Self-supervised learning (SSL) is an approach to pretrain models with unlabeled datasets and extract useful feature representations such that these models can be easily fine-tuned for various downstream tasks. Self-pretraining applies SSL on curated task-specific datasets without using task-specific labels. Increasing availability of public data repositories has now made it possible to utilize diverse and large, task unrelated datasets to pretrain models in the ”wild” using SSL. However, the benefit of such wild-pretraining over self-pretraining has not been studied in the context of medical image analysis. Hence, we analyzed transformers (Swin and ViT) and a convolutional neural network created using wild- and self-pretraining trained to segment lung tumors from 3D-computed tomography (CT) scans in terms of: (a) accuracy, (b) fine-tuning epoch efficiency, and (c) robustness to image acquisition differences (contrast versus non-contrast, slice thickness, and image reconstruction kernels). We also studied feature reuse using centered kernel alignment (CKA) with the Swin networks. Our analysis with two independent testing (public N = 139; internal N = 196) datasets showed that wild-pretrained Swin models significantly outperformed self-pretrained Swin for the various imaging acquisitions. Fine-tuning epoch efficiency was higher for both wild-pretrained Swin and ViT models compared to their self-pretrained counterparts. Feature reuse close to the final encoder layers was lower than in the early layers for wild-pretrained models irrespective of the pretext tasks used in SSL. Models and code will be made available through GitHub upon manuscript acceptance.

Computer-aided modeling and design (CAM/CAD) of patient anatomy from computed tomography (CT) imaging and 3D printing technology enable the creation of tangible, patient-specific anatomic models that can be used for surgical guidance. These models have been associated with better patient outcomes; however, a lack of CT imaging guidelines risks the capture of unsuitable imaging for patient-specific modeling. This study aims to investigate how CT image pixel size (X-Y) and slice thickness (Z) impact the accuracy of mandibular models. Six cadaver heads were CT scanned at varying slice thicknesses and pixel sizes and turned into CAD models of the mandible for each scan. The cadaveric mandibles were then dissected and surface scanned, producing a CAD model of the true anatomy to be used as the gold standard for digital comparison. The root mean square (RMS) value of these comparisons, and the percentage of points that deviated from the true cadaveric anatomy by over 2.00 mm were used to evaluate accuracy. Two-way ANOVA and Tukey-Kramer post-hoc tests were used to determine significant differences in accuracy. Two-way ANOVA demonstrated significant difference in RMS for slice thickness but not pixel size while post-hoc testing showed a significant difference in pixel size only between pixels of 0.32 mm and 1.32 mm. For slice thickness, post-hoc testing revealed significantly smaller RMS values for scans with slice thicknesses of 0.67 mm, 1.25 mm, and 3.00 mm compared to those with a slice thickness of 5.00 mm. No significant differences were found between 0.67 mm, 1.25 mm, and 3.00 mm slice thicknesses. Results for the percentage of points deviating from cadaveric anatomy greater than 2.00 mm agreed with those for RMS except when comparing pixel sizes of 0.75 mm and 0.818 mm against 1.32 mm in post-hoc testing, which showed a significant difference as well. This study suggests that slice thickness has a more significant impact on 3D model accuracy than pixel size, providing objective validation for guidelines favoring rigorous standards for slice thickness while recommending isotropic voxels. Additionally, our results indicate that CT scans up to 3.00 mm in slice thickness may provide an adequate 3D model for facial bony anatomy, such as the mandible, depending on the clinical indication.

Previous studies suggests that cone beam computerized tomography (CBCT) images could provide reliable information regarding the fate of bone grafts in the maxillofacial region, but no systematic information regarding the standardization of CBCT settings and properties is available, i.e., there is a lack of information on how the images were generated, exported, and analyzed when bone grafts were evaluated. The aim of this study was to (1) do a systematic review on which type of CBCT-based DICOM images have been used for the evaluation of the fate of bone grafts in humans and (2) use a software suggested in the literature to test DICOM-based data sets, exemplifying the effect of variation in selected parameters (windowing/contrast control, plane definition, slice thickness, and number of measured slices) on the final image characteristics. The results from review identified three publications that used CBCT to evaluate maxillofacial bone grafts in humans, and in which the methodology/results comprised at least one of the expected outcomes (image acquisition protocol, image reconstruction, and image generation information). The experimental shows how the influence of information that was missing in the retrieved papers, can influence the reproducibility and the validity of image measurements. Although the use of CBCT-based images for the evaluation of bone grafts in humans has become more common, this does not reflect on a better standardization of the developed studies. Parameters regarding image acquisition and reconstruction, while important, are not addressed in the proper way in the literature, compromising the reproducibility and scientific impact of the studies.

High-dose-rate (HDR) Magnetic Resonance (MR) guided brachytherapy (BT) is rapidly becoming the standard for treatment of locally advanced cervical cancer, globally. MR is an integral aspect of this treatment, enabling the level of soft tissue visualization required for precise delineation of organ and target contours with respect to the BT applicator or needles during treatment planning. The optimal slice thickness for MR datasets, and the role of super-resolved datasets are questions yet to be investigated. A digital phantom-based study assessed the impact of slice thickness on volumetric and geometric uncertainties in traditional MR datasets and estimated the resultant dosimetric uncertainty. Datasets with traditional slice thicknesses produced uncertainties up to 27% of the imaged structure volume, and contour uncertainty up to one third of the slice thickness This resulted in the exceeding of the American Association of Physicists in Medicine’s (AAPM) recommended dosimetric uncertainty in HDR BT. Trilinearly interpolated datasets reduced these uncertainties substantially, allowing imaging with 2.7 mm coarser slices while conferring an imaging time reduction of 6 minutes. The results of this thesis demonstrate that the recommended range of slice thicknesses introduces uncertainties on a level known to impact dosimetry more than 9%. Trilinearly interpolated datasets may thus confer benefit in this clinical setting.

High-dose-rate (HDR) Magnetic Resonance (MR) guided brachytherapy (BT) is rapidly becoming the standard for treatment of locally advanced cervical cancer, globally. MR is an integral aspect of this treatment, enabling the level of soft tissue visualization required for precise delineation of organ and target contours with respect to the BT applicator or needles during treatment planning. The optimal slice thickness for MR datasets, and the role of super-resolved datasets are questions yet to be investigated. A digital phantom-based study assessed the impact of slice thickness on volumetric and geometric uncertainties in traditional MR datasets and estimated the resultant dosimetric uncertainty. Datasets with traditional slice thicknesses produced uncertainties up to 27% of the imaged structure volume, and contour uncertainty up to one third of the slice thickness This resulted in the exceeding of the American Association of Physicists in Medicine’s (AAPM) recommended dosimetric uncertainty in HDR BT. Trilinearly interpolated datasets reduced these uncertainties substantially, allowing imaging with 2.7 mm coarser slices while conferring an imaging time reduction of 6 minutes. The results of this thesis demonstrate that the recommended range of slice thicknesses introduces uncertainties on a level known to impact dosimetry more than 9%. Trilinearly interpolated datasets may thus confer benefit in this clinical setting.