Umeå University's logo

Background and purpose: Accurate segmentation of the urethra is crucial for safe focal dose escalated radiotherapy, while prostate zone identification is important for prostate cancer diagnosis. Manual delineations on magnetic resonance imaging (MRI) are labour-intensive and variable, and while deep learning offers promise in automating this process, no available solution currently exists. This study aimed to develop and evaluate a deep learning model for automatic segmentation of the urethra, prostate and all prostate zones and benchmark its performance against inter-reader variability and assess generalisability to external data from a different MRI vendor.

Materials and methods: The public datasets ProstateZones and PROSTATEx included 200 magnetic resonance images with manual delineations, with 160 used for training/validation and 40 with independent duplicate segmentations used as a test set. A nnU-Net deep learning model was evaluated on the unseen test set and externally validated on a dataset with 55 samples. Performance was assessed using Dice Similarity Coefficient (DSC), Surface DSC, percentile Symmetric Surface Distance, and Center Line Distance (CLD) metrics.

Results: The model outperformed the inter-reader variability on multiple structures, and notably on all metrics for the urethra, with median CLD values of 2.8 and 2.9 mm compared to 3.6 mm for inter-reader variability. External validation showed robust generalisability to a dataset collected from a different vendor.

Conclusions: This study demonstrated that a deep learning model can achieve expert-level performance in automated segmentation of the urethra, prostate, and prostate zones. Robust performance on external data highlighted potential as a decision support solution.

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.

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.

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.

Purpose: The increasing importance of medical imaging in cancer treatment, combined with the growing popularity of deep learning gave relevance to the presented contributions to deep learning solutions with applications in medical imaging.

Relevance: The projects aim to improve the efficiency of MRI for automated tasks related to radiotherapy, building on recent advancements in the field of deep learning.

Approach: Our implementations are built on recently developed deep learning methodologies, while introducing novel approaches in the main aspects of deep learning, with regards to physics-informed augmentations and network architectures, and implicit loss functions. To make future comparisons easier, we often evaluated our methods on public datasets, and made all solutions publicly available.

Results: The results of the collected projects include the development of robust models for MRI bias field correction, artefact removal, contrast transfer and sCT generation. Furthermore, the projects stress the importance of reproducibility in deep learning research and offer guidelines for creating transparent and usable code repositories.

Conclusions: Our results collectively build the position of deep learning in the field of medical imaging. The projects offer solutions that are both novel and aim to be highly applicable, while emphasizing generalization towards a wide variety of data and the transparency of the results.

PURPOSE: During the acquisition of MRI data, patient-, sequence-, or hardware-related factors can introduce artefacts that degrade image quality. Four of the most significant tasks for improving MRI image quality have been bias field correction, super-resolution, motion-, and noise correction. Machine learning has achieved outstanding results in improving MR image quality for these tasks individually, yet multi-task methods are rarely explored.

METHODS: In this study, we developed a model to simultaneously correct for all four aforementioned artefacts using multi-task learning. Two different datasets were collected, one consisting of brain scans while the other pelvic scans, which were used to train separate models, implementing their corresponding artefact augmentations. Additionally, we explored a novel loss function that does not only aim to reconstruct the individual pixel values, but also the image gradients, to produce sharper, more realistic results. The difference between the evaluated methods was tested for significance using a Friedman test of equivalence followed by a Nemenyi post-hoc test.

RESULTS: Our proposed model generally outperformed other commonly-used correction methods for individual artefacts, consistently achieving equal or superior results in at least one of the evaluation metrics. For images with multiple simultaneous artefacts, we show that the performance of using a combination of models, trained to correct individual artefacts depends heavily on the order that they were applied. This is not an issue for our proposed multi-task model. The model trained using our novel convolutional loss function always outperformed the model trained with a mean squared error loss, when evaluated using Visual Information Fidelity, a quality metric connected to perceptual quality.

