An Anatomy-Aware Deep Learning Framework for Level-Wise and Side-Specific Classification of Lumbar Foraminal and Lateral Recess Stenosis on Magnetic Resonance Imaging

This section provides an overview of the data resources used and the methodology adopted to develop a DL-based system for three-class severity classification of lumbar spine stenosis.

2.1 Dataset descriptions

This study utilized a publicly available dataset of lumbar spine MRI scans obtained from the RSNA 2024 Lumbar Spine Degenerative Classification Challenge [25]. All cases were further reviewed by an experienced radiologist to ensure the reliability of the final dataset.

The MRI data were used in their original DICOM format, including both sagittal and axial T2-weighted sequences. Maintaining the data in DICOM format allowed preservation of important spatial and geometric information, which is essential for accurate localization. This approach reduces the risk of information loss and processing errors associated with early image format conversion.

The dataset from the RSNA 2024 Lumbar Spine Degenerative Classification Challenge had several issues. Some cases contained inconsistent labels, incomplete metadata, or missing imaging sequences, requiring careful filtering and validation before use.

In total, the dataset comprised 1,789 patients, from which 15,870 disc-level samples were derived. Each patient contributed five lumbar intervertebral disc levels (L1–L2 through L5–S1), resulting in 3,174 samples per level. Each sample corresponded to a single disc level annotated with a severity grade. The distribution of severity labels included 6,515 mild cases (41.0%), 6,850 moderate cases (43.2%), and 2,505 severe cases (15.8%). For each disc-level sample, left and right ROIs were additionally extracted to represent the foraminal and lateral recess regions.

The corresponding severity distributions for the four region-specific datasets (LF, RF, left subarticular/lateral recess (LS), and right subarticular/lateral recess (RS)) are presented in Table 1. These region-specific distributions reflect the side-specific annotations provided in the original RSNA dataset and were considered during model training.

To address the imbalance in severity distribution, first, samples were constructed at both the disc level and the anatomical side, increasing the representation of different severity categories. Also, during training, balanced mini-batch sampling was employed to reduce bias toward the dominant class. In addition, data augmentation techniques were applied to enhance variability and improve the generalization of the model.

Table 1. Distribution of severity labels across the four region-specific datasets

2.2 Data preparation

Figure 2 illustrates the overall workflow used for data preparation. In the first stage, lumbar spine MRI data in DICOM format are preprocessed to extract geometric information, correctly order the slices, and identify individual disc levels. A confidence estimation step is then applied to verify the reliability of the detected anatomical levels. The extracted metadata are then used to determine the spatial position and orientation of each slice, ensuring proper alignment between sagittal and axial MRI sequences.

In the second stage, the identified disc levels are mapped across sagittal and axial views. ROIs are then extracted around each disc and represented as 2.5D multi-slice stacks centered on the target disc level. Finally, these samples are organized into structured, level-wise datasets that are ready for model training and classification.

232323.png

Figure 2. The overall workflow of the data preparation phase

2.3 Methodology

A two-stage pipeline for automated three-class severity classification of lumbar foraminal and lateral recess stenosis from DICOM-format MRI images is proposed in this work.

In the first stage, intervertebral disc levels are identified using DICOM spatial metadata, followed by the extraction of side-specific ROIs corresponding to the foraminal and lateral recess zones. In the second stage, a 2.5D multi-slice representation is constructed and used to train a DL classifier for severity classification into three categories: mild, moderate, and severe. The overall workflow of the proposed work is shown in Figure 3.

33.png

Figure 3. The overall workflow of the proposed work

2.3.1 First stage

The first stage consists of four sequential steps: disc-level localization on sagittal T2-weighted images, sagittal-to-axial spatial mapping, confidence estimation, and side-specific ROI extraction. The procedure was implemented using image-intensity information together with the spatial geometry available in the original DICOM files.

