Juliette Moreau ^1* Laura MECHTOUFF ^1,2 David Rousseau ³ Omer Faruk EKER ^1,2 Yves Berthezene ^1,2 Tae-Hee Cho ^1,2 Carole Frindel ¹

• ¹ Université Claude Bernard Lyon 1, Lyon, France
• ² Hospices Civils de Lyon, Lyon, Rhône-Alpes, France
• ³ Université d'Angers, Angers, Pays de la Loire, France

Medical imaging plays a crucial role in stroke management, and machine learning has been increasingly used in this field, particularly for lesion segmentation. Despite advances in acquisition technologies and segmentation architectures, one of the main challenges of sub-acute stroke lesion segmentation in CT imaging is image contrast. To address this issue, we propose a method to assess the contrast quality of an image-dataset given a machine learning-trained model for segmentation. This method identifies the critical contrast level below which the model fails to learn meaningful content from images. Contrast measurement relies on the Fisher's ratio, estimating how well the stroke lesion is contrasted from the background. The critical contrast is found thanks to three methods: performance, graphical, and clustering analysis. Defining this threshold improves dataset design and accelerates training by excluding low-contrast images. Application of this method to brain lesion segmentation in CT imaging highlights a Fisher's ratio threshold value of 0.05, and training validation of a new model without these images confirms this with similar results with only 60% of the training data, resulting in an almost 30% reduction in initial training time. Moreover, the model trained without the low-contrast images performed equally well with all images when tested on another dataset.