Lymphoma is one of the most common hematologic cancers worldwide, with 544,352 new cases of non-Hodgkin lymphoma and 83,087 cases of Hodgkin lymphoma reported in 2020. High-income nations see more new diagnoses, while low-income countries face higher mortality rates. Accurate staging and histological assessment are critical for determining prognosis and guiding therapy, and Positron Emission Tomography combined with Computed Tomography (PET/CT) is the standard imaging modality for lymphoma patients.
Why PET/CT matters: PET scans detect metabolic activity that flags cancerous regions, while CT provides the anatomical detail needed to localize them precisely. Together, they offer complementary metabolic and structural information that is considered indispensable for treatment planning and restaging. However, segmenting lymph nodes from PET/CT scans is notoriously difficult because lymph nodes are often surrounded by soft tissues of similar intensity, making automated boundary detection unreliable.
Deep learning over traditional ML: Classical machine learning approaches require manual feature extraction, a process that converts image data into low-dimensional feature vectors. This introduces information loss and depends on expert knowledge of feature engineering. Deep learning (DL) methods bypass this limitation by learning features directly from high-dimensional PET/CT images. The paper argues that DL consistently outperforms traditional ML on large-scale PET/CT analysis tasks because it avoids the bottleneck of hand-crafted feature design.
This study introduces an Automatic Pre-Segmentation Model (APSM) built on the Swin Transformer architecture. The model segments lymphoma regions by detecting pixel-level metabolic changes across tissues and lymph nodes. It simultaneously trains for segmentation and classification, with the lymph node region as the primary target. The claimed improvements include 12.68% higher segmentation accuracy, 13.38% improved precision, and reductions in overhead (12.73%), error (9.27%), and segmentation time (10.23%) compared to existing methods.
The paper surveys a broad range of existing methods for lymphoma segmentation. AW-SDRLSE (Li et al.) used adaptive weighting and scalable distance regularized level set evolution on PET images, validated across 90 clinical cases. HD-RDS-UNet (Wang et al.) employed a recurrent dense Siamese decoder architecture with a weighted Dice loss to leverage spatial-temporal correlation between decoder feature maps, maintaining classification effectiveness while reducing inference time.
Multi-task and adversarial approaches: Liu et al. combined lymphoma segmentation with prognosis prediction using multi-task CNNs that share learned features across both tasks. Shi et al. proposed AMC-GAN, an auxiliary U-Net branch for anatomical-metabolic consistency that learns representations of normal metabolical and anatomical patterns using co-aligned whole-body PET/CT data. Zhu et al. developed CGBS-Net, a cruciform-guided boundary-optimized segmentation network.
Evidential and weakly supervised learning: Huang et al. introduced a deep evidential network that extracts semantic feature vectors from 3D PET-CT inputs and computes belief functions at each voxel to assess uncertainty. Their later work applied weakly supervised deep learning with an Atrous Spatial Pyramid Pooling (ASPP) module to reduce dependency on precisely labeled datasets. Wang et al. proposed PSR-Nets, which use prior shift regularization and reweighting loss functions to improve sensitivity.
Remaining gaps: The authors identify that existing single-layer or dedicated network approaches rely heavily on pixel distribution. When feature variation occurs across layers, overhead increases and errors compound through layered assessments. The proposed APSM addresses this by performing segmentation and detection as independent operations on pixel distributions, reducing the complexity of handling multi-sized images and preventing cascading errors from multiple feature assessments.
The proposed APSM operates on PET/CT images by first analyzing pixel changes across the input. If no variations are detected in a region, that region is deemed unaffected and skipped, reducing unnecessary computation. Only images showing metabolic variations are processed through the Swin Transformer for classification and segmentation.
Feature extraction: The model extracts features from PET/CT images and represents pixel changes as a sequence. For each region, the system computes fractal deviations and feature changes simultaneously. The average detection time for a region is calculated based on the fractal deviation relative to the segmentation time, the feature changes observed, and the probability of encountering complex situations. When the probability of complexity at a given region is less than 1, the model proceeds with standard processing. When lymph nodes are very large, pixel change and fractal deviation analysis in complex regions takes significantly longer.
