Umeå University's logo

Collecting large amounts of radiotherapy (RT) data from clinical systems is known to be a challenging task. Still, data collections outside the original RT systems are needed to follow-up on the quality of cancer care and to improve RT. This paper aims to describe how RT data is collected nationally in Denmark and Sweden for this purpose and gives an overview of the stored information in both countries' national data sources.

Although both countries have clinical national quality registries with broad coverage and completeness for many cancer diagnoses, some were initiated already in the seventies, and less than one in ten includes quantitative information on RT to a level of detail useful for more than basic descriptive statistics. Detailed RT data can, however, be found in Denmark's DICOM Collaboration (DcmCollab) database, initiated in 2009 and in Sweden's quality registry for RT launched in 2023 (SKvaRT). Denmark has collected raw DICOM data for all patients enrolled in clinical trials, with files being directly and automatically transferred to DcmCollab from the original data sources at each RT centre. Sweden collects aggregated RT data into SKvaRT for all patients undergoing RT in Sweden, with DICOM files being transferred and selected alpha-numeric variables forwarded via a local intermediate storage database (MIQA) at each hospital. In designing their respective solutions, both countries have faced similar challenges regarding which RT variables to collect and how to technically link clinical systems to their data repositories. General lessons about how flexibility currently is balanced with storage requirements and data standards are presented here together with future plans to harvest real-world RT data.

The world of Machine Learning (ML) offers great promise, but often there is a noticeable gap between claims made in research papers and the model's practical performance in real-life applications. This gap can often be attributed to systematic errors and pitfalls that occur during the development phase of ML models. This study aims to systematically identify these errors. For this, we break down the ML process into four main stages: data handling, model design, model evaluation, and reporting. Across these stages, we have identified fourteen common pitfalls based on a comprehensive review of around 60 papers discussing either broad challenges or specific pitfalls within ML pipeline. Moreover, Using the Brain Tumor Segmentation (BraTS) dataset, we perform three experiments to illustrate the impacts of these pitfalls, providing examples of how they can skew results and affect outcomes. In addition, we also perform a review to study the frequency of unclear reporting regarding these pitfalls in ML research. The goal of this review was to assess whether authors have adequately addressed these pitfalls in their reports. For this, we review 126 randomly chosen papers on image segmentation from the ICCV (2013-2021) and MICCAI (2013-2022) conferences from the last ten years. The results from this review show a notable oversight of these issues, with many of the papers lacking clarity on how the pitfalls are handled. This highlights an important gap in current reporting practices within the ML community. The code for the experiments is available at https://github.com/SG-Azar/BraTS-ML-Pitfalls-Experiments.

Background: Accurate diagnosis and staging are essential for optimal treatment planning of prostate cancer. By combining functional and anatomical imaging, PSMA-PET/mpMRI offers a potential to improve lesion detection and enhance staging accuracy. This study aimed to evaluate the diagnostic performance of lesion detection and local staging of prostate cancer using combined PSMA-PET/mpMRI compared to standalone mpMRI or PSMA-PET.

Results: Fifty-five patients with intermediate- to high-risk prostate cancer scheduled for robot-assisted laparoscopic radical prostatectomy were included. All patients underwent [68Ga]PSMA-PET/mpMRI prior to surgery. Whole-mount histopathology and surgical report served as reference standard. Two radiologists independently evaluated mpMRI, while two nuclear medicine physicians assessed PSMA-PET. For the PSMA-PET/mpMRI analysis, a consensus evaluation was performed by a new set of readers in two teams, each comprising one radiologist and one nuclear medicine physician. Lesion localization was reported based on the PI-RADS v2.1 sector map and compared to histopathology. Among 130 histopathologically confirmed lesions, mean detection rates were 38% (49.5/130) for PSMA-PET/mpMRI, 32% (41/130) for mpMRI and 32% (41/130) for PSMA-PET. For clinically significant prostate cancer (csPC) (≥0.5 ml, ≥ISUP 2; 42 lesions), mean detection rates were 85% (35.5/42) for PSMA-PET/mpMRI, 75% (31.5/42) for mpMRI and 70% (29.5/42) for PSMA-PET. The mean false discovery rates were 8% (PSMA-PET/mpMRI), 15% (mpMRI) and 12% (PSMA-PET). The likelihood of extraprostatic extension (EPE) and seminal vesicle invasion (SVI) were scored using a 5-point Likert scale, where scores of 1–3 were classified as negative and scores of 4–5 were considered positive. Sensitivity for EPE was 32% for PSMA-PET/mpMRI, 37% for mpMRI and 7% for PSMA-PET, with a specificity of 100%, 96% and 98%, respectively. For SVI, sensitivity was 50% for PSMA-PET/mpMRI and 38% for mpMRI and PSMA-PET, with a specificity of 100%, 95% and 97% respectively.

