MRI Plane Orientation Detection using a Context-Aware 2.5D Model

TL;DR

This work tackles the lack of MRI plane orientation metadata by introducing a 2.5D context-aware classifier that leverages multi-slice information to infer axial, coronal, and sagittal planes, improving accuracy over 2D baselines. Trained on BRISC and IXI datasets, the model achieves up to 99.99% accuracy on IXI and 99.74% on average, demonstrating strong cross-domain generalization. The authors validate metadata usefulness in a brain tumor detection task using a gated strategy based on predictive entropy, boosting test accuracy from 97.0% to 98.0% and reducing misdiagnoses by 33.3%. The system is deployed as an open-source, client-side web application (TensorFlow.js) enabling private, real-time plane inference, with plan to extend the approach to additional metadata modalities.

Abstract

Humans can easily identify anatomical planes (axial, coronal, and sagittal) on a 2D MRI slice, but automated systems struggle with this task. Missing plane orientation metadata can complicate analysis, increase domain shift when merging heterogeneous datasets, and reduce accuracy of diagnostic classifiers. This study develops a classifier that accurately generates plane orientation metadata. We adopt a 2.5D context-aware model that leverages multi-slice information to avoid ambiguity from isolated slices and enable robust feature learning. We train the 2.5D model on both 3D slice sequences and static 2D images. While our 2D reference model achieves 98.74% accuracy, our 2.5D method raises this to 99.49%, reducing errors by 60%, highlighting the importance of 2.5D context. We validate the utility of our generated metadata in a brain tumor detection task. A gated strategy selectively uses metadata-enhanced predictions based on uncertainty scores, boosting accuracy from 97.0% with an image-only model to 98.0%, reducing misdiagnoses by 33.3%. We integrate our plane orientation model into an interactive web application and provide it open-source.

Figures (5)

• Figure 1: System overview of the 2.5D context-aware plane detection pipeline. (1) Heterogeneous data from 2D-native and 3D-native sources are aggregated and (2) passed through a data cleaning pipeline. (3) The 2.5D model learns a unified strategy, processing 2D images with a static context and 3D images with a random context. (4) The final trained classifier generates the missing plane orientation metadata for any input slice.
• Figure 2: Comparing 2D vs. 2.5D errors on the BRISC test set brisc. The 2.5D model (right) corrects most of the 2D model's top misclassifications (left).
• Figure 3: Gated model accuracy plotted against the uncertainty threshold on the validation set.
• Figure 4: Qualitative analysis of tumor detection using Grad-CAM gradcam. The first column shows the location of the tumor (red) if one is present.
• Figure 5: Tumor Detection performance comparison. The Gated Metadata model achieves the highest accuracy (98.0%) and the fewest misdiagnoses (20), representing a 33.3% reduction in errors compared to the Image-Only model.