Image analysis by machine learning methods on microtomography data

The use of deep learning models in the field of image segmentation saves experts time and money in many areas, especially in medicine, and therefore its continuous development is of great importance. With the increase of data collected in the medical field today, the expectations of doctors and researchers for accurate diagnostic applications developed with deep learning models are increasing. For this reason, innovative applications in the field of health need to be developed and disseminated by overcoming various difficulties. This thesis study consists of two stages. In the first stage, in addition to the original data set provided, PH2, BOWL 2018 and CVC-ClinicDB medical data sets, which are open to everyone, were used for a comprehensive analysis of deep learning models. These freely accessible datasets were preferred due to their changing image content information and different difficulty levels in segmentation tasks. U-Net, one of the deep learning models, has inspired many studies due to its success in this field. Therefore, in the first stage of the thesis study; U-Net, U-Net++, Attention U-Net, Residual U-Net, Residual Attention U-Net, TransUNet, Swin-Unet deep learning models were used and the performances obtained with the modified Self Attention U-Net and the proposed MMIAU-Net (Mixed Modified Inception and Self Attention U-Net) architecture were compared. At this stage, when all experimental results of the study were examined; it was seen that in the segmentation process, the proposed MMIAU-Net architecture achieved the highest success with 90.18% IoU and 94.83% dice coefficient in the PH2 data set, 83.23% IoU and 90.82% dice coefficient in the BOWL 2018 data set, 86.05% IoU and 92.41% dice coefficient in the CVC-ClinicDB data set and 90.51% IoU and 94.68% dice coefficient in the original data set. In the second stage of the thesis study, analyses were performed on the original data set with the proposed MMIAU-Net model. In order to demonstrate the damage of phthalates to bone health, the MMIAU-Net model was trained on micro tomography images, then the weights were recorded and the bone masses of the control group, DBP group and DEHP group were calculated. The obtained results highlight the damage rates of phthalates on bone mass.