Application of an improved YOLOv5s-based deep learning model for automated detection of pulmonary adenocarcinoma in situ and minimally invasive adenocarcinoma
The improved YOLOv5s-based deep learning model performed better in automated detection, and the LCSA × ASFF combination improved localization accuracy while maintaining recall through the synergy between local contrast enhancement and adaptive scale fusion.
Abstract
Objective To investigate the performance and clinical application potential of an improved YOLOv5-based deep learning model for automated detection and classification of pulmonary adenocarcinoma in situ (AIS) and minimally invasive adenocarcinoma (MIA), with particular emphasis on attention mechanisms and multiscale feature fusion for small pulmonary nodule detection. Methods CT images of 225 pathologically confirmed AIS or MIA lesions treated at the Affiliated People's Hospital of Fujian University of Traditional Chinese Medicine between December 2019 and June 2025 were retrospectively collected. In the deep learning workflow, the 225 lesions were first divided at the lesion level: 27 lesions (11 MIA and 16 AIS) were set aside, and the 218 original PNG images derived from these lesions were reserved as an independent test set that was excluded from all data augmentation and model training. The remaining lesions were converted into PNG format and further augmented using horizontal flipping, yielding a final dataset of 3,000 images for training and validation. Subsequently, the dataset was partitioned into training and validation sets at an 8:2 ratio. During this process, the origins of the augmented lesion images were manually verified to ensure that an original image and its corresponding augmented variants were not assigned to different subsets simultaneously. YOLOv5s was used as the baseline model. Local contrast-spatial attention modules (LCSA and LCSAv2), BiFPN, and ASFF were incorporated for attention enhancement and multiscale feature fusion. Five-fold cross-validation was performed. Mean average precision (mAP), precision, and recall were evaluated across IoU thresholds from 0.6 to 0.9, and ablation experiments were conducted to quantify the contribution of each module. Results In the baseline model selection stage, comparison under identical conditions showed that YOLOv5s achieved detection performance comparable to YOLOv11s while reducing GFLOPs by approximately 25%; therefore, YOLOv5s was selected as the improvement baseline. In five-fold cross-validation at IoU = 0.6, baseline YOLOv5s achieved an mAP@50 (AP at the matching IoU of 0.6) of 0.947, precision of 0.849, and recall of 0.932. After LCSA x ASFF was introduced, recall increased to 0.941 and mAP@50 was maintained at 0.946. Under the stringent localization criterion of IoU = 0.9, LCSA × ASFF achieved an mAP@50 of 0.878, a 3.8-percentage-point higher value than the baseline (0.846); precision (0.879) and recall (0.773) were also the highest among all models, with the smallest performance attenuation in the high-IoU range. Ablation experiments showed that the combined use of LCSA and ASFF produced stronger synergistic effects in the high-IoU range than either module alone, whereas LCSAv2 combined with ASFF did not show the expected synergistic gain. On the independent test set, baseline YOLOv5s achieved precision, recall, and mAP@50 values of 0.747, 0.715, and 0.753, respectively; the improved LCSA × ASFF model achieved corresponding values of 0.693, 0.753, and 0.788. Recall and mAP@50 increased simultaneously, indicating better overall detection performance than the baseline. The improved YOLOv5s-based deep learning model performed better in automated detection, and the LCSA × ASFF combination improved localization accuracy while maintaining recall through the synergy between local contrast enhancement and adaptive scale fusion. Conclusion The improved YOLOv5s model achieved automated detection and classification of small pulmonary nodules through end-to-end training, showing clear advantages in clinical workflow automation without manual intervention. The synergistic effect of local contrast enhancement and adaptive scale fusion effectively improved AIS/MIA localization precision and classification accuracy, supporting its potential as an assistive tool for early lung adenocarcinoma screening, pending external multi-centre validation.
