Source-linked AI summary

GRIPNet: Gaussian Radial Intensity Prior Guided Architecture for Pulmonary Nodule Detection in CT

Yang, Haojie, Su, Ran

arXiv:2609.11312v1cs.CVcs.AI

TL;DR

Small pulmonary nodules remain difficult to detect because generic detectors ignore the Gaussian-like radial physics of their CT appearance. GRIPNet measures that prior and builds it into convolution, frequency decomposition, attention, and loss design, reaching strong mAP@0.5 across three benchmarks at real-time speed.

  • Problem

    Nodules below six millimeters remain difficult to detect because conventional methods treat them as generic objects and disregard the physical process shaping their CT appearance.

  • Method

    GRIPNet maps the Gaussian radial intensity prior to pinwheel convolutions, dual-frequency features, dilated masked attention, and adaptive loss.

  • Results

    GRIPNet reaches mAP@0.5 values of 95.3%, 91.6%, and 97.9% on KanserSet, LUNA16, and Lung-PET-CT-Dx, respectively.

  • Takeaways & Limitations

    Across 18,218 lesions, the Gaussian radial pattern is stable across datasets and size strata, supporting a physics-guided detector that operates at real-time speed.

  • Takeaways & Limitations

    The reported transfer evidence is limited to three public benchmarks and relies on the assumption that shared volumetric physics explains their cross-dataset consistency.

Abstract

from arXiv · show

Lung cancer causes more deaths than any other malignancy, and low-dose CT screening is the main pathway to early diagnosis. That pathway hinges on the smallest lesions, yet nodules below six millimeters remain hard to detect, because most methods treat a nodule as a generic object and ignore the imaging physics behind its appearance. We show that this appearance is highly regular. Intensity peaks at the geometric center of a nodule and decays radially in a Gaussian pattern, and a fit to 18,218 annotated lesions from three public benchmarks yields a mean radial coefficient of determination above 0.86 in every dataset and size stratum. A square convolution samples both axes uniformly and is mismatched to this radial signal, most severely for small nodules. Guided by this evidence, we propose GRIPNet (Gaussian Radial Intensity Prior Network), a detector in which every module maps to a measurable property of the intensity distribution. Pinwheel convolutions decompose radial gradients, a dual-frequency module separates boundary detail from structural context, dilated masked attention matches the decay extent, and an adaptive loss reweights samples by conspicuity. GRIPNet raises mAP@0.5 to 95.3, 91.6 and 97.9 percent on KanserSet, LUNA16 and Lung-PET-CT-Dx while sharpening high-IoU localization at real-time speed.

I. INTRODUCTION

Pulmonary nodules have a measurable Gaussian-like radial intensity structure that generic detectors overlook, especially for small lesions. GRIPNet converts this physical prior into architecture choices and reports strong accuracy at real-time speed.

  • Micro-nodules below 6 mm achieve only 65–75% sensitivity in routine reading, while conventional extractors also struggle with faint-margin ground-glass opacities.
  • Gaussian fitting shows peak intensity at the nodule center with radial attenuation, but square convolutions are mismatched to these gradients, particularly for small nodules.
  • GRIPNet maps measurable intensity properties to pinwheel convolutions, dual-frequency processing, dilated masked attention, and adaptive loss on a YOLOv11 backbone.
  • 89.2% mAP@0.75 is reached on KanserSet at real-time throughput, with consistent improvement over strong and lung-nodule-specific baselines.

III. GAUSSIAN RADIAL INTENSITY PRIOR

The paper tests whether pulmonary-nodule intensity follows a Gaussian radial model by fitting radial HU profiles across three public CT benchmarks. The fit is consistently strong across lesion strata, motivating prior-guided architectural design.

  • The protocol extracts concentric-annulus HU profiles around each bounding-box centroid, fits a Gaussian by nonlinear least squares, and estimates amplitude, spread, and R2.
  • Mean radial R2 exceeds 0.86 in every benchmark stratum, and more than 89% of nodules exceed R2=0.70.
  • The radial spread scales with lesion diameter, from about 3 pixels for sub-6 mm nodules to 12 pixels for nodules larger than 15 mm.
  • The measured regularity is attributed to a dense central core and partial-volume averaging that produces values declining with distance.
  • The prior motivates asymmetric operators for radial gradients and explicit frequency separation between the smooth Gaussian envelope and high-frequency boundary irregularities.

