Use of superpixels for improvement of inter-rater and intra-rater reliability during annotation of medical images.

In the context of automatic medical image segmentation based on statistical learning, raters' variability of ground truth segmentations in training datasets is a widely recognized issue. Indeed, the reference information is provided by experts but bias due to their knowledge may affect the quality of the ground truth data, thus hindering creation of robust and reliable datasets employed in segmentation, classification or detection tasks. In such a framework, automatic medical image segmentation would significantly benefit from utilizing some form of presegmentation during training data preparation process, which could lower the impact of experts' knowledge and reduce time-consuming labeling efforts. The present manuscript proposes a superpixels-driven procedure for annotating medical images. Three different superpixeling methods with two different number of superpixels were evaluated on three different medical segmentation tasks and compared with manual annotations. Within the superpixels-based annotation procedure medical experts interactively select superpixels of interest, apply manual corrections, when necessary, and then the accuracy of the annotations, the time needed to prepare them, and the number of manual corrections are assessed. In this study, it is proven that the proposed procedure reduces inter- and intra-rater variability leading to more reliable annotations datasets which, in turn, may be beneficial for the development of more robust classification or segmentation models. In addition, the proposed approach reduces time needed to prepare the annotations.

Fully automated 3D body composition analysis and its association with overall survival in head and neck squamous cell carcinoma patients.

Objectives: We developed a method for a fully automated deep-learning segmentation of tissues to investigate if 3D body composition measurements are significant for survival of Head and Neck Squamous Cell Carcinoma (HNSCC) patients. Methods: 3D segmentation of tissues including spine, spine muscles, abdominal muscles, subcutaneous adipose tissue (SAT), visceral adipose tissue (VAT), and internal organs within volumetric region limited by L1 and L5 levels was accomplished using deep convolutional segmentation architecture - U-net implemented in a nnUnet framework. It was trained on separate dataset of 560 single-channel CT slices and used for 3D segmentation of pre-radiotherapy (Pre-RT) and post-radiotherapy (Post-RT) whole body PET/CT or abdominal CT scans of 215 HNSCC patients. Percentages of tissues were used for overall survival analysis using Cox proportional hazard (PH) model. Results: Our deep learning model successfully segmented all mentioned tissues with Dice's coefficient exceeding 0.95. The 3D measurements including difference between Pre-RT and post-RT abdomen and spine muscles percentage, difference between Pre-RT and post-RT VAT percentage and sum of Pre-RT abdomen and spine muscles percentage together with BMI and Cancer Site were selected and significant at the level of 5% for the overall survival. Aside from Cancer Site, the lowest hazard ratio (HR) value (HR, 0.7527; 95% CI, 0.6487-0.8735; p = 0.000183) was observed for the difference between Pre-RT and post-RT abdomen and spine muscles percentage. Conclusion: Fully automated 3D quantitative measurements of body composition are significant for overall survival in Head and Neck Squamous Cell Carcinoma patients.

Benchmarking of Deep Architectures for Segmentation of Medical Images.

In recent years, there were many suggestions regarding modifications of the well-known U-Net architecture in order to improve its performance. The central motivation of this work is to provide a fair comparison of U-Net and its five extensions using identical conditions to disentangle the influence of model architecture, model training, and parameter settings on the performance of a trained model. For this purpose each of these six segmentation architectures is trained on the same nine data sets. The data sets are selected to cover various imaging modalities (X-rays, computed tomography, magnetic resonance imaging), single- and multi-class segmentation problems, and single- and multi-modal inputs. During the training, it is ensured that the data preprocessing, data set split into training, validation, and testing subsets, optimizer, learning rate change strategy, architecture depth, loss function, supervision and inference are exactly the same for all the architectures compared. Performance is evaluated in terms of Dice coefficient, surface Dice coefficient, average surface distance, Hausdorff distance, training, and prediction time. The main contribution of this experimental study is demonstrating that the architecture variants do not improve the quality of inference related to the basic U-Net architecture while resource demand rises.