Due to its late identification and challenging diagnosis, lung cancer continues to be one of the top causes of death for cancer patients globally, positioning it as one of the most critical concerns. Timely identification of cancerous nodules is essential for enhancing the patient’s survival likelihood CT image analysis by hand is not very productive and significantly relies on a specialist’s expertise. In this study, we offer an autonomous lung cancer classification method based on explainable deep learning. The popular DenseNet121 network serves as the foundation for our deep learning model, which is enhanced by the Convolutional Block Attention Module (CBAM). To improve feature extraction of significant spatial and channel properties of input data, attention techniques are added. Furthermore, our method is interpretable because the Grad-CAM technique makes it possible to explain the choices made by a machine learning system. A database of CT scans, comprising 4,598 pictures categorized by large cell carcinoma, adenocarcinoma, and healthy lungs, was utilized. Our evaluations show the model’s effectiveness with an accuracy rate of 94.6\%.
S. Jegadeesan, S. Matheswaran, R. Palanivelrajan· International Conference on...· 0 citations
The proposed system demonstrates potential as an assistive tool for automated lung cancer screening, warranting further validation on larger and multi-institutional datasets before clinical application.
Vishwas V. Patange, Jagadish B. Jadhav, Sanjay L. Nalbalwar et al.· Scientific Reports· 0 citations
INTRODUCTION/OBJECTIVE
Many other diseases can produce a similar pattern; therefore, the diagnosis of IPF is made by exclusion, based on the absence of alternative pathologies. To differentiate IPF from other pathologies, the study developed a deep learning model with an attention mechanism to improve performance.
METHODS
This retrospective, single-center study included 96 patients (46 IPF and 50 non-IPF) with a typical UIP pattern on HRCT. The data were split at the patient level into training (70%), validation (15%), and test (15%) sets using stratified randomization (seed = 42). Squeeze-and-Excitation (SE) blocks were incorporated after each convolutional stage of VGG-16 to enable channel-wise feature recalibration. The model was trained using the AdamW optimizer (learning rate = 1×10-4, batch size = 32, 50 epochs, early stopping with a patience of 7), with augmentation applied only to the training set. Test predictions were generated from original, unaugmented slices, and patient-level labels were determined via majority voting. Gradient-weighted Class Activation Mapping (Grad-CAM) was applied post hoc for heatmap visualization. Model performance was evaluated using accuracy, sensitivity, specificity, F1-score, Area Under the Curve (AUC), and Cohen's kappa.
RESULTS
The SE-VGG-16 model achieved a patient-level accuracy of 87.2%, sensitivity 86.1%, specificity 88.4%, F1-score 0.870, and AUC 0.91. Ablation studies demonstrated a 4.1% accuracy improvement over baseline VGG-16. Grad-CAM heatmaps consistently highlighted subpleural reticular opacities and honeycombing regions, aligning with established radiological criteria.
DISCUSSION
The study has successfully distinguished IPF from other interstitial diseases with a high degree of accuracy. Despite the existence of numerous studies in the literature on the differential diagnosis of IPF, research on deep learning and studies combining deep learning with attention maps are quite limited. Unlike traditional deep learning models, the use of an attention mechanism has enabled the model to focus on pathological regions, thereby producing more reliable results. The findings present a potential approach that could be used in clinical decision support systems.
CONCLUSION
Based solely on radiological images, the attention-enhanced VGG-16 model achieved high accuracy in the differential diagnosis of IPF.
