6-8^th November 2024

MetricsReloaded is a collaborative scientific endeavour that recently published guidelines regarding the use of evaluation metrics in medical imaging processing tasks. We have developed a software tool to accompany these guidelines which is hosted on the MONAI repository. This includes the library of metrics but also the associated end-to-end evaluation pipelines.

For people to use the software, we need to make it easy and reliable through
1. Implementation of tests for the evaluation pipelines
2. Creation of didactic jupyter notebooks to guide the user through instance segmentation and object detection evaluation procedures
3. Improvement on the documentation to relate more closely to the guideline papers.
This is the first hackathon of 3 across the UK, Germany and France with the MetricsReloaded collaborators. Ultimately we aim to release a (relatively) stable version of the work by January 2025 and will try and submit a paper detailing specifically the software capabilities to which active members of the hackathons could be invited to contribute.

The MetricsReloaded software is the companion code to the guidelines described extensively here and the pitfalls explained here. Code can be found at https://github.com/Project-MONAI/MetricsReloaded

Adaptive Normative Modelling with Real-Time Quality Control Feedback focuses on enhancing the reliability of brain MRI-based normative models by integrating dynamic quality control (QC) feedback. The main aim is to ensure accurate results from normative models, even when working with lower quality MRI scans, by conducting QC both before and after Freesurfer parsing. This approach allows adaptive handling of issues like motion artifacts or noise in real-time, reducing the need to discard valuable data. The challenge lies in creating a robust feedback loop that adjusts how the data is processed based on QC metrics. We will use brain MRI data, leveraging software such as Freesurfer/Fastsurfer, VisualQC and PyTorch/Tensorflow for potential future model adaptations.

The diaphragm, a key respiratory muscle, plays a crucial role in lung function and can be affected by various pulmonary diseases. Accurate measurement of diaphragm height on chest CT scans can provide valuable insights into respiratory health and the diagnosis of conditions like chronic obstructive pulmonary disease (COPD) and diaphragm paralysis.

This project aims to develop an automated system for diaphragm height measurement using lung masks extracted from CT chest scans. By leveraging advanced image processing and AI techniques, the system will identify and segment the diaphragm, compute height measurements, and provide quantitative assessments. Machine learning or deep learning models can be trained to enhance the accuracy of diaphragm height estimation. The project aims to streamline the assessment of diaphragm function, reduce the need for manual measurements, and support clinicians in diagnosing respiratory conditions more efficiently.

Accurate segmentation of pulmonary airways and vessels is crucial for the diagnosis and treatment of pulmonary diseases. However, current deep learning approaches suffer from disconnectivity issues that hinder their clinical usefulness.

Topology repairing for vessels focuses on correcting discontinuities within segmented vessel structures in medical imaging. This approach addresses segmentation errors, such as disconnected vessel segments or small gaps, by reconstructing the topology of vascular networks. Through keypoint detection or graph-based techniques, the method identifies critical points or endpoints in segmented vessels and reconnects them to restore continuity. This process typically involves using a neural network to detect and predict connection points or applying algorithms to optimize connectivity across the network. Topology repair enhances the accuracy and anatomical coherence of segmented vessels, which is essential for reliable clinical analysis and diagnostics.

Blender is a free and open-source software for 3D geometry rendering. Uses include synthetic datasets generation which is of particular interest in the field of medical imaging, where often there is limited real data that can be used to train machine learning models. By creating large amounts of synthetic but realistic data, we can improve the performance of models in tasks such as polyp detection in image guided surgery. Synthetic data generation has other advantages since using tools like Blender gives us more control and we can generate a variety of ground truth data from segmentation masks to optic flow fields, which in real data would be very challenging to generate or would involve extensive time consuming manual labelling. Another advantage of this approach is that often we can easily scale up our synthetic datasets by randomising parameters of the modelled 3D geometry. There can be challenges to make the data realistic and representative of the real data.

This project would be using a previously created baseline dataset of the colon and/or liver. This synthetic data provided will need some additional modification i.e. geometry, texture, lighting, and camera settings to make it more realistic. The challenge will be to generate synthetic data that is realistic enough to be useful for training models, while also being able to generate enough data to train the models effectively. A Blender add-on (or plug-in) under development for the purpose of data generation which would be a starting point for the project (https://github.com/UCL/Blender_Randomiser) with the aim to check data quality and if time can pre-train models for polyp detection (YoloV7) and organ segmentation (UNet), with performance evaluated on open-source datasets.

Disclosure: This project is part of the BHF Data Science Centre ECG Challenge.

The aim of this challenge is to create models which can be used to identify and classify important disorders of the heart from electrocardiogram (ECG, EKG) images. The dataset includes synthetic images of ECGs which simulate artefacts which might occur when taking a photo of the ECG with a smart phone. Models to automate diagnoses from these images could be used to speed up heart disease care around the world.

This challenge will use the GenECG dataset. This dataset contains synthetically generated ECG images which have been created to incorporate artefacts common on paper ECG images which are typically scanned or photographed. The wave form data to create this dataset is part of the PTB-XL dataset containing diagnoses that have been adjudicated by two cardiologists. The GenECG dataset contains 21,799 ECG images which have passed a clinical Turing test where expert observers could not distinguish between synthetic and real ECGs.

This challenge will focus on the identification of five important heart disease diagnoses. These include myocardial infarction, atrial fibrillation, hypertrophy, conduction disturbance, and ST/T change. Further information on these diagnoses can be found in the data section.