Deep learning works well for the classification of medical images, but the reason why it has not received much acceptance in the medical industry lies in the opacity and fragility of the model when any kind of input alterations occur. Here, a method has been proposed to measure the amount and graphically represent the displacement of the embeddings in convolutional neural networks that have undergone transformations named Blur, Canny, and Sobel. Results obtained by using the concepts of t-SNE, PCA, and case-by-case analysis have identified the relationship between the displacement of the embeddings and the error margin when making any kind of medical diagnoses, like the breast cancer versus Acute Lymphoblastic Leukemia dataset. The proposed method enables transparent diagnostic decision-making while incorporating transformation-aware training. Consequently, it facilitates thedevelopment of trustworthy and robust medical systems, reducing potential risks associated with the integration of deep learning into clinical workflows.