Lung cancer screening relies on CT imaging, but small nodules are easily missed. CT has become the standard modality for lung cancer screening due to its high sensitivity, but manual interpretation of CT scans is time-consuming, subject to inter-observer variability, and limited in detecting subtle or low-contrast lesions. Small pulmonary nodules, particularly those 6 mm or smaller, present the greatest diagnostic challenge because they represent the earliest detectable stage of malignancy.
Existing AI detection models fail in three key areas for clinical use. Current YOLO-based object detection models struggle with identifying small nodules, handling low-contrast lesions against complex anatomical backgrounds, and reducing false positives caused by vessels, airway walls, and pleural interfaces that visually mimic nodules. These limitations have prevented widespread clinical adoption of automated detection systems.
HARM-YOLO is designed to address all three limitations simultaneously. Built on the YOLOv10 framework, HARM-YOLO introduces four targeted architectural innovations: a multi-dimensional receptive field feature extractor, a decoupled neck for multi-scale semantic fusion, series and parallel receptive field enhancement modules, and a background attention mechanism. Together, these components aim to achieve high sensitivity for small nodules while maintaining real-time inference speed.
Standard convolutional networks have fixed receptive fields that are insufficient for lung nodule detection. Conventional CNN backbones process images using fixed 3x3 kernels that capture only local neighborhoods, missing the fine-grained textural patterns of early-stage nodules and the broader anatomical context needed to distinguish them from normal structures. This is a fundamental limitation when nodules can range from 3 mm to several centimeters.
The C2f-MDR module integrates dilated convolutions with multiple dilation rates. By applying convolutions with dilation rates of 1, 2, and 4 in combination, the module simultaneously captures fine-grained local texture and broader spatial context. The effective receptive field is the union of all individual receptive fields across the different dilation rates, allowing the model to encode hierarchical representations from minute nodule textures to broader anatomical structures.
Channel-to-feature connections reduce redundancy and improve spatial coherence. The module combines receptive field expansion with 1x1 convolutions that reduce inter-channel redundancy while enhancing spatial feature coherence. Multi-level feature aggregation then combines feature maps from different network layers using learned weights, enabling the model to learn hierarchical representations robust to variations in tumor size, shape, and contrast.
All combined features pass through ReLU activation for non-linear refinement. After dilated convolutions, channel-to-feature connections, and multi-dimensional aggregation, the final feature map is processed through a ReLU activation function, introducing non-linearity and ensuring the model retains only the most discriminative information for tumor detection before passing features to downstream layers.
The Decoupled Neck module resolves semantic conflicts in multi-scale feature fusion. Traditional YOLO neck architectures combine deep high-semantic features with shallow low-semantic features through simple concatenation, creating semantic conflicts that degrade detection of subtle nodules. The DENeck module separates high-semantic (global context) and low-semantic (local detail) pathways into distinct branches, processes them independently, and fuses them using learned weights, preventing interference between these complementary information types.
SRFEM expands receptive fields sequentially through three dilation rates. The Series Receptive Field Enhancement Module applies dilated convolutions with dilation rates of 1, 3, and 5 in sequence, encoding hierarchical features at progressively larger scales. A SiLU activation function with a residual connection combines the sequentially processed features, enabling the model to integrate fine structural details with global contextual information critical for identifying small nodules with weak contrast or irregular boundaries.
PRFEM captures multiple spatial resolutions within a single layer through parallelism. The Parallel Receptive Field Enhancement Module runs three convolutional branches with different dilation rates simultaneously, concatenating their outputs into a unified feature map. This parallel design allows the network to capture features at different spatial resolutions in a single pass, improving sensitivity to nodules of varying sizes and morphologies without the sequential processing overhead of SRFEM.
SRFEM and PRFEM together provide complementary multi-scale coverage. The sequential design of SRFEM progressively builds context across scales, while the parallel design of PRFEM ensures simultaneous multi-scale representation. Together, they improve both sensitivity and specificity, a property of particular importance for lung cancer screening where small nodules frequently exhibit low contrast against surrounding tissue, contributing an absolute gain of 4.8% in mAP@0.5 according to ablation studies.
The Background Attention Module directly suppresses false positives from anatomical structures. Vessels, airway walls, and pleural interfaces in lung CT scans can exhibit intensity patterns similar to nodules, causing spurious detections that undermine clinical trust. The Background Attention Module applies a sigmoid-activated convolution that produces attention weights, then multiplies them element-wise with the feature map, adaptively down-weighting activations associated with irrelevant background patterns while strengthening sensitivity to subtle nodules.
The Multi-level Detection Head aggregates features from multiple network depths. Unlike traditional detectors that use a single feature level for prediction, HARM-YOLO's detection head combines feature maps from multiple layers, each weighted by learned parameters. Higher layers provide global semantic context while lower layers provide fine-grained local detail, enabling the head to detect tumors across a wide range of sizes and contrast levels within a single unified prediction stage.
The combined loss function jointly optimizes localization and classification. Model training uses a composite loss that combines bounding box regression loss (squared error between predicted and ground-truth coordinates), classification loss (cross-entropy between predicted class probabilities and radiologist-annotated ground truth), and a regularization term that penalizes excessive weight magnitudes to prevent overfitting. The ground truth annotations from experienced radiologists in LIDC-IDRI and LUNA16 serve as the supervisory signal throughout.
