Early pulmonary nodule detection is life-saving but technically demanding. Lung cancer mortality is closely tied to disease stage at diagnosis, and detection of small pulmonary nodules on CT imaging before they progress is the primary mechanism by which screening reduces mortality. However, automated nodule detection faces a fundamental challenge: models accurate enough to reliably identify small nodules are typically slow and computationally expensive, while fast models sacrifice detection accuracy.
YOLO models are fast but struggle with small nodules and false positives. The YOLO family of one-stage object detectors converts detection into a regression problem, enabling real-time inference without a separate region proposal step. While this speed advantage is clinically valuable, YOLO models applied to lung CT images tend to miss small-sized pulmonary nodules and generate false-positive detections, which undermine clinical utility and drive unnecessary follow-up procedures.
Two-stage detectors sacrifice speed for accuracy. Two-stage models such as Faster R-CNN achieve higher detection accuracy than one-stage methods but require 80 times more parameters and operate at only 13 frames per second, making them impractical for real-time clinical deployment on standard hospital hardware. Existing improvements to YOLO for lung nodule detection either boosted accuracy while doubling model size or improved speed while reducing accuracy below clinically acceptable thresholds.
Three simultaneous improvements were needed to break the tradeoff. Prior work addressed the accuracy-speed conflict by modifying a single component, achieving only marginal gains. This study proposed combining three complementary improvements: multidimensional attention convolution for richer feature extraction, lightweight bidirectional feature pyramid fusion for multiscale detection, and an improved loss function specifically targeting small nodule detection and class imbalance.
Standard convolution applies fixed kernels that miss contextual variability. Conventional convolutional layers learn fixed kernel weights that cannot adapt to differences among input features. Dynamic convolution methods like CondConv and DyConv improved on this by learning combinations of multiple kernels, but both focused only on the kernel-number dimension while increasing the total parameter count substantially.
ODConv captures attention across all four convolution dimensions simultaneously. Omni-dimensional dynamic convolution (ODConv) extends dynamic convolution by applying independent attention weights across four dimensions: the spatial size of the convolutional kernel, the input channel count, the output channel count, and the number of kernels. These four attention weights are computed in parallel using a shared global average pooling operation followed by fully connected branches with Sigmoid activation, and are multiplied progressively to generate adaptive, input-dependent convolution kernels.
The C3_ODC module replaces standard convolutions in the YOLOv5 backbone. The standard C3 bottleneck module in YOLOv5s uses conventional CBS (convolution-batch normalization-SiLU) blocks. The proposed C3_ODC module replaces the inner convolution layers with ODConv-based OBS blocks, enabling the backbone network to capture richer contextual information between CT image slices while adapting to the diverse appearance of pulmonary nodules across patients and scanning parameters.
Ablation confirmed C3_ODC was the dominant accuracy contributor. Ablation experiments on the LUNA16 dataset showed that adding C3_ODC alone to YOLOv5s improved mAP from 84.1% to 90.9%, a gain of 6.8 percentage points. Across five alternative attention mechanisms tested under identical conditions, C3_ODC achieved the highest mAP of 92.8%, outperforming CBAM and coordinate attention by more than 3 percentage points, at the cost of only a modest increase in parameter count.
BiFPN replaces the FPN-PANet neck for bidirectional weighted feature fusion. The standard YOLOv5 neck uses a feature pyramid network with path aggregation (PANet) for multiscale feature fusion. The proposed GS-BiFPN structure replaces this with a weighted bidirectional feature pyramid network (BiFPN), which adds both top-down and bottom-up feature pathways with learnable per-feature weighting. Unlike PANet, BiFPN removes low-contribution single-input nodes and adds skip connections within the same resolution level, enabling more efficient multiscale information propagation.
GSConv replaces standard convolutions in the neck to reduce parameters and increase speed. Within the BiFPN neck, standard convolutional layers are replaced with lightweight GSConv modules. Each GSConv layer performs a standard convolution followed by a depth-wise convolution, concatenates both outputs, and applies a shuffle-blending operation that integrates channel information from both branches. This design reduces redundant parameters while preserving inter-channel connectivity, substantially increasing detection speed without degrading multiscale feature quality.
The F-NWD loss function addresses small nodule detection and class imbalance simultaneously. The standard CIoU loss function used in YOLOv5s is highly sensitive to positional deviations in small objects, degrading detection of tiny nodules. The proposed focal-normalized Wasserstein distance (F-NWD) loss combines two complementary components: Focal Loss down-weights easy negative samples and focuses gradient updates on difficult positive samples to address the overwhelming ratio of background to nodule regions, while the normalized Wasserstein distance models bounding boxes as 2D Gaussian distributions and measures similarity via the Wasserstein distance, which is less sensitive to absolute size differences than IoU-based metrics.
The optimal Focal Loss to NWD weighting ratio was determined empirically. Weight analysis experiments varied the ratio of Focal Loss to NWD contribution across multiple training runs. A ratio of 0.1 Focal Loss to 0.9 NWD achieved the lowest overall loss value at convergence, because NWD provides accurate positional deviation correction for small targets while Focal Loss handles the class imbalance through a smaller but essential contribution. This ratio was used for all subsequent training and evaluation experiments.