IV. PRIOR-GUIDED ARCHITECTURE

GRIPNet translates the radial prior into a YOLOv11-based architecture, using asymmetric operators to capture directional components of nodule gradients. Pinwheel convolution reconstructs the radial field with fewer parameters and a larger receptive field.

  • GRIPNet implements four prior-derived modules on a YOLOv11 backbone with progressive downsampling and feature fusion at P3–P5.
  • Pinwheel Convolution uses four directionally padded asymmetric branches to capture radial gradients that standard 3 × 3 convolution may miss.
  • The branch outputs are concatenated and refined by a 2 × 2 convolution before PWConv is used at stride-2 downsampling stages.
  • The four branches reconstruct the radial field with 22% fewer parameters and a 177% larger receptive field than a 3 × 3 convolution.

B. Dual-Frequency Feature Decomposition

DualFreqC3k2 separates smooth low-frequency structure from high-frequency nodule morphology, while the broader architecture places this decomposition within a multi-scale GRIPNet pipeline. The supplied passages identify fitting statistics and feature-routing roles rather than a comparative table outcome.

  • B. Dual-Frequency Feature Decomposition: The high-frequency path targets spiculations, lobulations, and calcifications, while the low-frequency path represents the smooth Gaussian envelope.
  • B. Dual-Frequency Feature Decomposition: Table I reports radial-fit statistics across three benchmarks, including R2, mean spread, and the fraction of nodules with R2 > 0.70.
  • B. Dual-Frequency Feature Decomposition: DualFreqC3k2 separates high-frequency morphology from low-frequency context at P4 and P5 within GRIPNet’s multi-scale architecture.

C. Dilated Masked Attention for Scale-Calibrated Fusion

C2DTAB calibrates fusion attention to the Gaussian decay scale by combining dilated channel dependencies with sparsely sampled masked-window attention.

  • C. Dilated Masked Attention for Scale-Calibrated Fusion: C2DTAB combines dilated grouped channel self-attention with masked-window self-attention that samples sparsely through even-coordinate positions.This design models inter-feature dependencies while reducing dense-attention cost.
  • C. Dilated Masked Attention for Scale-Calibrated Fusion: A 3 × 3 feed-forward window expands to 7 × 7 with dilation rate d=2, covering the 3σ range of 6 to 15 mm nodules.The module replaces C2PSA in the neck.
  • C. Dilated Masked Attention for Scale-Calibrated Fusion: The attention design aggregates peripheral context across the Gaussian envelope without incurring the quadratic cost of dense attention.

D. Scale-Adaptive Detection Heads

Scale-adaptive heads combine grouped spatial and convolutional attention with dynamically reweighted samples and centroid-aware regression to handle nodules across sizes and difficulties.

  • D. Scale-Adaptive Detection Heads: At P3, Efficient Multi-scale Attention emphasizes micro-nodule boundary detail, whereas at P5 it prioritizes context separating large masses from confluent vessels.The module splits inputs into eight groups and preserves PWConv’s orthogonal gradient components.
  • D. Scale-Adaptive Detection Heads: Efficient Multi-scale Attention splits each input into eight groups processed by parallel spatial and convolutional pathways with softmax-normalized weights.A sigmoid gate combines the pathway outputs.
  • E. Adaptive Loss with Dynamic Difficulty Tracking: EWMASlide dynamically tracks prediction IoU with threshold µt and shifts sample emphasis toward challenging examples as training progresses.The loss uses an exponentially weighted moving average rather than a fixed difficulty threshold.
  • E. Adaptive Loss with Dynamic Difficulty Tracking: The WiseIoU regression penalty uses centroid distance, matching nodules whose Gaussian peak defines a natural center; the unified objective sets λcls=1.0 and λbox=0.05.

A. Datasets and Implementation

The study evaluates GRIPNet on three public benchmarks using physics-respecting augmentation and compares it with general and lung-nodule-specific detectors. GRIPNet achieves the strongest reported benchmark performance overall, especially at stricter localization thresholds, while trading some KanserSet recall for localization precision.