Grad-CAM and Score-CAM visualizations confirm the model focuses on nodule regions. Interpretability analysis using gradient-weighted class activation maps shows that HARM-YOLO's high-attention regions consistently align with radiologist-annotated nodule locations rather than background anatomical structures. This provides clinicians with intuitive visual validation that the model's decisions are grounded in clinically meaningful image features, addressing a key barrier to clinical acceptance of automated detection systems.
HARM-YOLO achieves 90.8% mAP@0.5 on LIDC-IDRI, outperforming all baselines. Against classical two-stage detectors like Faster R-CNN (84.2% mAP) and one-stage detectors like RetinaNet (84.0% mAP), HARM-YOLO shows 6 to 7 percent absolute improvement. Among YOLO-based methods, the closest competitor ELCT-YOLO achieves 88.1% mAP, still 2.7% below HARM-YOLO. Segmentation-based models nnUNet and Swin-UNet reach 86.2% and 86.5% mAP respectively, despite higher raw recall, due to excessive false positives from redundant bounding boxes derived from pixel-level predictions.
On LUNA16, HARM-YOLO achieves 92.5% mAP@0.5 with 91.2% precision and 88.9% recall. HARM-YOLO outperforms YOLOv5 by 1.8% in precision and 1.7% in recall, and improves over MSG-YOLO and ELCT-YOLO by 1.1 to 1.6% across all metrics. The segmentation models nnUNet and Swin-UNet achieve higher recall (89.1% and 88.5%) but significantly lower precision (84.7% and 85.3%), a 6.5% precision disadvantage that corresponds to approximately 28 more false positives per test set, strongly favoring HARM-YOLO for clinical deployment where false alarms undermine radiologist trust.
Cross-dataset generalization experiments confirm real-world applicability. When trained on LIDC-IDRI and tested on LUNA16 (or vice versa), HARM-YOLO consistently outperforms all baseline methods across both transfer directions. Training on LIDC-IDRI and testing on LUNA16 yields 87.3% precision, 84.9% recall, and 88.6% mAP@0.5. This robustness to domain shift caused by differences in imaging protocols, scanner types, and annotation standards is critical for deployment across multiple clinical centers.
FROC analysis shows 89.5% sensitivity at 1 false positive per scan, a 5.4% improvement over the strongest baseline. Free-response receiver operating characteristic analysis at clinically relevant operating points confirms that HARM-YOLO achieves approaching-clinical-threshold sensitivity while maintaining acceptable false positive rates. This metric directly addresses the screening program requirement that automated systems must be sufficiently sensitive to justify deployment without overburdening radiologists with false alarms.
Small nodule detection (6 mm or less) is where HARM-YOLO shows the most clinically important gains. For nodules at or below 6 mm in diameter, HARM-YOLO achieves 85.7% precision, 82.1% recall, and 84.9% mAP@0.5. These are the hardest cases in clinical screening, representing the earliest detectable malignancies, and where detection failures have the greatest impact on patient outcomes through missed early-stage cancers.
HARM-YOLO outperforms YOLOv5 by 2.8% in precision and 2.3% in recall for small nodules. Compared to MSG-YOLO and ELCT-YOLO, improvements range between 1.5 and 2.0% across all metrics. These gains are attributable specifically to the SRFEM and PRFEM receptive field enhancement modules and the background attention mechanism, which together enable the model to detect subtle textures and boundary irregularities that characterize early-stage malignant nodules.
Segmentation models achieve higher recall but suffer from severe precision loss on small nodules. nnUNet and Swin-UNet reach 83.4 to 83.9% recall on small nodules, but their precision falls to 77 to 78%, producing a large number of false positives. This imbalance makes segmentation-based approaches poorly suited for clinical deployment in small nodule screening, where each false alarm requires additional follow-up imaging and clinical evaluation, increasing patient anxiety and healthcare system burden.
HARM-YOLO is designed specifically for large-scale screening program deployment. The model's real-time inference speed, capable of processing over 100 CT volumes per hour on standard clinical hardware, combined with its high sensitivity for small nodules, makes it a viable tool for population-level lung cancer screening programs. This addressees the throughput requirements that have limited adoption of more accurate but slower architectures like Swin-UNet.
The 5.4% improvement in sensitivity at 1 false positive per scan approaches clinical screening requirements. Achieving 89.5% sensitivity at the 1 FP/scan operating point is clinically meaningful because it represents the threshold at which automated systems can meaningfully augment radiologist reading without generating excessive follow-up burden. This performance level supports HARM-YOLO's role as a first-read AI system that flags cases for radiologist review rather than serving as a standalone diagnostic tool.
Cross-dataset generalization reduces barriers to multi-center deployment. Robust performance under domain shift conditions, where training and testing data come from different institutions with different imaging protocols and scanners, is a prerequisite for practical clinical deployment. HARM-YOLO's demonstrated generalization from LIDC-IDRI to LUNA16 and vice versa indicates that a single trained model could potentially be deployed across diverse clinical environments without institution-specific retraining.
Interpretable attention maps support radiologist trust and workflow integration. The Grad-CAM and Score-CAM visualizations that show the model's attention concentrated on annotated nodule regions provide clinicians with a validation mechanism that goes beyond accuracy metrics alone. When a model's decision can be visually grounded in the same anatomical features that radiologists themselves examine, it becomes easier to integrate the tool into clinical workflows as a trusted decision-support system rather than a black box.