MedGRU-SVC: A Hybrid ConvGRU and Support Vector Clustering Framework for Interpretable Anomaly Detection in Medical Radiographs

Medical imaging has become one of the most critical tools in modern healthcare, accounting for more than 70% of hospital diagnostic procedures worldwide [1]. Modalities such as chest X-rays, mammograms, and CT scans provide invaluable insights for disease detection and treatment planning. However, manual interpretation by radiologists remains prone to fatigue-induced errors and subjectivity, particularly when anomalies are subtle or rare. Based on research, initial reading of chest radiographs can miss as many as 30% of actionable pulmonary nodules [2, 3]. These limitations illustrate the critical need for advanced, automated diagnostic technologies that can enhance clinical decision-making's accuracy, reliability, and consistency.

One of the distinctive challenges of medical imaging is the capture of temporal dynamics—e.g., lesion evolution in CT sequences or infiltrate formation over serial chest X-rays. Temporal information such as this is crucial for the early detection of disease patterns, such as fibrosis and developing pneumonia. Temporal modeling tools like LSTMs and GRUs have enhanced progression prediction tasks by 6–10% on F1-score [4]. But traditional recurrent networks tend to flat spatial data and hence yield suboptimal performance for radiology tasks. Convolutional Gated Recurrent Unit (ConvGRU) overcomes this by modeling spatial and temporal dependencies simultaneously [5], which makes ConvGRU a good candidate for dynamic anomaly detection.

Although they have their benefits, ConvGRU-based approaches are often used in supervised learning pipelines that need masses of labeled data. This is an issue with medical imaging since expert annotations are limited—only 2–5% of radiology datasets are annotated for cost and time reasons [6]. In addition, purely supervised models tend to be "black boxes," providing minimal interpretability. This lack of clarity erodes clinician trust, which is critical to the adoption in healthcare practice [7]. These concerns stimulate a high need for explainable, unsupervised, or weakly-supervised methods that have the ability to significantly make use of unlabeled data while being clinically interpretable.

Support Vector Clustering (SVC), an unsupervised kernel-based method, provides mathematically sound means to detect dense regions in feature space without the need for labels. Previous work has shown that deep-SVC hybrids can obtain 4–6% higher Area Under the Curve (AUC) than fully supervised baselines with applications to real-world anomaly detection, such as rare infiltrate and nodule detection [8]. This implies that incorporating ConvGRU embeddings with SVC clustering may provide a framework that is both accurate and interpretable [9]. Radiology is confronted with an acute data challenge: labels are available on only 2–5% of images because annotating them is very expensive [10]. Furthermore, supervised deep learning models tend to be non-interpretable, restricting their adoption in clinical decision-making where black-box predictions wear away trust. This has driven the need for explainable unsupervised or weakly supervised approaches [11].

SVC is an unsupervised kernel-based method that locates dense clusters in feature space without labels and has been found to perform well on noisy or imbalanced datasets [12]. Coupled with deep temporal embeddings of ConvGRU, SVC can detect subtle pathological features like infiltrates, nodules, and masses. Previous work measures AUC gains of 4–6% compared to supervised baselines, especially in identifying rare disease patterns in chest radiographs [13]. Based on this, we introduce Medical Gated Recurrent Unit–SVC (MedGRU-SVC), a hybrid model incorporating ConvGRU for spatial-temporal feature learning and SVC for unsupervised abnormality clustering. Adaptive histogram equalization is utilized for preprocessing to improve local contrast and inhibit artifacts, which is particularly helpful for low-quality chest X-rays. Trained on 10,000 radiographs from the National Institutes of Health (NIH) ChestX-ray14 dataset, the model uses Gradient-weighted Class Activation Mapping (Grad-CAM) visual attribution to emphasize clinically useful areas of attention, including lung regions, rib fractures, and soft-tissue opacities. Interestingly, Grad-CAM heatmaps revealed more than 87% correspondence with expert annotations, consolidating the interpretability of the model [14].

