Developing uncertainty methods to determine trustworthiness of segmentation masks generated from prostate MR images

Prostate cancer is the second most common type of cancer in men worldwide, accounting for about 26% of all cancer cases. The prostate gland plays an important role in the male reproductive an excretory system and prostate cancer occurs as a result of uncontrolled proliferation of cells in it. As prostate cancer can progress slowly, it can metastasize to bones and other organs by showing an aggressive spread.Therefore, early diagnosis and the right treatment methods are critical to the life span and quality of patients. Magnetic Resonance Imaging (MRI) is an imaging modality that is widely used in diagnosing prostate cancer and offers high-resolution anatomical details. MRI allows for a detailed examination of the prostate region and the identification of cancerous and healthy tissues. However, manually dividing prostate MRI images is a time consuming, tiring and high risk of errors for medical professionals. The structural differences of the prostate and uncertainties in tissue boundaries make it difficult to automatically analyze images. Therefore, there is growing interest in deep learning based automated partitioning methods in the field of medical image processing. The field of medical image processing has been significantly advanced by deep learning models in recent years, and they have often been utilized to segment prostate MRI images. Deep neural network architectures, such as U-Net, have the capacity to accurately determine the anatomical structure of prostate images. However, the shortcomings of current models in uncertainty management and their variable performance on different datasets are an important limitation for reliable use in clinical applications. Accurate measurement of uncertainties in the segmentation masks predicted by deep learning models can increase the reliability of models and enable healthier assessments in clinical decision support systems. In this study, it was aimed to estimate the division uncertainties in prostate MRI images and to examine the effects of uncertainties on out of distribution data. Using methods such as Monte Carlo Dropout (MCDO), Test-Time Data Augmentation (TTA), Ensemble, Ensemble and Orthogonality, Snapshot Ensemble and Snapshot Ensemble and Orthogonality. The durability and generalization capacity of these methods were analyzed and model performances were compared, especially with the out-of distribution data. Prediction of uncertainty in prostate MRI images has frequently been investigated in distribution data in the literature, but its effects on out-of-distribution data have not been thoroughly examined. With this study, success performances of different uncertainty prediction methods on out of distribution data were analyzed and their advantages and disadvantages were revealed in terms of model reliability. According to the study, the methods used perform differently on different datasets. While the Promise 12 and UCL datasets are generally high-performing, the results obtained on the HK dataset indicate a lower performance. This suggests that model performance is directly related to the characteristics and distribution of the data set used. In particular, Ensemble and Orthogonality and Snapshot Ensemble and Orthogonality methods have yielded more successful results in terms of uncertainty prediction compared to other methods. It has been observed that these methods are more effective in managing uncertainty and can produce more reliable results in clinical decision support systems. As a result, the effects of uncertainty prediction methods on in-distribution and out-of distribution data were examined in detail in prostate MRI images. The findings of the study suggest that uncertainty prediction could contribute to clinical processes by increasing model reliability and could help achieve more accurate results in diagnosing prostate cancer. The increase in uncertainty, especially in out-of-distribution data, is due to the data structures and limited data diversity that the model does not encounter during the training process. This suggests that uncertainty forecasting methods should be implemented correctly and that using them in clinical decision support systems can improve model reliability.