An Explainable Plane-Wise ConvNet Approach for Detecting Femoral Head Osteonecrosis from Magnetic Resonance Images Authors: Şükrü Demir, Mehmet Vural, Buğra Can, Fatih Demir, Abdulkadir Sengur Published: 2026 ABSTRACT Background/Objectives: Osteonecrosis of the femoral head (ONFH) is difficult to diagnose, particularly in the early stages, because radiological findings may be subtle. Delayed or inaccurate staging may increase the risk of femoral head collapse and functional loss. Although magnetic resonance imaging is highly sensitive for early-stage lesion detection, interpretation may vary depending on observer experience. Therefore, reliable and explainable automated decision support approaches are needed. Methods: In this study, a deep learning-based approach was proposed to classify ONFH into early and late stages according to the Ficat–Arlet staging system. Stage I–II cases were defined as early-stage, whereas Stage III–IV cases were defined as late-stage. Axial and coronal MR images were evaluated separately to investigate plane-dependent classification performance. The images were converted into a three-channel format, resized to a common spatial resolution, normalized, and augmented during training. Feature extraction was performed using transfer learning with modern convolutional neural network architectures. ConvNeXt Tiny was used as the main classification backbone. Weighted loss was applied to reduce the effect of class imbalance, and the decision threshold was optimized on validation data to reduce missed clinically critical late-stage cases. Results: A dataset collected from the Orthopedics and Traumatology Department of Firat University Hospital was used in the experimental evaluation. The dataset was divided into training and test sets using an 80:20 split, and 10-fold cross-validation was additionally performed to assess model stability. In the hold-out test, the axial plane model achieved 94.51% accuracy, 96.80% sensitivity, 93.49% specificity, 0.9162 F1-score, and 0.981 AUC. In the coronal plane model, 92.84% accuracy, 96.13% sensitivity, 90.96% specificity, 0.9072 F1-score, and 0.988 AUC were obtained. The 10-fold cross-validation results provided a more conservative estimate of generalization performance. Conclusions: The findings indicate that deep learning-based plane-wise analysis of MR images can distinguish early- and late-stage ONFH with high performance. Grad-CAM-based visual explanations showed that the model focused mainly on clinically relevant subchondral and weight-bearing regions of the femoral head. The proposed approach may serve as an explainable decision support tool for reducing observer-dependent variability in clinical staging. Future studies should validate the method using external, multicenter datasets and paired patient-level axial–coronal images. 1. Introduction Femoral head osteonecrosis (ONFH) is a progressive disease, commonly described as avascular necrosis, resulting from decreased and impaired blood flow to the subchondral bone [1]. This ischemic process leads to osteocyte necrosis, microstructural weakening, and often loss of integrity of the spherical structure of the femoral head. This clinical course is particularly important in young and middle-aged adults because femoral head collapse leads to secondary osteoarthritis and usually results in total hip arthroplasty [2]. In cases of late diagnosis, the effectiveness of joint-preserving treatment is significantly reduced, making the process leading to arthroplasty unavoidable [2,3]. The underlying causes of ONFH are heterogeneous and multidimensional. In non-traumatic cases, high-dose corticosteroid uses or prolonged excessive alcohol consumption are the most consistent risk factors. However, metabolic, autoimmune, and hemoglobinopathies, as well as prothrombotic clinical conditions, are also associated with the disease [1,4]. The pathogenesis is multifactorial and involves vascular occlusion, increased intraosseous pressure, bone marrow fat-cell hypertrophy, endothelial injury, and impaired bone remodeling, ultimately leading to subchondral ischemia [4,5]. ONFH should be considered in the differential diagnosis of patients with function-related hip pain, particularly when relevant risk factors are present, even if radiographic findings are inconclusive [2,5]. At this stage, imaging plays a fundamental role in both diagnosis and staging of the disease. Although plain radiographs can provide diagnostically meaningful information after structural deformations have occurred, they may often appear normal in early-stage ONFH cases [2,3]. MRI is adopted as the most sensitive diagnostic method for ONFH, especially in the radiographically inconsistent stage. The ability to reveal the distribution of damaged tissue, subchondral fracture, and early collapse is vital in determining the natural clinical course of the disease [3,5,6,7]. The serpiginous lesion margin and the “double-line sign” on T2-weighted images are well-known diagnostic features [7]. In addition, the most decisive factor in the treatment decision-making process is whether the femoral head