(A). Disc-level localization

Sagittal T2-weighted images were used to identify the five lumbar intervertebral disc levels (L1–L2, L2–L3, L3–L4, L4–L5, and L5–S1). When multiple sagittal T2-weighted slices were available, the slice used for localization was selected automatically according to an image-quality score defined as the sum of the normalized image intensity standard deviation and 0.5 times the mean absolute image-gradient magnitude. The sagittal image was then normalized to the range [0,1] and smoothed using a Gaussian filter with a standard deviation of 1.2 pixels.

To identify candidate disc levels, the analysis was restricted to the central 30% of the image width (approximately 35–65%), where the lumbar spinal column is expected to be located. For each image row, the brightest 15% of pixels within this central band were selected, and their mean intensity was calculated, producing a one-dimensional vertical intensity profile. The profile was subsequently smoothed using a 15-pixel moving-average filter. Only the vertical region between approximately 25% and 95% of the image height was considered for peak detection.

(B). Sagittal-axial slice mapping

The detected disc locations were converted from sagittal image pixel coordinates to patient-space coordinates using the DICOM Image Position Patient (IPP), Image Orientation Patient (IOP), and Pixel Spacing parameters. For a detected pixel position ((x,y)), the corresponding three-dimensional patient-space coordinate (P) was calculated as:

$P=I P P+(x-1) {Col} \Delta x+(y-1) R o w \Delta y$

where, Row and Col are the row and column direction-cosine vectors obtained from IOP, and (Δx) and (Δy) are the corresponding pixel-spacing values.

For the axial T2-weighted series, the normal vector of the axial imaging plane was calculated from the cross product of the row and column direction-cosine vectors:

$n_{{axial}}=\frac{{Row} * {Col}}{\| {Row} * {Col} \|}$

For each detected disc-level coordinate P_disc, the perpendicular spatial distance to every axial slice was then calculated as:

$D_i=\left|\left(I P P_i-P_{{disc}}\right) \cdot n_{{axial}}\right|$

where, IPP_i denotes the IPP of the i-th axial slice. The axial slice having the minimum distance d_iwas selected as the center slice corresponding to the detected disc level. Five consecutive axial slices were subsequently selected around this center slice, consisting of the center slice and two neighboring slices on either side. This procedure establishes the spatial correspondence between the sagittal localization and the axial classification images without relying on image filename ordering.

(C). Confidence estimation

A localization confidence score was calculated to quantify the reliability of the automatically detected five disc levels. The score combines the strength of the detected intensity peaks with the consistency of the vertical spacing between consecutive detected levels.

The peak-strength component was calculated as:

$S_{\text {peak}}=\frac{\text {mean}\left(R_{\text {loc}}\right)}{\text { mean}\left(R_{\text {search}}\right)}$

where, R_loc represents the row-intensity scores at the five selected locations and R_search represents the scores throughout the permitted search region.

The spacing-consistency component was calculated as:

$S_{\text {spacing }}=1-\frac{\operatorname{std}(\Delta r)}{\operatorname{mean}(\Delta r)+\varepsilon}$

where, Δr denotes the vertical distances between consecutive detected levels. The final confidence score was constrained to the range [0,1] by clamping each component before combining them with equal weights:

$C=0.5~ clamp_{[0,1]}\left(S_{\text {peak}}\right)+0.5 ~clamp_{[0,1]}\left(S_{\text {spacing}}\right)$

A minimum confidence threshold of 0.55 was used for quality-control flagging. Cases with (C < 0.55) were marked as low-confidence and recorded in the processing log for subsequent quality assessment. DICOM series for which the required spatial geometry could not be read, or for which an adequate number of axial slices was unavailable, were not processed for level-wise extraction.

(D). ROI extraction

Following axial slice selection, side-specific ROIs were extracted separately for the left and right lateral recess regions. The ROI definitions were based on fixed image-relative anatomical regions rather than a separately trained segmentation model.

