An Intensity-Based Cropping Approach for Fast, Interpretable, and Robust Localization of the Knee Joint in...

Effective image preprocessing is critical for ensuring the robustness and generalizability of downstream models by preventing shortcut learning on spurious features. Knee joint localization is essential for reliable pathology assessment by automatically isolating clinically meaningful joint anatomy. Accurate selection of region of interest improves the training quality and robustness of deep learning (DL) models, making it a key step for reliable diagnostics and ensuring consistent and stable performance across different clinical tasks and settings. We introduce a fully automated, intensity-based cropping algorithm and compare its performance with established SVM- and DL-based methods. A deterministic, intensity-based algorithm using anatomical landmarks without requiring annotations or training was developed. Cropping performance was evaluated on OAI and MRKR datasets. Localization accuracy was measured using intersection over union (IoU), Dice, and mAP@0.5 scores. Computational efficiency was assessed under single-threaded conditions. Clinical utility was tested by training ConvNeXt models for Kellgren-Lawrence Grade prediction and compartment-specific osteoarthritis classification on crops from each method, with cross-testing to assess generalizability. Distributional consistency was analyzed using ResNet-18 embeddings and t-SNE clustering. The DL-based method achieved the highest localization accuracy (IoU, 0.737; mAP@0.5, 0.947; Dice, 0.842), followed by the intensity-based approach (IoU, 0.692; mAP@0.5, 0.882; Dice, 0.809). The intensity-based method was markedly faster (0.047 s/image) than SVM (0.131 s) and DL (6.217 s) and successfully processed 100% of OAI images. Models trained on intensity-based crops achieved high sensitivity and stable performance across test conditions, with minimal drop on cross-method evaluation. Visual analyses confirmed that intensity-based crops yielded more homogeneous and anatomically consistent distributions with reduced inclusion of artifacts and spurious features. The proposed intensity-based algorithm offers a robust, annotation-free alternative to supervised pipelines. Its deterministic design and minimal computational footprint support its use in large-scale research and potential deployment in clinical settings with limited data availability and access to advanced AI tools.