Quality Control Menu

Quality Control menu, version 3.0.1 Alpha

Current version 3.0.1 Alpha Quality Control menu screenshot.

Label Selected Spectra

Starts a labelling session for the currently selected spectra. The application steps through the selected voxels, displays the corresponding spectrum, and asks the user to assign a quality label. During this process the active voxel is selected in the image viewer so the spectral label remains connected to its spatial position.

The labelling handler extracts spectral quality features for each selected voxel and stores the session information together with patient, study, and series context. These labels can later be reviewed in the labelling analyser or used to train a classifier.

Internally, the selected voxels are visited in grid order. For quality labelling, spectrIm-QMRS computes the same spectral quality feature vector that is later passed to the automatic classifier. The active voxel is highlighted while the labelling popup asks for the expert decision, so the spectrum, image position, and stored label remain tied to the same coordinate.

Label Selected Spectra - Consensus

Starts a consensus labelling workflow. This mode compares a new labelling session with a reference labelling file, making it useful when several raters annotate the same spectra or when a project needs agreement with an established reference set.

Consensus labelling is especially relevant for building robust quality-control datasets. It helps separate genuine spectral quality problems from subjective differences between raters.

Labelling Analyser

Opens the labelling session analyser. The analyser can load labelling files, list sessions, show patient/study/series information, count labelled voxels, and summarize values such as the mean label value and standard deviation per session.

This view is useful for checking whether a labelling campaign is balanced, whether raters used the same classes consistently, and whether enough labelled spectra are available before training or validating a classifier.

Labelling files use the .lbls extension. The analyser can load one or more labelling-session files, refresh the session overview, and remove selected sessions when a labelling attempt should not be kept. The table records the session file, patient, study, series, start date and time, duration, number of labelled voxels, mean label value, standard deviation, and storage location.

The analyser also reports the number of used label classes and the actual class values found in the loaded sessions. This is useful before classifier training, because imbalanced labels or unexpected class values can otherwise silently reduce the usefulness of the training set.

To be defined: add a current screenshot of the Labelling Analyser window.

Classify Spectral Quality Using Custom Classifier

Applies a user-selected quality classifier to the loaded MRSI spectra. The classifier file is selected through a file chooser and is expected to be an .RData file. The output is stored as spectral quality scores and can be mapped or used in downstream quality-based selection.

The current implementation writes the spectral-quality feature table of the selected spectra to a temporary CSV file, loads the selected R Random Forest model, and asks the model for the class-2 probability. In this context the score is interpreted as the estimated probability that a rater would accept the spectrum. Scores are stored per voxel and can be displayed as a quality map or used by selection and processing actions.

Version 3.0.1 Alpha checks several prerequisites before classification: an Rscript.exe path must be configured in Settings > Other, the frequency-domain signal-free region must be valid for the current spectrum, a valid .RData classifier file must be supplied, and the required R packages, randomForest and Runiversal, must be available in the user Documents/R library path.

In processing actions, classifier output can be combined with a quality threshold. This makes it possible to keep only spectra whose predicted quality exceeds the configured threshold, while still keeping the quality map available for review.

To be checked manually: document which validated classifier files are distributed or recommended for version 3.0.1 Alpha. Until this is defined, classifier files should be treated as project- and protocol-specific.

Scientific background

The automatic quality-control workflow is related to the work by Pedrosa de Barros, McKinley, Knecht, Wiest, and Slotboom on automatic quality control in clinical proton MRSI of brain cancer. The formal citation and related quality-management references are collected on the Citing spectrIm-QMRS page.

In Nuno Pedrosa de Barros' thesis, the Random Forest quality classifier was deliberately trained on compact spectrum-derived descriptors rather than on every individual spectral data point. This was done to capture robust information about spectral quality while reducing sensitivity to small spectral misalignment, temperature or pH related shifts, and irrelevant details of individual spectra.

The current spectrIm-QMRS feature table contains 47 features. They include maximum peak SNR and mean SNR in frequency-domain ranges covering lipid, lactate, NAA, Glx, creatine, choline, and residual-water/lipid regions; global frequency-domain SNR descriptors and ratios; TD and FD maximum, mean, standard deviation, skewness, and kurtosis; TD mean SNR in 25 ms intervals from 0 to 200 ms; and TD relative-change features in the same time intervals.

These feature families are also exposed elsewhere in the application as spectral feature maps. They can therefore be used interactively for voxel selection, and they can also serve as input variables for automatic MRS(I) quality filtering when a suitable classifier is available.