Fractal geometry for texture analysis: The model leverages multistage and multidimensional fractal geometry to classify lymphoma tissue. Fractal analysis captures self-similarity at multiple scales, which is particularly relevant for lymphoma tumors that exhibit fractal-like borders. By computing the fractal dimension at various scales, the model captures both fine and coarse structural information, including the irregularities and roughness of tumor boundaries. This fractal feature set fluctuates across the tumor area, providing a rich descriptor of tissue morphology.
Pre-segmentation pipeline: The full pipeline involves magnifying PET and CT scans, ensuring co-registration of functional and anatomical data, using thresholding to isolate high-uptake areas in PET scans, and then feeding the resulting patches into the Swin Transformer. The transformer output classifies maximum deviations and changes, which are used to identify the root causes of cancer by continuously adapting the correlation between testing and training inputs.
Variation detection: The model detects variations across regions by analyzing sequential pixel changes and feature differences. For each lymph node, the system examines size, color, and intensity based on pixel changes from PET images. A processing rate is assigned to each input, and the system computes the addition of segmentation and detection rates across all features in a given region. If pixel changes are observed at a rate of 0.07, for example, that indicates 7% of the fractal deviations observed from the instance contribute to region detection precision and processing overhead.
Pixel changes in identified regions: Once a problem location is identified through pixel changes, the model uses fractal deviations and feature changes through the Swin Transformer to adapt the correlation between training and testing data. Two diagnostic pathways are followed: first, analyzing the image to verify whether variation is occurring and guiding treatment for affected patients; second, if maximum deviation is observed in a lymph node between identified regions, flagging high overhead and processing time. The model formulation maximizes the detection rate by optimizing region features against pixel change constraints across all regions.
Swin Transformer for region detection: The Swin Transformer (Shifted Window Transformer) partitions the input image into non-overlapping local windows and computes self-attention within each window. The shifted window mechanism creates cross-window connections that improve computational efficiency while capturing global context. For lymphoma segmentation, the model uses the Swin Transformer as a backbone to extract multiscale and multidimensional features from PET/CT patches. A combination of Dice loss and cross-entropy loss handles class imbalance, and post-processing steps like erosion and dilation refine segmentation mask boundaries.
Training termination strategy: A distinctive feature of this model is that it terminates training of the matching transformer network when it identifies new deviations, alterations, or both. This prevents the system from overfitting to previously seen patterns and forces it to adapt to new pixel distributions. The probability of pixel changes is computed sequentially across lymphoma regions, ensuring continuous detection based on maximum deviation and changes observed from the instance.
The experimental analysis used a publicly available whole-body FDG-PET/CT dataset from The Cancer Imaging Archive. The dataset contains 501 input images acquired at 200 mAs reference dosage and 120 kV tube voltage, captured 60 minutes post intravenous injection. Image sizes used for analysis ranged from 1x1, 2x2, 4x4, 8x8, to 16x16 pixels, extracted from larger source images of 400x400 pixels. All processing was performed in MATLAB.
Analysis I, Variation and Regions: The first experimental stage examined the relationship between fractal deviation values and the number of training epochs. The optimal condition for deviation is when the fractal deviation falls within the [0, 1] range, which produces less variation across three epochs. High convergence occurs when the fractal deviation is below 0, and low convergence when it exceeds 1. The number of epochs and training recurrences are determined by the complexity of the feature changes observed across regions.
Analysis II, Region Detection Based on Maximum Deviation: The second stage performed region detection using combined fractal deviation and complexity probability inputs, tested across 10 iterations from the original input. Six maximum deviation values were analyzed under different conditional combinations. The regions identified represent those expressing maximum deviation, which are then distinguished from precision regions.
Analysis III, Precision Region Detection: The third stage validated conditions from Analysis II by examining training, validation, and testing outputs across 3 epochs for 6 identified regions. Unlike the conditional analysis of prior stages, this stage cumulatively identified the best solution and convergence. The precision regions exhibited distinct changes in lymphoma detection compared to the deviation-observed regions from earlier stages.
The proposed APSM-ST was benchmarked against three existing methods: AW-SDRLSE, DFE+ES, and PSR-Nets. Five metrics were evaluated across five region sizes (1x1, 2x2, 4x4, 8x8, and 16x16). Features were varied from 1 to 14 across 5 regions.