collapse has progressed. While femoral head-preserving treatment strategies can be evaluated in the pre-collapse stage, treatment methods after collapse increasingly focus on arthroplasty [2,3,8]. For this reason, the early-late distinction is not only descriptive but also the main determinant of the treatment strategy. Different staging systems have been proposed for the systematic analysis of ONFH. The Ficat–Arlet classification is widely preferred in clinical practice, and MRI findings are heavily included in this system to contribute to the detection of early-stage disease in current practice [6,8]. The ARCO staging system, revised with international consensus, offers a more structured, image-based framework and is gaining increasing acceptance [8]. However, a significant limitation of staging systems is observer dependence. Previous studies have shown only poor-to-fair reproducibility for both Ficat and ARCO staging systems, with reported inter-observer kappa values of approximately 0.31–0.39 and intra-observer kappa values of approximately 0.43–0.52, particularly in borderline or intermediate cases. This variability makes clinical standardization difficult, reduces reproducibility in research, and further strengthens the rationale for automated decision support systems [9]. Recently, deep learning approaches have played a major role in the classification of medical images [10,11]. Research on musculoskeletal imaging, in particular, is noteworthy [10,11]. In studies using multicenter datasets in the field of ONFH, models capable of performing automated staging and disease severity assessment based on magnetic resonance imaging have been reported to exhibit performance close to expert assessments [12]. Although methods for lesion detection and classification based on CNN for early-stage ONFH have been defined, it is seen that most of these approaches in the current literature do not include a standard staging process [11]. Furthermore, it is observed that holistic models that address early and late stages in a way that directly reflects the classical Ficat stages in our daily practice, interpret many image planes together, and are compatible with clinical decision support systems are limited in the literature [2,3,8,9,11,12]. The main objective of this study is to develop a system that can automatically classify ONFH cases into Ficat stages based on MR slices and make a clinically significant distinction between early (I–II) and late (III–IV) stages [2,5,6,8]. The separate evaluation of axial and coronal sections was performed to investigate plane-dependent classification performance, since lesion visibility and subchondral changes may vary across anatomical planes. The spread characteristics of lesions located in the anterosuperior weighted region and the detection of subchondral depression can show varying characteristics in different image sections [3,8]. The main goal of this approach is to reduce observer-induced variability in ONFH staging and to establish a more repeatable decision mechanism [9,11]. Early and accurate staging of ONFH lesions is crucial for timely access to joint-preserving treatment options. Clinically, the determining threshold is often the difference between the pre-collapse and post-collapse phases. Therefore, reliable detection of early-stage disease is considered a fundamental requirement that shapes not only diagnosis but also treatment. In this regard, numerous studies based on deep learning and artificial intelligence, aiming to overcome the limitations of traditional radiological evaluations, have been published in the literature in recent years. The studies reported in the literature do not exhibit a homogeneous structure in terms of imaging modality, methodological approach, and the type of clinical problem targeted. Table 1 presents a comparison of deep learning-based studies in the diagnosis of ONFH. The table summarizes the model and methods used in each study, the imaging modalities, performance percentages, and associated limitations. Developments in the literature show that MR imaging methods are far more dominant compared to other methods. The main reason for this is that MRI technology offers high sensitivity in early-stage cases, allowing for the visualization of a critical process such as subchondral insufficiency. One concrete example of this field is the study by Shen et al. [12], which was conducted with a CNN model inspired by the DenseNet architecture for a multi-class severity analysis based on the JIC system. The model, presented with multicenter cross-sectional data, reportedly achieved an accuracy of 87.8% in the expansion internal test sections. However, the retrospective nature of the study and the insufficient representation of the early-stage sample limit its reliability regarding early diagnosis. A similar situation was observed in the study reported by Klontzas et al. [13]; although high AUC and success rates were obtained with the CNN-ensemble approach from STIR sequences, the limited sample size caused a significant decrease in external test performance and raised doubts about the