Conclusions: PSMA-PET/mpMRI provided higher and a more consistent performance in localized prostate cancer detection and staging without increasing false-positive findings.

DeepSMOTE is an oversampling method that combines autoencoders and the Synthetic Minority Over-sampling Technique (SMOTE) in the autoencoder's latent space to address class imbalances. A key component is a penalty term intended to increase sample diversity, but its formulation and impact are not well understood. We examined two versions: the one described in the original SMOTE paper and the one in the official SMOTE code. We formally analyze both and show that the paper version, contrary to its goal, makes reconstructions within each class more similar, thus reducing diversity. The implemented version instead increases the frequency of class sampling, implicitly rebalancing class contributions to the loss. Building on this analysis, we propose a simple refinement that better matches the intended purpose. Experiments on MNIST, FMNIST, CIFAR-10, and SVHN validate our findings. Code is available at: https://github.com/SG-Azar/DeepSMOTE-penalty.

We present an automated radiation oncology treatment planning pipeline that operates between segmentation and plan review, minimizing manual interaction and reliance on traditional planning systems. Two AI models work in sequence: the first generates a dose distribution, and the second creates a deliverable DICOM-RT plan. Trained and validated on 276 plans, and tested on 151 datasets, the system produced clinically deliverable plans—complete with all VMAT parameters—in about 38 s. These plans met target coverage and most organ-at-risk constraints. This proof-of-concept demonstrates the feasibility of generating high-quality, deliverable DICOM plans within seconds.

BACKGROUND AND PURPOSE: The study aims to evaluate dosimetric properties of hypofractionated treatment plans integrating focal boost, using registered whole-mount histopathology (WMHP) as reference standard.

METHODS: Fifteen men from the PAMP trial (EudraCT: 2015-005046-55) were included. Participants had ≥ 1 ISUP Grade group ≥ 4 lesion and underwent [⁶⁸Ga]prostate-specific membrane antigen (PSMA) positron emission tomography/multiparametric magnetic resonance imaging (PET/mpMRI) and [¹¹C]Acetate-PET/computed tomography before radical prostatectomy. Four radiation oncologists delineated gross tumor volumes (GTVs) on PSMA-PET/mpMRI. Sixty treatment plans were optimized, one per GTV and patient. Prostate planning target volumes were prescribed 42.7 Gy in seven fractions, with a simultaneous GTV boost up to 49.0 Gy, prioritizing organs at risk (OARs). Digital WMHP provided Gleason grading and was co-registered with in-vivo imaging. Target coverage for GTVs and voxels sharing Gleason patterns (GPs) was assessed via dose-volume histogram (DVH) analysis. Interobserver agreement in GTV-delineations was quantified with Fleiss' kappa.

RESULTS: The median GTV dose per plan (D50) ranged from 48.3 to 49.1 Gy. For voxels with the highest GP, D₅₀ was 42.9-49.2 Gy, exceeding 47.2 Gy in all except one plan. In lowest pattern voxels, D₅₀ was 42.5-49.3 Gy, and below 43.4 Gy in over half the plans. Significant positive correlations between Fleiss' kappa and DVH parameters appeared only for GP 5 regions, specifically for Fleiss' kappa and D₅₀ for two observers and the average D₅₀ across observers.

INTERPRETATION: The histologically confirmed tumor was only partially boosted. Regions with more aggressive disease received better coverage. These findings provide a rational for prioritizing OARs in treatment planning.

BACKGROUND: The delineation of intraprostatic lesions is vital for correct delivery of focal radiotherapy boost in patients with prostate cancer (PC). Errors in the delineation could translate into reduced tumour control and potentially increase the side effects. The purpose of this study is to compare PET-based delineation methods with histopathology.

MATERIALS AND METHODS: The study population consisted of 15 patients with confirmed high-risk PC intended for prostatectomy. [68Ga]-PSMA-PET/MR was performed prior to surgery. Prostate lesions identified in histopathology were transferred to the in vivo [68Ga]-PSMA-PET/MR coordinate system. Four radiation oncologists manually delineated intraprostatic lesions based on PET data. Various semi-automatic segmentation methods were employed, including absolute and relative thresholds, adaptive threshold, and multi-level Otsu threshold.