The proposed model achieved 92.8% mAP at 116.7 frames per second. Evaluated on the LUNA16 dataset containing 1,186 lung nodules from 888 CT scans, the combined model with all three improvements achieved a mean average precision of 92.8% and a detection speed of 116.7 frames per second, with 6.64 million parameters and 14.6 billion floating-point operations. This represented an 8.7% improvement in mAP and a 5.1% improvement in FPS over the YOLOv5s baseline, while simultaneously reducing parameters by 5.4% and FLOPs by 8.2%.
The model substantially outperformed all comparable-size YOLO variants. Among models with similar parameter counts, the proposed model outperformed YOLOv5s (84.1%), YOLOv7-Tiny (74.7%), and YOLOv9s (86.5%) by 8.7, 18.1, and 6.3 percentage points in mAP respectively. Against heavier models, the proposed model exceeded YOLOv5l (90.5%) despite using only 14% as many parameters, and outperformed YOLOv7 (88.4%) while running 2.7 times faster at 116.7 FPS versus 42.9 FPS.
Qualitative detection showed specific gains for challenging nodule types. Visual comparison of detection results confirmed that YOLOv5s generated false-positive detections in multi-nodule images that the proposed model correctly resolved. The proposed model additionally detected extremely small nodules that YOLOv5s missed entirely, which was directly attributed to the NWD component of the F-NWD loss function. For isolated nodules, near-vascular nodules, and near-lung-wall nodules, both models detected the nodules but the proposed model assigned higher confidence scores.
All three improvement components contributed independently and synergistically. Full ablation results confirmed additive contributions from each module. C3_ODC alone added 6.8 mAP points, F-NWD alone added 5.7 mAP points, and GS-BiFPN alone reduced FLOPs by 3.5G while contributing 2.4 mAP points. The full combination achieved the highest mAP of 92.8% with a well-balanced parameter and FLOPs profile, confirming that the three components addressed complementary aspects of the detection problem without interfering with each other.
LUNA16 is the world's largest publicly available lung nodule benchmark. The LUNA16 dataset is a subset of the LIDC-IDRI collection, containing 1,186 annotated lung nodules in 888 lung CT scans. Nodule coordinates and diameters are recorded in a standardized annotation file, and each CT scan is stored in paired .mhd and .raw format files. The dataset was split 80/20 into training and test sets using random stratified assignment.
Lung parenchyma extraction removed confounding anatomical structures. Before training, each CT image underwent a preprocessing pipeline to extract the lung parenchyma and remove interfering structures including bones, blood vessels, and surrounding soft tissue. The pipeline applied denoising, density range adjustment, image binarization to separate lungs from background, morphological operations to remove noise and fill voids, and connected-region marking to define lung boundaries. The resulting parenchymal masks substantially reduced false-positive detections from non-nodule structures.
Six performance metrics balanced accuracy, efficiency, and clinical relevance. Model evaluation used precision, recall, and mean average precision (mAP) to assess detection quality; parameter count (Params in millions) and floating-point operations (FLOPs in gigaflops) to assess computational efficiency; and frames per second (FPS) to assess real-time suitability. This multi-metric framework allowed direct comparison of the accuracy-speed-efficiency tradeoff across competing models rather than optimizing for a single metric.
Training used standard deep learning infrastructure with fixed hyperparameters. Experiments were conducted on an NVIDIA GeForce RTX 3070 Ti GPU with PyTorch 1.13.1. Input images were resized to 640 by 640 pixels. Stochastic gradient descent was used with momentum 0.937, weight decay 0.0005, and an initial learning rate of 0.01 over 200 epochs with batch size 8. All ablation and comparison experiments used the same hardware and hyperparameters to ensure fair comparison.
The model is designed for resource-constrained clinical environments. The primary motivation for lightweight design was suitability for hospitals with large medical image datasets but limited hardware infrastructure. With only 6.64 million parameters and real-time inference at 116.7 FPS, the model can operate on commonly available clinical GPU hardware, removing the barrier that prevents deployment of higher-accuracy two-stage models in community hospital settings.
Combining all three improvements eliminated individual component tradeoffs. Prior works improved accuracy at the cost of speed or reduced parameters at the cost of detection performance. The proposed model demonstrates that carefully designed complementary improvements, where C3_ODC addresses feature extraction quality, GS-BiFPN addresses multiscale fusion efficiency, and F-NWD addresses loss function sensitivity, can simultaneously advance accuracy and efficiency beyond what any single improvement achieves alone.
3D multi-slice feature integration is identified as the primary future direction. The current model converts 3D CT volumes into 2D axial slices for detection, which loses cross-slice feature information that could improve detection of nodules spanning multiple slices. The authors identified extraction of consistent feature information across multiple 2D slices as the primary next step, along with expanding the training dataset to include more diverse and novel nodule morphologies for improved model robustness.