures in the ovary that contain oocytes (eggs). Dominant follicles are physiologically selected for preferential development and ovulation. Statistically significant differences between the two classes of follicles were observed for two co-occurrence matrix derived texture features and two edge-frequency based texture features which allowed accurate distinction of healthy and atretic follicles of similar diameters. Trend analysis revealed consistent turning points in time series of texture features between 3 and 4 days prior to ovulation coinciding with the time at which follicles are being biologically "prepared" for ovulation.

Enhancing ultrasound texture differences for developing an in vivo 'virtual histology' approach to bovine ovarian imaging

A ‘virtual histology’ can be thought of as the ‘staining’ of a digital ultrasound image via image processing techniques in order to enhance the visualisation of differences in the echotexture of different types of tissues. Several candidate image-processing algorithms for virtual histology using ultrasound images of the bovine ovary were studied. The candidate algorithms were evaluated qualitatively for the ability to enhance the visual differences in intra-ovarian structures and quantitatively, using standard texture description features, for the ability to increase statistical differences in the echotexture of different ovarian tissues. Certain algorithms were found to create textures that were representative of ovarian micro-anatomical structures that one would observe in actual histology. Quantitative analysis using standard texture description features showed that our algorithms increased the statistical differences in the echotexture of stroma regions and corpus luteum regions. This work represents a first step toward both a general algorithm for the virtual histology of ultrasound images and understanding dynamic changes in form and function of the ovary at the microscopic level in a safe, repeatable and non-invasive way.