  • Datasets and implementation: The evaluation uses KanserSet, LUNA16, and Lung-PET-CT-Dx, covering lesions, radiologist-annotated nodules, and four histological subtypes.Training uses PyTorch 2.1.0 on an NVIDIA RTX 4090 with SGD and cosine annealing.
  • Datasets and implementation: Augmentation preserves the radial prior through rotation, scaling, flipping, scanner-related brightness shifts, contrast scaling, and mild Gaussian noise.Mosaic augmentation is excluded because it creates anatomically impossible layouts and disrupts contextual reasoning.
  • Benchmark comparison: GRIPNet reaches the highest mAP@0.5 on all three benchmarks, scoring 95.3% on KanserSet and 97.9% on Lung-PET-CT-Dx.It also leads all six metrics on LUNA16.
  • Benchmark comparison: At mAP@0.75, GRIPNet reaches 89.2% on KanserSet while every other method remains below 79%.This is the clearest separation from the compared detectors.
  • Benchmark comparison: GRIPNet trails YOLOv8 in KanserSet recall, 89.9% versus 92.5%, while providing higher-IoU localization precision for measuring nodule extent.The comparison describes this as a trade-off between recall and localization precision.
  • Benchmark comparison: MSDet is the strongest competitor on KanserSet and Lung-PET-CT-Dx, but GRIPNet leads it by 2.2 and 1.7 points of mAP@0.5.On LUNA16, GRIPNet leads MSDet and improved YOLOv11 by 1.1 points of mAP@0.5 while achieving 87.8% recall.

C. Ablation Study

The ablation study tests GRIPNet components on KanserSet and compares the complete prior-guided model with YOLOv11 in accuracy, complexity, and throughput. The curriculum and center penalty jointly produce the strongest high-IoU localization, while the full model remains real-time.

  • Component ablation: The EWMASlide curriculum and WiseIoU center penalty together raise full-model mAP@0.75 to 89.2%, whereas omitting the curriculum reduces it to 67.8%.The ablation attributes the localization gain to their combination rather than to the center penalty alone.
  • Component ablation: PWConv produces the sharpest single ablation gain, lifting mAP@0.5 to 92.0% on KanserSet.The sequential ablation also reports 92.9% for the EMA-augmented baseline and intermediate changes from DualFreqC3k2 and C2DTAB.
  • Efficiency: The prior-guided model increases parameters from 2.58 M to 4.94 M and GFLOPs from 6.3 to 8.5 versus YOLOv11.These added costs correspond to mAP@0.5 gains of 1.9 to 7.4 points.
  • Efficiency: Throughput remains above real time at 230 to 309 FPS across the three benchmarks despite the added model complexity.Table IV reports FPS and mAP@0.5 in KanserSet / LUNA16 / Lung-PET-CT-Dx order.
  • Training behavior: Training stabilizes in about 100 epochs on KanserSet, roughly 266 epochs on LUNA16, and near epoch 200 on Lung-PET-CT-Dx.The early oscillation is attributed to the adaptive threshold shifting during optimization.

F. Qualitative Analysis

Qualitative analyses show that GRIPNet’s representations progress from radial boundary responses to peri-nodular frequency-separated context and broader anatomical neighborhoods, while detections remain tightly localized across datasets and lesion conditions.

  • The detector preserves tight boxes for faint sub-centimeter lesions, pleural or vessel-adjacent nodules, and four histological subtypes across the benchmarks.The reported behavior is attributed to WiseIoU center anchoring and dilated masked attention constrained to the Gaussian decay envelope.
  • GRIPNet’s feature maps progress from radial rim activation at P3, through a peri-nodular 3σ ring at P4, to bronchial and vascular context at P5 across all three datasets.P4 separates smooth interior structure from sharp spiculation and lobulation detail.
  • Representative detections are shown across the three benchmarks under diverse imaging conditions, complementing the feature-map progression and localization analysis.
  • The mAP@0.75 increase from 84.1% to 89.2% occurs only when EWMASlide and the center penalty act together, while neither component alone exceeds 80%.This ablation links high-IoU localization to jointly learning sample difficulty and center position.
  • GRIPNet’s 2D isotropic assumption can fail for strongly anisotropic nodules, and its KanserSet recall remains slightly below YOLOv8 despite high-IoU precision.The authors identify an ellipsoidal volumetric prior and a gentler difficulty schedule as possible next steps.

VII. CONCLUSION

The paper identifies a stable Gaussian radial intensity law in pulmonary nodules and converts it into the GRIPNet architecture. The detector achieves state-of-the-art accuracy at real-time speed across three benchmarks, with especially wide margins at high IoU.

  • GRIPNet reaches mAP@0.5 values of 95.3%, 91.6% and 97.9% on KanserSet, LUNA16 and Lung-PET-CT-Dx, respectively.The detector maps radial gradients, frequency structure, decay extent and lesion conspicuity to corresponding architectural or training components.
Loading 2609.11312v1…