generalizability of the model. In the MRI-based literature, more advanced architectures focusing directly on early-stage ONFH, rather than just general staging, have also been reported. For example, Yang et al. [14] proposed a two-stage three-dimensional approach (3D-ONFHNet) for automatic staging between Ficat–Arlet stages 0–II with early ONFH detection and reported 93.83% accuracy in 381 MRI volumes from four centers. However, stage I cases were included in a limited number in this study, and the high computational load and increased memory load of three-dimensional architectures created significant limitations in terms of clinical integration. Uemura et al. [15] focused on automating the Steinberg volumetric classification, one of the indicators of collapse risk. For this purpose, a classification based on necrotic areas and femoral head segmentation was developed using Dynamic U-Net. However, the fact that the study only included pre-collapse cases and was single-center raises questions about its widespread applicability in different protocols. Segmentation and quantitative studies are also being conducted using MRI images. Ruckli et al. [17] facilitated the surgical decision-making process by segmenting healthy and necrotic bone sections with 2D-3D segmentation using the nnU-Net base. Gao et al. [16] aimed to make inferences with 3D ROI radiographic features using nnU-Net and to estimate collapse risk with LightGBM in their study and reported AUC = 0.851 in the external test. In these studies, although it is important as an infrastructure for medical decision support, the direct staging decision differs methodologically compared to other models. In addition, Vezirhüyük et al. [18] examined the consistency of different staging systems comparatively; however, the limited sample size and the lack of correlation with clinical outcomes reduce the power of the results obtained. While the literature generally reports high accuracy/AUC values on specific datasets, four common gaps stand out in terms of clinical application: (i) generalizability problems due to the small number of early-stage (especially stage I) image slices, (ii) the dominance of single-center/retrospective designs and the limitations of prospective validation, (iii) clinical integration difficulties due to computational-memory costs in 3D architectures, and (iv) the risk of inter-observer variability appearing as “tag noise” in the model in tasks open to interpretation, such as staging. Therefore, approaches that target Ficat staging and its binary distinction of early (I–II)/late (III–IV), evaluate different MRI planes separately, and aim to improve generalizability in a way that is compatible with clinical decision-making are particularly meaningful for real-world decision support. 2. Materials and Methods This study utilizes a plane-wise convolutional neural network approach for the Ficat stage classification problem. The aim of the proposed method is to classify ONFH cases into early- and late-stage groups by evaluating axial and coronal MR images separately. No feature-level fusion or patient-level multi-view integration was performed because paired axial–coronal images from the same patients were not available. The same classification pipeline was applied independently to each imaging plane, and the resulting performances were reported separately for axial and coronal images. Before model training, input MR images were converted into a three-channel format compatible with the network architectures and resized to the required input resolution. Intensity normalization was then applied to standardize the pixel-value scale and improve training stability. During training, augmentation techniques were used to improve generalization and reduce overfitting. No explicit segmentation was performed within the deep learning pipeline. In addition, no separate ROI-based cropping step was applied unless such cropping was already present in the curated dataset. Separate models were trained and evaluated for axial and coronal images using the same preprocessing and training protocol. In the feature extraction phase, feature maps obtained from axial and coronal images are represented through deep network backbones. For this purpose, high-level distinguishing features were obtained separately from axial and coronal images using ConvNeXt Tiny and EfficientNetV2 S architectures [19]. Final classification is performed with a softmax-based output layer, and probability values are generated for each class. The class with the highest probability is reported as the model’s prediction. Figure 1 provides a general overview of the proposed method. The method’s performance is evaluated using several performance metrics. Overall classification performance is reported using the accuracy metric, while a confusion matrix is used to examine error types on a class basis. The area under the ROC curve is calculated to measure the discrimination power independent of the threshold value. The area under the PR curve, summarizing the relationship between precision and sensitivity, is also used as an additional evaluation metric, considering the possibility of class imbalance. This evaluation framework aims to reveal not only the average success level of the method but also its inter-class discrimination capacity and its resilience to potential data imbalances. 