Segmentation accuracy: APSM-ST achieved the highest accuracy at 93.05% for the 1x1 region, compared to 86.5% for PSR-Nets, 76.3% for DFE+ES, and 72.5% for AW-SDRLSE. Across all region sizes, APSM-ST consistently led, with accuracy ranging from 86.09% (2x2) to 93.05% (1x1). The improvement over existing methods was 12.68% for varying features and 13.75% for varying sizes.
Precision: The model reached a maximum precision of 0.9678 at the 16x16 region size, compared to 0.926 for PSR-Nets, 0.802 for DFE+ES, and 0.762 for AW-SDRLSE. Precision gains were 13.38% for features and 14.81% for sizes. Overhead: APSM-ST reduced computational overhead to as low as 0.211 seconds (2x2 region), compared to 0.33 seconds for PSR-Nets, 0.41 for DFE+ES, and 0.58 for AW-SDRLSE. Overhead reductions were 12.73% and 11.27% for features and sizes respectively.
Error and segmentation time: Error rates dropped to 0.0588 at the 1x1 region (vs. 0.0656 for PSR-Nets, 0.081 for DFE+ES, 0.11 for AW-SDRLSE), representing reductions of 9.27% for features and 11.05% for sizes. Segmentation time fell to 0.82 seconds for the 2x2 region and 1.217 seconds for the 1x1 region, compared to 2.1-2.8 seconds for PSR-Nets at the same sizes. Time reductions were 10.23% and 11.14% for features and sizes.
The confusion matrix analysis revealed that APSM-ST achieved the highest true positive count (780) and true negative count (1,100) among all tested models, with the fewest misclassifications: only 20 false positives and 45 false negatives. This indicates strong balanced detection across both object and background classes, with neither class disproportionately affected by errors.
Comparison to other models: PSR-Nets, DFE+ES, and AW-SDRLSE all produced higher counts of false positives and false negatives. The gap in true positive and true negative counts highlights that APSM-ST is not just marginally better on aggregate metrics but fundamentally more consistent in its classification decisions. The low false negative count (45) is particularly important in a clinical context, where missed lymphoma regions could delay diagnosis and treatment.
The comparative performance table confirmed that APSM-ST consistently achieved superior results across all metrics and all region sizes. The model reached its best numbers at specific scales: peak accuracy (93.05%) at 1x1, peak precision (0.9678) at 16x16, lowest overhead (0.211s) at 2x2, lowest error (0.0588) at 1x1, and fastest segmentation (0.82s) at 2x2. These findings support the claim that the model is suitable for both accurate and efficient medical image segmentation in real-time applications with high reliability and minimal resource utilization.
Single-dataset evaluation: All experiments were conducted on a single publicly available FDG-PET/CT dataset with 501 images. While this dataset is well-established in the research community, the absence of multi-dataset or multi-center validation limits confidence in how the model would generalize to clinical data from different scanners, imaging protocols, or patient populations. Real-world lymphoma PET/CT data can vary considerably in resolution, noise levels, and reconstruction algorithms.
Lymph node size variability: The authors acknowledge that initial segmentation procedures, including classification and detection, are influenced by differences in lymph node diameters. Variations in region widths create distribution differences in pixel sequences, which can introduce errors. This is a fundamental challenge because lymph nodes in real patients range from a few millimeters to several centimeters, and the model's performance may degrade at extreme size ranges not well-represented in the training data.
Proposed solution, regressive segmentation: To address the size-dependency issue, the authors propose a future method called regressive segmentation. This approach would analyze pixel distribution independently according to region size, handling each scale separately rather than applying a uniform analysis. The goal is to tackle pre-segmentation limitations and improve diagnostic accuracy across the full range of lymph node sizes encountered in clinical PET/CT imaging.
Broader considerations: The paper does not address prospective clinical validation, integration into radiology workflows, or how the model handles edge cases such as co-occurring pathologies or treatment-altered anatomy. The computational requirements for real-time deployment are also not discussed in detail, though the reduced segmentation time (as low as 0.82 seconds) suggests feasibility for clinical settings. Future work would benefit from multi-center trials and comparison against radiologist-level segmentation performance on matched datasets.