VN October 2025

Vetnews | Oktober 2025 24 « BACK TO CONTENTS The Dice Index is particularly useful in medical imaging as it penalises small discrepancies between the predicted and actual segmentation more strongly than IoU, thus providing a more stringent assessment of segmentation performance. • Pixel Accuracy: This metric calculates the proportion of correctly classified pixels across the entire image, both for foreground and background classes. The formula (13) is: Where TN refers to true negatives (Figure 21). While Pixel Accuracy offers a broad overview of the model’s ability to classify pixels, it may be less sensitive to rare conditions when most of the image consists of healthy tissue. • Sensitivity (Recall): Sensitivity, also known as the true positive rate, measures the proportion of actual positives (symptomatic regions) that the model correctly identifies. It is a critical metric in medical diagnostics because it indicates the model’s ability to detect all possible symptomatic areas. The formula (14) is: A high sensitivity score is crucial for ensuring that no symptomatic regions are missed. • Specificity: Specificity, or the true negative rate, evaluates how well the model identifies negative regions, which are areas of the image that are free from symptoms. The formula (15) is: Specificity is important for minimising false positives, ensuring that the model does not incorrectly classify healthy regions as symptomatic. Both Sensitivity and Specificity work together to provide a comprehensive evaluation, ensuring that the model not only identifies symptomatic regions but also avoids misclassifying healthy areas. For all these metrics, a score closer to 1 indicates better performance, signifying near-perfect alignment between the predicted segmentation and the ground truth mask. These metrics will be used to compare the performance of U-Net with other models, providing a detailed analysis of how well the models generalise across different cases. 4.2.2 Text similarity metrics In the second part of the experiment, the models’ability to interpret images and generate accurate textual diagnoses is evaluated. This is particularly important for models like GPT-4 and GPT-4o, which generate text-based interpretations of the symptoms seen in the images. A suite of metrics commonly used in natural language processing (NLP) is employed to compare the generated text with the reference diagnosis. • BERTScore: BERTScore is a sophisticated metric that uses BERT embeddings to evaluate the semantic similarity between the generated text and the ground truth. It measures the cosine similarity between the tokens of the predicted and reference sentences (16): where TTBB and RRBB represent the precision and recall of the BERT embeddings, respectively. BERTScore captures subtle semantic similarities that traditional n-gram-based metrics may overlook. The formulas for Precision (17) and Recall (18) are: Where C represents the set of tokens in the candidate sentence and R the set of tokens in the reference sentence, the term cosρρSSee (SS), ρρSSee (SS) denotes the cosine similarity between the embeddings of tokens t and t’. The F1 Score (FFBB) is computed as follows (19): • CLIPScore: Leveraging the multimodal CLIP model, CLIPScore measures the similarity between the vector representations of the generated text and the reference text. It is calculated using the cosine similarity between the two vectors (20): Where VVccccnncciiccccttcc and VVccrrccrrccnncccc represent the vector embeddings of the candidate and reference texts. This metric is particularly effective for evaluating descriptions in the context of images, making it ideal for diagnosing medical conditions from image input. • BLEU: This metric evaluates the overlap of n-grams (sequences of words) between the candidate and reference texts, penalising shorter sentences to ensure that full descriptions are considered. The formula is (21): BLEU is a popular metric for evaluating machine translation and text generation models. • METEOR: This metric computes the harmonic mean of unigram precision and recall, leveraging linguistic resources such as stemmers and synonyms. The formula is (22): METEOR provides a more nuanced evaluation by considering word stems and paraphrasing. Leading Article 12 13 14 15 16 17 18 19 20 21 22

RkJQdWJsaXNoZXIy OTc5MDU=