RESULTS: The gross tumour volumes (GTVs) delineated by the oncologists showed a moderate level of interobserver agreement with Dice similarity coefficient (DSC) of 0.68. In comparison with histopathology, manual delineations exhibited the highest median DSC and the lowest false discovery rate (FDR) among all approaches. Among semi-automatic approaches, GTVs generated using standardized uptake value (SUV) thresholds above 4 (SUV > 4) demonstrated the highest median DSC (0.41), with 0.51 median lesion coverage ratio, FDR of 0.66 and the 95th percentile of the Hausdorff distance (HD95%) of 8.22 mm.

INTERPRETATION: Manual delineations showed a moderate level of interobserver agreement. Compared to histopathology, manual delineations and SUV > 4 exhibited the highest DSC and the lowest HD95% values. The methods that resulted in a high lesion coverage were associated with a large overestimation of the size of the lesions.

Background and purpose: Dose escalation in external radiotherapy of prostate cancer shows promising results in terms of biochemical disease-free survival. Boost volume delineation guidelines are sparse which may cause high interobserver variability. The aim of this research was to characterize gross tumor volume (GTV) delineations based on multiparametric magnetic resonance imaging (mpMRI) and prostate specific membrane antigen-positron emission tomography (PSMA-PET) in relation to histopathology-validated Gleason grade 4 and 5 regions.

Material and methods: The study participants were examined with [68Ga]PSMA-PET/mpMRI prior to radical prostatectomy. Four radiation oncologists delineated GTVs in 15 study participants, on four different image types; T2-weighted (T2w), diffusion weighted imaging (DWI), dynamic contrast enhanced (DCE) and PSMA-PET scans separately. The simultaneous truth and performance level estimation (STAPLE) algorithm was used to generate combined GTVs. GTVs were subsequently compared to histopathology. We analysed how Dice similarity coefficient (DSC) and lesion coverage are affected by using single versus multiple image types as well as by adding a clinical target volume (CTV) margin.

Results: Median DSC (STAPLE) for different GTVs varied between 0.33 and 0.52. GTVPSMA-PET/mpMRI generated the highest median lesion coverage at 0.66. Combining different image types achieved similar lesion coverage as adding a CTV margin to contours from a single image type, while reducing non-malignant tissue inclusion within the target volume.

Conclusion: The combined use of mpMRI or PSMA-PET/mpMRI shows promise, achieving higher DSC and lesion coverage while minimizing non-malignant tissue inclusion, in comparison to the use of a single image type with an added CTV margin.

Manual segmentations are considered the gold standard for ground truth in machine learning applications. Such tasks are tedious and time-consuming, albeit necessary to train reliable models. In this work, we present a dataset with expert segmentations of the prostatic zones and urethra for 200 randomly selected patients from the PROSTATEx dataset. Notably, independent duplicate segmentations were performed for 40 patients, providing inter-reader variability data. This results in a total of 240 segmentations. This dataset can be used to train machine learning models or serve as an external test set for evaluating models trained on private data, thereby addressing a current gap in the field. The delineated structures and terminology adhere to the latest Prostate Imaging Reporting and Data Systems v2.1 guidelines, ensuring consistency.

The use of synthetic CT (sCT) in the radiotherapy workflow would reduce costs and scan time while removing the uncertainties around working with both MR and CT modalities. The performance of deep learning (DL) solutions for sCT generation is steadily increasing, however most proposed methods were trained and validated on private datasets of a single contrast from a single scanner. Such solutions might not perform equally well on other datasets, limiting their general usability and therefore value. Additionally, functional evaluations of sCTs such as dosimetric comparisons with CT-based dose calculations better show the impact of the methods, but the evaluations are more labor intensive than pixel-wise metrics.

To improve the generalization of an sCT model, we propose to incorporate a pre-trained DL model to pre-process the input MR images by generating artificial proton density, T1 and T2 maps (i.e. contrast-independent quantitative maps), which are then used for sCT generation. Using a dataset of only T2w MR images, the robustness towards input MR contrasts of this approach is compared to a model that was trained using the MR images directly. We evaluate the generated sCTs using pixel-wise metrics and calculating mean radiological depths, as an approximation of the mean delivered dose. On T2w images acquired with the same settings as the training dataset, there was no significant difference between the performance of the models. However, when evaluated on T1w images, and a wide range of other contrasts and scanners from both public and private datasets, our approach outperforms the baseline model. Using a dataset of T2w MR images, our proposed model implements synthetic quantitative maps to generate sCT images, improving the generalization towards other contrasts. Our code and trained models are publicly available.