CONCLUSION: We trained two models for multi-task MRI artefact correction of brain, and pelvic scans. We used a novel loss function that significantly improves the image quality of the outputs over using mean squared error. The approach performs well on real world data, and it provides insight into which artefacts it detects and corrects for. Our proposed model and source code were made publicly available.

Concerns about the reproducibility of deep learning research are more prominent than ever, with no clear solution in sight. The Medical Imaging with Deep Learning (MIDL) conference has made advancements in employing empirical rigor with regards to reproducibility by advocating open access, and recently also recommending authors to make their code public—both aspects being adopted by the majority of the conference submissions. We have evaluated all accepted full paper submissions to MIDL between 2018 and 2022 using established, but adjusted guidelines addressing the reproducibility and quality of the public repositories. The evaluations show that publishing repositories and using public datasets are becoming more popular, which helps traceability, but the quality of the repositories shows room for improvement in every aspect. Merely 22% of all submissions contain a repository that was deemed repeatable using our evaluations. From the commonly encountered issues during the evaluations, we propose a set of guidelines for machine learning-related research for medical imaging applications, adjusted specifically for future submissions to MIDL. We presented our results to future MIDL authors who were eager to continue an open discussion on the topic of code reproducibility.

In magnetic resonance imaging (MRI), bias fields are difficult to correct since they are inherently unknown. They cause intra-volume intensity inhomogeneities which limit the performance of subsequent automatic medical imaging tasks, \eg, tissue-based segmentation. Since the ground truth is unavailable, training a supervised machine learning solution requires approximating the bias fields, which limits the resulting method. We introduce implicit training which sidesteps the inherent lack of data and allows the training of machine learning solutions without ground truth. We describe how training a model implicitly for bias field correction allows using non-medical data for training, achieving a highly generalized model. The implicit approach was compared to a more traditional training based on medical data. Both models were compared to an optimized N4ITK method, with evaluations on six datasets. The implicitly trained model improved the homogeneity of all encountered medical data, and it generalized better for a range of anatomies, than the model trained traditionally. The model achieves a significant speed-up over an optimized N4ITK method—by a factor of 100, and after training, it also requires no parameters to tune. For tasks such as bias field correction - where ground truth is generally not available, but the characteristics of the corruption are known - implicit training promises to be a fruitful alternative for highly generalized solutions.

 The contrast settings to select before acquiring magnetic resonance imaging (MRI) signal depend heavily on the subsequent tasks. As each contrast highlights different tissues, automated segmentation tools for example might be optimized for a certain contrast. While for radiotherapy, multiple scans of the same region with different contrasts can achieve a better accuracy for delineating tumours and organs at risk. Unfortunately, the optimal contrast for the subsequent automated methods might not be known during the time of signal acquisition, and performing multiple scans with different contrasts increases the total examination time and registering the sequences introduces extra work and potential errors. Building on the recent achievements of deep learning in medical applications, the presented work describes a novel approach for transferring any contrast to any other. The novel model architecture incorporates the signal equation for spin echo sequences, and hence the model inherently learns the unknown quantitative maps for proton density, 𝑇1 and 𝑇2 relaxation times (𝑃𝐷, 𝑇1 and 𝑇2, respectively). This grants the model the ability to retrospectively reconstruct spin echo sequences by changing the contrast settings Echo and Repetition Time (𝑇𝐸 and 𝑇𝑅, respectively). The model learns to identify the contrast of pelvic MR images, therefore no paired data of the same anatomy from different contrasts is required for training. This means that the experiments are easily reproducible with other contrasts or other patient anatomies. Despite the contrast of the input image, the model achieves accurate results for reconstructing signal with contrasts available for evaluation. For the same anatomy, the quantitative maps are consistent for a range of contrasts of input images. Realized in practice, the proposed method would greatly simplify the modern radiotherapy pipeline. The trained model is made public together with a tool for testing the model on example images.