For foraminal stenosis, the image midpoint was used as the left–right reference. The ROI width was set to approximately 45% of the half-image width, with a vertical extent from approximately 22% to 80% of the image height. The left and right ROIs were shifted laterally by approximately 20% of the half-image width to position the crops over the corresponding foraminal regions.

For subarticular/lateral recess stenosis, the ROI width was set to approximately 70% of the half-image width, with a vertical extent from approximately 18% to 85% of the image height. The image midpoint separated the left and right ROIs.

The four resulting anatomical regions were therefore defined as: (1) LF, (2) RF, (3) LS, and (4) RS. Extracting these regions separately provides side-specific inputs and reduces the amount of irrelevant background information presented to the classification models.

2.3.2 Second stage

In the second stage of the proposed work, a 2.5D representation is constructed, and the classification begins after organizing the input data into four groups.

(A). 2.5D multi-slice representation

A 2D model processes a single slice and ignores information from neighboring slices, while a 3D model analyzes the entire volume and requires high computational resources. The proposed approach, referred to as 2.5D, uses several adjacent slices stacked together, allowing the model to capture limited depth information while still applying 2D convolutions. This step is illustrated in Figure 4.

4.png

Figure 4. Illustration of the 2.5D magnetic resonance imaging (MRI) input used for classification

For each disc level, a stack of K = 5 consecutive axial T2-weighted slices centered at the disc plane is extracted. These slices are combined to form a multi-channel input. The proposed approach uses axial slices for classification, and sagittal slices help identify the correct disc level.

(B). Dataset organization

After ROI extraction, the dataset is divided into four independent region-specific subsets corresponding to LF, RF, LS, and RS. Each sample includes:

•Disc level

•Anatomical side

•Severity label (mild, moderate, severe) from the corresponding severity annotation in the original RSNA dataset.

This structured organization allows each model to learn region-specific characteristics effectively.

(C). Deep feature extraction and classification

A CNN based on ResNet-18 was employed for severity classification. ResNet-18 was chosen because it can learn important image patterns effectively without being too complex. Compared to heavier models, ResNet-18 requires less computational power and memory, making it appropriate for our dataset size and hardware setup. Figure 5 shows the ResNet-18 model with a 2.5D input approach.

5.png

Figure 5. ResNet-18 model with 2.5D approach

To accommodate the 2.5D input:

•The first convolutional layer is modified to accept five input channels instead of three.

•The final fully connected layer is replaced with a three-class classifier.

•A softmax activation function is used to produce class probabilities.

The same network architecture is trained separately on each region-specific dataset (LF, RF, LS, RS), resulting in four specialized classifiers.

The four ResNet-18 classifiers were trained for 30 epochs using Adam, with an initial learning rate of 3 × 10⁻⁴ and a mini-batch size of 15. Balanced sampling and focal loss (γ = 1.0) were used to address class imbalance, with gradient clipping applied for training stability. Pretrained ResNet-18 weights were used, while the input and final classification layers were modified for the 5-channel input and three-class output. Training was implemented in MATLAB R2024b using dlnetwork and MAT-file streaming on a PC equipped with an 11th-generation Intel Core i5-1145G7 processor and 16 GB RAM, running Windows 11 Pro.

The dataset is split at the patient level into:

•70% training

•15% validation

•15% testing

To address class imbalance:

•Balanced mini-batch sampling was used during training to reduce bias toward the dominant class.

•Data augmentation is performed on-the-fly. Data augmentation was applied only to the training set. Small rotations, shifts, brightness changes, and horizontal flips were used to increase image variability. These transformations were applied to increase image variability while preserving the relevant anatomical and pathological characteristics of the original images. The RSNA dataset includes MRI examinations acquired using different scanner configurations, providing some degree of acquisition heterogeneity; however, scanner-related variability was not systematically analyzed in the present study.

Training is conducted using a custom MATLAB dlnetwork loop with MAT-file streaming, enabling efficient handling of large datasets without exceeding memory limitations.

