DEEP-LEARNING DETECTION OF INTRA-ARTICULAR CALCIFICATION ON KNEE CT

Intra-articular calcification may represent a distinct radiological osteoarthritis phenotype associated with structural joint and compositional tissue changes. Advanced image analysis and deep learning approaches can extract high-dimensional imaging features that are difficult to quantify visually and may improve the detection and characterization of calcification-related osteoarthritis phenotypes.

As a first step in this broader planned work, we aimed to develop an initial deep learning model to classify knees with and without intra-articular calcifications on knee CT scans from the MOST Study.

We analyzed knee CT data from the MOST Study after image quality control. 4,014 knee CTs were read by a musculoskeletal radiologist using BUCKS in a semi-quantitative manner. Among these, 3,584 knees had no mineralization, and 430 had either cartilage or meniscal mineralization (BUCKS>0 in different meniscus and/or cartilage subregions). CT images were preprocessed by standardizing orientation, cropping to the knee region, resizing to 224 × 224 × 128 voxels, and applying min-max normalization to scale intensities to 0-1. Data were partitioned into training, validation, and hold-out test sets using stratified sampling to preserve the proportion of mineralization cases across all splits: 2,508 training samples, including 301 with mineralization; 602 validation samples, including 64 with mineralization; and 603 testing samples, including 65 with mineralization. We implemented a modified 3D ResNet-18 architecture, which has been previously used for CT images. Specifically, this consisted of 3D convolutional layers, residual blocks, global average pooling, dropout regularization with p = 0.2, and a final fully connected output layer for binary classification of intra-articular calcification. To address the severe class imbalance during training, we applied balanced class weights to the binary cross-entropy loss function, heavily penalizing misclassifications of the minority calcification class.

The modified 3D ResNet-18 demonstrated strong discrimination for IA calcification detection. Across training epochs, validation AUC increased to a peak of 0.91, with validation accuracy reaching 0.919. On the independent hold-out test set, the model achieved an PR-AUC of 0.809, AUC of 0.921 and an accuracy of 0.949, with balanced performance, as reflected in robust sensitivity and specificity for detecting calcification. These results indicate that 3D CT-based deep learning can identify imaging patterns associated with IA calcification despite class imbalance.

A modified 3D ResNet-18 model achieved promising performance for automated IA calcification detection on knee CT scans, although external validation and more granular characterization of calcification subtypes remain important next steps. This approach supports the feasibility of deep learning-based imaging biomarkers for IA calcification and provides a foundation for future work that incorporates subchondral bone segmentation, model interpretability via saliency or activation maps, separate classifiers for cartilage and meniscal calcification, and comparative CT/MRI model refinement. Ultimately, these methods may help define biologically meaningful osteoarthritis phenotypes, including insights into whether intra-articular calcification reflects a broader alteration in bone mineralization. Additionally, this work will lay the foundation for identifying relevant features on other commonly acquired imaging modalities (radiographs, MRI) to ultimately help improve risk stratification for patients with IA calcification.