VN October 2025

Vetnews | Oktober 2025 20 « BACK TO CONTENTS Instead, prompts need to be rephrased in simpler terms, such as “watery eyes” or “eye inflammation,” to ensure the model can process and detect the relevant symptoms. Figure 15. [63] demonstrates how the structure and content of the text prompt significantly influence the model’s output, highlighting the importance of careful prompt formulation. 3.2.4 Single Shot Multibox Detector (SSD) for Eye Detection The Single Shot Multibox Detector (SSD) [90] was utilised to pinpoint the eye region in canine images, serving as an assisting tool for the improved segmentation task. SSD stands out for its real-time performance, accomplishing object localisation and classification simultaneously in a single forward pass of the network. This efficiency is achieved by omitting a separate region proposal generation stage, a common feature in other detection systems like Mask R-CNN, which simplifies the detection process and reduces computational overhead. The architecture of SSD is built upon a robust base convolutional neural network, which is employed for feature extraction from the input image. Following this feature extraction phase, SSD utilises additional convolutional layers to produce feature maps at multiple scales. This multi-scale feature extraction allows SSD to detect objects of various sizes effectively. Key components of the SSD workflow include: • Feature Extraction: which utilises a pretrained network like VGG16 or ResNet to derive rich feature maps from the input image. These features form the foundation for detecting objects at various scales. • Detection Heads: a segment where each feature map level has associated detection heads that output scores and bounding boxes for potential objects at that scale. • Non-Maximum Suppression (NMS): this step refines the detection by removing overlapping boxes, ensuring that each detected object is represented by the single best bounding box. The process flow of the SSD model is illustrated in the provided diagram in Figure 16, which clarifies the stages from image input through to the final predictions. The diagram shows how the image is processed through the VGG16/ResNet layers for feature extraction, followed by multi-scale detection heads, and culminating in the NMS stage to produce the final object predictions. Given the absence of publicly available models for detecting canine eyes, a custom SSD model was trained specifically for this task. The training dataset consisted of 99 images of dog eyes from various breeds and medical conditions, including healthy eyes, each annotated with ground truth bounding boxes in XML format. This approach was designed to ensure that the model could accurately identify eye regions across diverse scenarios. The performance of the custom-trained SSD model was evaluated with impressive outcomes. The model achieved an average Intersection over Union (IoU) of 0.92 with a standard deviation of 0.08, and an average Dice Score of 0.96 with a standard deviation of 0.05. An F1 score of 0.9, calculated using Precision (0.97) and Recall (0.84), defines a well-suited detection model. These results highlight the SSD model’s high precision in localising the eye region Leading Article Figure 14. Grounding Dino and SAM integration flowchart a b c d Figure 15. Grounding Dino prompting examples showing diversity with resulting detection and its prediction score. Prompts used for generating detection from top to bottom are: (a) "colored mass protrusion in the corner of the eye", (b) "colored mass protrusion in the eye" and (c) "red mass protrusion". Third prompt (c) is used for segmentation by SAM implementation (d). Figure 16. Single Shot Multibox Detector diagram for eye detection

RkJQdWJsaXNoZXIy OTc5MDU=