Model performance is monitored using:

•Training loss

•Validation accuracy

•Validation prediction distribution

The model achieving the highest validation Macro-F1 score is selected for final evaluation.

This study proposes a DL framework for automated classification of lumbar foraminal and lateral recess stenosis using MRI images. The proposed system achieved a mean test accuracy of 80.17% and a mean Macro-F1 score of 0.802 for all four independent anatomical regions.

There is slightly better performance in the foraminal models compared with the lateral recess models, which may be due to anatomical differences. The neural foramen is more clearly defined on MRI images, while the lateral recess shows more complex and variable anatomy, making it more challenging to classify. This increased anatomical complexity in lateral recess narrowing may explain the slight reduction in performance in left and right lateral recess models.

Confusion matrices show that misclassifications mainly occurred between adjacent severity levels (mild-moderate and moderate-severe). This is clinically reasonable because stenosis severity level changes gradually rather than in clearly separated categories. The confusion matrices show that misclassifications between mild and severe stenosis were relatively infrequent compared with errors involving adjacent severity categories.

Previous studies have demonstrated the feasibility of automated lumbar stenosis assessment using DL and anatomical localization. Hallinan et al. [19] developed a two-component framework using ROI detection followed by CNN-based classification for central canal, lateral recess, and neural foraminal stenosis, with four-level severity grading and independent external testing. Bharadwaj et al. [18] subsequently used anatomical segmentation and geometric localization followed by DL classification, with an emphasis on interpretable assessment of central canal stenosis and binary classification of neural foraminal stenosis.

Table 4 summarizes the main differences between the proposed framework and previous deep-learning studies for automated lumbar spinal stenosis assessment.

Previous studies have established anatomical localization and region-specific classification for automated lumbar stenosis assessment. The proposed work builds on these approaches by integrating DICOM-based spatial correspondence, anatomy-guided side-specific ROI extraction, and a structured 2.5D representation for independent LF, RF, LS, and RS grading. Thus, the main contribution lies in the integration and task-specific formulation of these components rather than in introducing a new DL architecture.

Table 4. Comparison of the proposed framework with previous studies

Note: BiT = Big Transfer, CNN = convolutional neural network, ROI = region-of-interest, DICOM = Digital Imaging and Communications in Medicine, IPP = Image Position Patient, IOP = Image Orientation Patient, LF = left foraminal, RF = right foraminal, LS = left subarticular/lateral recess, RS = right subarticular/lateral recess.

Although direct numerical comparison is not appropriate because of differences in datasets, grading schemes, and evaluation metrics, the results demonstrate that the proposed framework can achieve consistent three-class classification performance across four anatomically distinct regions. The mean Macro-F1 of 0.802, together with the relatively narrow range of performance across the four models, suggests that the proposed spatially guided formulation provides a consistent basis for automated assessment of foraminal and lateral recess stenosis.

The present evaluation used a single patient-level 70%/15%/15% train-validation-test split. While this prevents patient-level data leakage, a single split does not assess variability across different partitions. Future work will evaluate repeated patient-level splits or cross-validation to further assess model robustness and generalizability.

The present study should be interpreted as a retrospective technical evaluation rather than a clinical validation study. Although the RSNA 2024 Lumbar Spine Degenerative Classification Challenge dataset provides a relatively large and standardized benchmark for evaluating the proposed framework, performance on a benchmark dataset does not necessarily represent performance in routine clinical practice. In particular, this study did not include external validation using independent hospital or multi-center datasets, systematic analysis of scanner- and institution-related variability, or prospective evaluation within a clinical workflow. Potential uncertainty or variability in the original severity annotations should also be acknowledged, as independent multi-reader agreement was not systematically evaluated in this study. Therefore, the reported results demonstrate the technical feasibility and consistency of the proposed approach but do not establish clinical effectiveness or readiness for deployment as a radiological decision-support system. Future work should evaluate the framework on independent datasets acquired from different institutions and MRI scanners and should include prospective clinical testing to assess robustness, generalizability, and practical utility.