Hüseyin Alper Kızıloğlu, Kenan Zengin· Current medical imaging· 0 citations
Globally, colorectal cancer (CRC) remains a key contributor to cancer‐related death, with most malignancies developing through the progression of colorectal polyps. Early detection and accurate histological classification of polyps during colonoscopy are essential for effective CRC screening and prevention. However, conventional colonoscopy may fail to detect certain lesions and exhibits variability in diagnostic performance due to operator dependence and challenging imaging conditions. To address these limitations, this study proposes a hybrid deep learning (DL) model that integrates YOLOv10 for polyp detection and feature extraction with a customized Convolutional Neural Network (CNN) for the histological classification of colorectal polyps into hyperplastic and adenomatous categories. A total of 6000 endoscopic images obtained from the Harvard Dataverse PolypsSet repository were used for model development and evaluation. To improve robustness and generalization, data augmentation techniques were applied during training, and stratified 5‐fold cross‐validation was employed to prevent data leakage between training and validation sets. Experimental results demonstrated that the proposed YOLOv10–CNN model achieved an average detection mAP@50 of 0.9848 and a classification accuracy of 0.9913 across the cross‐validation folds. External validation on an independent dataset achieved mAP@50 of 0.926, indicating good generalization ability to unseen data. Furthermore, the model achieved an inference speed of approximately 120 frames per second (FPS), demonstrating efficient computational performance. A web‐based graphical user interface was also developed to facilitate visualization of detection and classification results from colonoscopy videos. The findings suggest that the proposed hybrid model provides accurate and efficient polyp detection and classification while maintaining stable performance across internal and external evaluations. The proposed approach may serve as a supporting computer‐aided analysis tool for colorectal polyp screening.
Yao-Tien Chen, Debalke Embeyale Sahilu· International journal of ima...· 0 citations
Lung cancer (LC) is among the leading causes of cancer-related deaths, and early identification of pulmonary nodules plays a crucial role in reducing mortality. This work proposes a novel hybrid framework that combines preprocessing, segmentation, feature extraction, and classification for accurate nodule detection from computed tomography scans. Digital imaging and communications in medicine images are first preprocessed using median filtering to remove noise and contrast limited adaptive histogram equalization to enhance contrast. Segmentation of lung regions is performed through histogram-based thresholding and connected component analysis. To improve feature quality, a minimum repetition and a wolf search algorithm are applied for heuristic feature selection, followed by feature learning using AdaBoost. The selected features are then classified using deep learning architectures, including LeNet, AlexNet, and VGG16, with softmax for final prediction. Experimental results show that AlexNet with SGD achieved the best performance with 97.42 % accuracy and an F1-score of 97.58 %, outperforming Adam and LeNet. Although LeNet (SGD) reached 95.9 % accuracy, its sensitivity (94.76 %) was slightly lower than that of AlexNet. The Adam optimizer provided competitive results but generally underperformed compared to SGD in both architectures. Overall, SGD-optimized AlexNet offers the most reliable balance of sensitivity, specificity, and predictive values for LC nodule detection. This integration of ensemble learning and deep neural networks enhances detection accuracy, reduces reliance on handcrafted features, and offers a robust solution to support radiologists in early LC diagnosis.
Ghousia Usman, Usman Ahmad, Huma Tauseef et al.· Journal of Testing and Evalu...· 0 citations
This study aims to explore and develop advanced methods for the effective detection of breast pathologies using state-of-the-art machine learning techniques, specifically YOLOv8 and Faster R-CNN. Traditional approaches to breast disease diagnosis are critically reviewed, and their effectiveness is evaluated in comparison to modern automated methods. The proposed models are applied to mammographic images to identify and categorize pathological patterns into six distinct levels, considering variations in severity and disease characteristics. This multi-level classification allows for a more precise assessment of disease progression and provides critical information for personalized treatment planning.
Experimental results demonstrate that the proposed approach achieves high accuracy and fast image processing, enabling reliable and rapid detection of potential breast abnormalities. These findings suggest that machine learning algorithms can significantly enhance the diagnostic process, providing clinicians with more accurate and timely information. Furthermore, the study highlights the potential of automated detection systems to improve early diagnosis, optimize treatment strategies, and ultimately enhance patient outcomes. The results emphasize the growing role of artificial intelligence in medical imaging and its transformative impact on the future of breast disease management.
A. Orazayeva, N. Smakova, K. Maksutova et al.· Bulletin of Shakarim Univers...· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
