Automatic Lung Nodules Detection Using a Modified YOLOv5

Lung cancer diagnosis involves a detection strategy to determine the specific location of the abnormality and its likelihood whether cancerous or non-cancerous. In existing methods such as Endobronchial ultrasound-guided transbronchial needle extraction (EBUS-TBNA) method requires specific equipment and trained operators. The limits in biomarker discovery begin with sample collection, transportation, representative tissue processing, reference standards, assay sensitivity and specificity. Liquid biopsy method requires tissue biopsy for initial diagnosis and may increase the likelihood of false negatives and false positives. To overcome these challenges, initially in this paper, three detectors are trained for nodule detection i.e. You Only Look Once version 3 (YOLOv3), YOLOv4 and YOLOv5. YOLOv3 achieved precision of 72%, recall of 75%, mean Average Precision (mAP) of 70%, F1 Score of 0.73 and Giga Floating-Point Operations (GFLOPs) of 30. In contrast, YOLOv4 achieved 85% precision, 70% recall, 80% mAP, F1 score of 0.76 and 65 GFLOPs. On the other hand, YOLOv5 achieved precision of 90%, recall of 80%, mAP of 85%, F1 Score of 0.85 and FLOPs of 217. These three detectors also faced few challenges like complexity, have high computation time and low performance. So, to overcome the problems of YOLO based methods, a modified YOLOv5 model has been proposed for the automatic detection of lung nodules in CT scans. Key modifications include enhanced feature extraction layers and customized anchor boxes tailored for small nodule detection. These modifications demonstrate the model's potential for reliable and efficient lung cancer screening, aligning with the manuscript's focus on advancing detection techniques through customized YOLOv5 enhancements. The modified model achieves a 90% precision, 85% recall, 88% mAP, 0.87 F1 score and 35.2 GFLOPs. These results represent an improvement in accuracy and increase in sensitivity compared to the standard YOLOv5 model. Also, GFLOPs have been reduced which demonstrates low computing requirement for the proposed model. The proposed model could be further used as clinical tool for lung cancer diagnosis.