Comprehensive testing illustrates the dominance of MedGRU-SVC. Against a CNN-SVM baseline, the system elevated specificity from 94.0% to 96.7%, eliminating false positives in high-throughput clinical pipelines. Recall was lifted 6.5% above SVMs with manually crafted features, a vital improvement for detecting anomalies at early stages. Performance was stable across a wide range of conditions such as cardiomegaly, effusions, and infiltrates [15]. For overcoming these issues, we suggest MedGRU-SVC, a new hybrid system that combines ConvGRU for temporal–spatial feature extraction and SVC for unsupervised anomaly clustering. The system uses adaptive histogram equalization to augment contrast in noisy radiographs and Grad-CAM visual attribution for clinical interpretability. Tests on the NIH ChestX-ray14 dataset reflect better accuracy, sensitivity, and AUC performance compared to competitive baselines like CNN-SVM and LSTM-CNN hybrids. Additionally, qualitative outcomes reflect more than 87% overlap between Grad-CAM heatmaps and human annotations, supporting model transparency in predictions.

Research gap and objectives: Although existing studies have explored ConvGRU for temporal modeling and SVC for clustering, their integration in a unified framework for unsupervised medical anomaly detection remains underexplored. Current solutions either (i) rely heavily on labeled datasets, (ii) underutilize temporal dynamics, or (iii) fail to provide interpretability. This gap motivates the design of MedGRU-SVC as a scalable and clinically viable solution.

Contributions of this study: The proposed MedGRU-SVC is a novel hybrid framework that combines ConvGRU-based spatiotemporal feature extraction with SVC for unsupervised anomaly detection in medical imaging. The model attains state-of-the-art performance on NIH ChestX-ray14, outperforming competitive baselines in terms of accuracy, sensitivity, and AUC while minimizing false positives. Increased clinical interpretability through Grad-CAM visualizations, providing clear explanations to ensure transparent predictions that concur with expert annotations and facilitate clinician trust.

3.1 MedGRU-SVC model implementation

The MedGRU-SVC is a hybrid pipeline for medical radiograph anomaly detection, combining deep spatial-temporal feature learning with unsupervised clustering. It starts with input image sequences like chest X-rays or CT slices, which are processed with adaptive histogram equalization for enhanced local contrast and background noise suppression, improving the visibility of disease-related features.

Preprocessed image sequences are input to a ConvGRU network in which convolutional layers learn spatial features and recurrent gates learn temporal dependencies between frames. This enables the model to detect progression-based patterns, e.g., infiltrates or fractures that might develop over time. The ConvGRU hidden state at the last time step is reduced to a fixed-length embedding vector via global average pooling, capturing both spatial and temporal features of potential anomalies.

These embeddings are then processed by an SVC module with a Radial Basis Function (RBF) kernel. The SVC identifies natural cluster structures in the high-dimensional feature space by forming a minimum enclosing hypersphere, optimized using Bayesian hyperparameter tuning. This unsupervised clustering approach makes the pipeline suitable for datasets with limited or weak labels.

To ensure clinical interpretability, Grad-CAM is applied to the ConvGRU layers, producing heatmaps that highlight the regions contributing most to model decisions. The modular design of MedGRU-SVC allows the integration of high-performance anomaly detection with transparent visual explanations, making it a robust and interpretable solution for real-world clinical decision support systems, as illustrated in Figure 1.

Figure 1. Process flow of MedGRU-SVC model implementation

Each of the frames in the sequence is passed through the ConvGRU at time steps t = 1, 2, ..., T, in which the hidden state updates encode disease context over time. We global average pool over the last hidden state to produce a fixed-length embedding vector which captures both temporal patterns and spatial signatures of potential anomalies. The suit has been labeled by the current module with the embedding and it is used as an input to the SVC module. RBF kernel in SVC maps in a high-dimensional space, where it fits a smallest (in a Euclidean sense) container hyper-sphere. Data points that belong to similar manifolds in this representation space would be clustered into a cluster and can be treated as undesired anomalous types. The clustering process is unsupervised, so the model is suitable for unlabeled or weakly labeled.

Let I $\in$ ℝ^H×W×C×T denote a sequence of T medical images (e.g., X-rays or CT slices), where each image has height H, width W, and C channels. Adaptive Histogram Equalization (AHE) is applied to enhance contrast:

${I}'=AHE\left( I \right)$                   (1)

This step improves visibility of structural patterns by adjusting local contrast.