2.1. Preprocessing The raw MR images were first converted into a three-channel format and resized to 224 × 224 pixels to ensure compatibility with the deep learning architectures used in this study. Intensity normalization was then applied to reduce variations in signal range across images and to provide a more standardized input for training. No explicit segmentation of the femoral head or necrotic area was performed before classification. Similarly, no additional ROI-based cropping centered on the femoral head was applied during preprocessing, unless such cropping was already present in the curated dataset. Slice selection was determined during dataset preparation, where the slice showing the clearest visualization of the femoral head and the most evident subchondral involvement was chosen for analysis; when needed, adjacent slices were also reviewed to support the staging decision [20]. 2.2. Model Architecture This study aims to automatically detect whether MR images belong to two distinct phases of clinical significance. The problem is defined as a binary classification problem addressed under a supervised learning approach. The dataset is defined as a sample set consisting of a limited number of image label pairs: (1)D={xi, yi|i=1,2,…,N} Here, xi represents a high-dimensional medical image, and yi∈{0,1} represents the corresponding clinical stage label. The main goal is to learn a generalizable function that associates images in the data field with labels in the output field. Medical images can have wide density distributions due to different devices, imaging parameters, and patient characteristics. This can lead to numerical instability in the learning process and overfitting of the model on certain samples. To reduce this effect, all images have been statistically normalized. The normalized representation for each image sample is expressed as follows [20]: (2)x~i=xi−μσ Here, μ and σ represent the mean and standard deviation values calculated over the entire dataset. This process ensures that different samples are represented in a common statistical space, contributing to a more stable optimization process. 2.3. Representation of Hierarchical Characteristics Medical images contain multi-layered information, ranging from low-level structural features (edges, density transitions) to high-level semantic structures (anatomical and pathological patterns). In this study, a nonlinear, hierarchical transformation function was used to model this multi-level information [21]. This transformation is expressed as follows: (3)hi=Φx~i;Θ Here, Φ(⋅) represents a multi-layered transformation that extracts distinctive representations from the image, and Θ represents the set of parameters optimized during the learning process. Based on this structure, the model can interpret local patterns within a global context and learn representations that strengthen the distinction between classes. The learned high-level representations are projected onto the decision space for classification purposes. This projection is expressed by a linear transformation [22]: (4)zi=Whi+b The values obtained at this stage represent the discrimination scores for each class. The magnitude of the scores indicates the proximity to the relevant class but does not yet contain a probabilistic interpretation. To make the decision scores clinically interpretable, they need to be converted into probability distributions. For this purpose, the scores are normalized to obtain class-conditional probabilities [23]: (5)P(c|xi)=expzi,c∑k=12expzi,k This definition guarantees that the sum of the probabilities of each image belonging to either class is equal to one. This makes uncertainty analysis and threshold-based decision mechanisms possible. One of the fundamental problems frequently encountered in medical datasets is the imbalance in the number of samples between classes. This can lead to biased results favoring the dominant class during the learning process. To mitigate this effect, a weighted risk function has been used: (6)L=−1N∑i=1N∑C=12ωC∥yi=clogPyi=cxi Here, ωc is the coefficient that balances the contribution of the relevant class to the learning process. This approach aims to ensure that rare classes are adequately represented by the model. Model parameters have been updated using an adaptive and moment-based optimization strategy. The general parameter update rule is expressed as follows: (7)Θt+1=Θt−ηtm^tv^t+ε+λΘt This formulation is effective in terms of both rapid convergence and preventing overlearning. The regularization term increases the generalizability of the model by preventing excessive parameter growth. Keeping the learning rate constant can lead to performance loss at different stages of the training process. Therefore, a two-stage scheduling strategy has been adopted. In the first stage, the learning rate is gradually increased: (8)ηt=ηmax.tTω, t