Work

41 Automatic corneal cell detection and semi-supervised segmentation on histological images using deep learning techniques

Year: 2026

Type: conference-abstract

Abstract: Purpose The analysis of corneal cultures is key for determining its suitability for transplantation. Currently, it is performed by manually counting cells to estimate the density, as well as studying the morphology of the cells, on high resolution histological images of the cornea. This is a time-consuming task that can also be affected by human subjectivity, motivating the design of automatic methods that can reduce the workload of the experts and potentially improve the quality of the analyses.Methods In this work we propose an automatic method, based on Deep Learning techniques, for the detection of corneal cells on histological images, trained using pseudo-centroids annotated by the experts. Some areas that may be ambiguous due to blurriness were also indicated. In this context, we also propose a refinement method for these excluded areas in order to improve the robustness of the proposed methods. In addition, we address the cell segmentation task with self-supervised techniques, removing the necessity of any segmentation annotations from the experts.Results The achieved results show that the proposed detection model is adequate for cell-counting tasks, estima-ting an average of 0.004085 cells per micron on the test set with a ground truth of 0.004098, committing an error of only 0.38% with respect to the annotated ground truth. In addition, the proposed detection model is able to provide reasonable detections in the excluded abmiguous areas. Moreover, our self-su-pervised segmentation approach achieves satisfactory qualitative results, allowing a feasible morphologi-cal analysis of the cells.Conclusions Our proposed methods successfully estimate the cell density on histological images of corneal cultures, achieving a performance similar to the experts. It is expected that the proposed methods will be leveraged in the clinics, as it provides a significant improvement in a tedious repetitive task such as cell-counting, ne-cessary for the assessment of tissue quality. In addition, we also expect that our proposed self-supervised segmentation model will be of help for the clinicians in more complex tasks such as the study of corneal cell morphology.Abstract 41 Figure 1

Source: BMJ Open Ophthalmology

DOI: 10.1136/bmjophth-2026-eeba.41