Each frame I'_t  at time step t is passed through a Convolutional GRU. The ConvGRU updates its hidden state h_t as:

${{z}_{t}}=\sigma \text{(}{{W}_{z}}*{{{I}'}_{t}}+{{U}_{z}}*{{h}_{t-1}}+{{b}_{z}})$                                (2)

${{r}_{t}}=\sigma ({{W}_{r}}*{{{I}'}_{t}}+{{U}_{r}}*{{h}_{t-1}}+{{b}_{r}})$                           (3)

 ${{\tilde{h}}_{t}}=\tanh \left( W\text{*}I_{t}^{'}+U\text{*}\left( {{r}_{t}}\odot {{h}_{\left\{ t-1 \right\}}} \right)+b \right)$                 (4)

${{h}_{t}}=\left( 1-{{z}_{t}} \right)\odot {{h}_{\left\{ t-1 \right\}}}+{{z}_{t}}\odot {{\tilde{h}}_{t}}$                        (5)

where, σ is the sigmoid function, $\odot $ denotes element-wise multiplication, and * represents convolution operations.

Embedding generation:

After the final time step T, the hidden state h_T is used as the embedding vector e $\in$ℝ^d representing the sequence:

$e=Flatten\left( {{h}_{T}} \right)$                        (6)

SVC:

The embedding vector e is mapped to a high-dimensional space using an RBF kernel:

$K\left( {{e}_{i}},~{{e}_{j}} \right)=\exp \left( -\gamma {{\left| \left| {{e}_{i}}-{{e}_{j}} \right| \right|}^{2}} \right)$                    (7)

SVC finds a minimal enclosing sphere in this feature space, then maps back to identify cluster boundaries in input space.

3.2 Optimization and classification

Bayesian optimization is used to tune the RBF kernel parameter γ and the sphere radius ν. After clustering, each embedding e is assigned a cluster label corresponding to different anomaly types (e.g., nodules, infiltrates).

To visualize the regions contributing most to a prediction, Grad-CAM is applied on ConvGRU outputs. This produces a localization heatmap L $\in$ ℝ^{H×W}:

$L=ReLU\left( {{\Sigma }_{k}}{{\alpha }_{k}}{{A}^{k}} \right)$                      (8)

where, A^k is the activation map of the k-th channel, and α_k is its importance weight computed via gradients. In the proposed MedGRU-SVC framework, the input at each time step is denoted by X_t, representing an image or feature map in the sequence. The hidden state of the ConvGRU network at time t is represented by h-th, which captures the spatial-temporal features from the input sequence. The temporal weighting factor γ controls the influence of previous hidden states on the current state, enabling the network to balance past and present information. The attention map at layer k is denoted as A^k, highlighting the spatial regions that contribute most to the anomaly detection score. The network parameters are optimized using a learning rate ν, while σ represents the activation function used in the ConvGRU, such as ReLU or sigmoid.

To promote clinical interpretability, Grad-CAM is applied to the ConvGRU’s intermediate feature maps, producing heatmaps that highlight the localized regions contributing to the clustering decision. Although Grad-CAM is conventionally applied to supervised CNNs to visualize class-discriminative regions, in our unsupervised anomaly detection pipeline, it is adapted to highlight regions that contribute most to the anomaly score. Specifically, after processing input sequences through the ConvGRU network, Grad-CAM generates spatial attention maps using gradients of the unsupervised loss with respect to feature maps. This provides a visual interpretation of which spatial-temporal regions the model considers anomalous, facilitating explainability without requiring labeled data. This not only aids radiologists in verifying predictions but also provides visual transparency into the model’s decision-making process. The entire architecture is implemented using Python with PyTorch for the deep learning modules and Scikit-learn for the SVC component. Training is performed on GPUs for accelerated computation, and the framework is modularized to enable future expansion, such as real-time deployment on hospital PACS systems or integration with multi-modal medical data (e.g., patient history, lab reports). Overall, the MedGRU-SVC pipeline delivers a high-performing, explainable, and clinically viable solution for automated anomaly detection in medical radiography data. The kernel parameters γ and penalty term CCC are automatically tuned using Bayesian optimization, improving the stability and compactness of clusters.

Table 1 summarizes the key hyperparameters used in the MedGRU-SVC pipeline, including learning rate, batch size, and convolutional kernel specifications. These parameters were selected to optimize training stability and model performance.

Table 1. Hyperparameters