LSTAC provides an efficient and accurate solution for automated CEUS analysis, reducing manual workload and improving consistency in thyroid nodule assessment, with potential for clinical application.
Abstract
Background
AND
Objective
Contrast-enhanced ultrasound (CEUS) is widely used for evaluating thyroid nodule malignancy, but conventional time-intensity curve (TIC) analysis is labor-intensive and operator-dependent. This study proposes LSTAC, an automated framework for nodule segmentation and TIC analysis in CEUS videos.
Methods
LSTAC integrates an improved YOLOv5-based segmentation network with a peak intensity frame (PIF) extraction algorithm to enable automatic nodule localization, TIC generation, and PIF identification. The framework was trained using CEUS data from 623 patients collected across three hospitals and evaluated on both internal and external validation cohorts.
Results
LSTAC achieved 3-10× higher efficiency than VueBox in PIF extraction while maintaining strong temporal accuracy (0.94, 0.77, 0.79) and structural similarity (SSIM: 0.80, 0.60, 0.67). In malignancy prediction based on PIF features, LSTAC outperformed VueBox in two of three validation sets, with AUCs of 0.8279 vs. 0.8226 and 0.8000 vs. 0.7000.
Conclusion
LSTAC provides an efficient and accurate solution for automated CEUS analysis, reducing manual workload and improving consistency in thyroid nodule assessment, with potential for clinical application.
Breast cancer is one of the leading causes of death in women and early diagnosis and correct diagnosis is important. Although ultrasound imaging is widely used due to its non-invasive nature and is low cost, suitable for real time clinical assessment, lesion segmentation is challenging because of speckle noise, low contrast, shadowing, fuzzy boundaries and variations in lesion size and shape. This paper proposes an improved DeepLabv3+ segmentation framework using EfficientNetB4 as the encoder to localize the breast lesions in ultrasound images. It was chosen because of its ability to achieve a high accuracy to computation ratio via compound scaling and can also be used to generate rich multi-scale feature representations, using fewer parameters than its heavier counterparts. The public Breast Ultrasound Images (BUSI) dataset is subjected to image resizing, normalization, Contrast Limited Adaptive Histogram Equalization (CLAHE), synchronized augmentation and pixel-wise mask learning during the proposed pipeline. CLAHE has the effect of increasing the contrast of the lesion boundary region in it and decreasing the ambiguity of the boundaries prior to feature extraction which enhances the interpretability of the segmentation output. Experimental analysis shows an accuracy of 94.57%, precision of 84.11%, recall of 42.59% and F1-score of 56.55%. While recall is still moderate for difficult lesions, the model shows stable convergence, high pixel-level accuracy and clinically useful lesion localization. The new study also provides more insight into the split of the data set and the composition of the data set and the comparison with other latest segmentation models.
B. B. Jayasingh, N. Monika· 2026 4th International Confe...· 0 citations
In order to improve patient outcomes, rapid and repeatable characterisation of hepatic tumours by computed tomography (CT) is essential. Liver cancer is still one of the most deadly cancers in the world. Although traditional volumetric (3D) segmented networks accurately identify lesions, their utility for real time screenings on commodity hardware is restricted due to their high computational cost and delay. A compact a convolutional neural network (CNN) which classifies hepatic regions as benign, malignant, or normal is the main contribution of this works lightweight classification framework. An upstream, anchor less YOLOv8 stage provides candidate regions for lesion localisation. A semi supervised U-Net pipeline is used to transform volumetric research from the LiTS, 3DIRCADb, along with CHAOS collections into annotated 2D axial slices. While the upstream localisation step runs at about 30 ms per slice, the CNN classification stage achieves a mean per class accuracy of about 95% (0.96 normal, 0.94 benign, and 0.95 malignant) with minor, symmetric inter class confusion on the curated test set. A Flask web interface that overlays areas, confidence scores, as well as lesion labels in real time, provides access to the framework. The findings show that early liver cancer screening is a good fit for a lightweight, scalable classification pipeline, future research should focus on improving small lesion localisation.
Megha J., Gopalakrishna Madigondanahalli Thimmaiah, Vishruth B. Gowda· International journal of com...· 0 citations
Computed tomography (CT) images have poor tissue contrasts, irregular appearance of lesions, and unclear tumor borders which limit reliable diagnosis of liver malignancy. This paper introduces an attention-controlled automated liver tumor segmentation and classification based on a customized Mask Region Convolutional Neural Network (tm-RCNN) with an addition of Multi-Scale Attention Gate (MSAG). To increase the local contrast and reduce artifacts caused by acquisition, adaptive histogram equalization is used. The suggested MSAG selectively elevates boundary-sensitive features in a variety of spatial resolutions, which allow accurate delineation of lesions within the tm-RCNN decoder. Deep residual features, geometric form descriptors and enhanced median binary pattern (e-MBP) textures are extracted using segmented areas and then classified using a hybrid SqueezeNet DeepMaxout ensemble with score-level fusion. It has been experimentally validated on two benchmarking CT sets with a higher performance that achieves a Dice coefficient of 0.9587, classification accuracy of 0.936, sensitivity of 0.961 and less computational time of 64.21 s than the state-of-the-art. These findings affirm the usefulness and clinical appropriateness of the suggested framework.
Babeetta Bbhagat, Mohini Kumbhar, Swati Powar et al.· Journal of Intelligent Decis...· 0 citations
Objectives: Ultrasound is the primary modality for salivary gland tumor (SGT) evaluation, yet its reliance on subjective interpretation can lead to diagnostic variance. This study aims to develop and validate a two-stage deep learning system to automate SGT detection and classification. Methods: The study compiled a dataset of ultrasound images from patients with pathologically confirmed SGTs across three sequential cohorts: a training set (687 images, 2007–2020), a validation set (78 images, 2021), and a testing set (101 images, 2022). A YOLOv8 model was trained for tumor detection, and a modified ResNet50V2 model was utilized for benign versus malignant classification. The resulting two-stage pipeline was deployed on a local desktop system and further evaluated using two independent datasets: an internal validation set (56 images, 2023) and an external online dataset (57 images). Results: On the testing set, the YOLOv8 model achieved a bounding-box precision of 0.94 and a recall of 0.95 for tumor detection. When integrated with the classification model, the two-stage desktop system yielded an accuracy of 84%, sensitivity of 74%, and specificity of 87%. This system maintained comparable performance, demonstrating accuracies of 82% and 81%, sensitivities of 100% and 71%, and specificities of 81% and 86% on the internal and external validation sets, respectively. Conclusions: This study introduced a two-stage deep learning desktop system for automated SGT diagnosis. The edge-deployed system may serve as an objective adjunct to conventional ultrasound interpretation, potentially assisting clinicians during preoperative evaluation.
RATIONALE AND OBJECTIVES
Follicular neoplasms are difficult to classify by ultrasound, as they often present as thyroid adenomatoid nodules (TANU). Furthermore, fine-needle aspiration cytology (FNAC) exhibits limited accuracy in differentiating the nature of follicular lesions. This study aimed to develop and validate a novel model to optimize the workflow for distinguishing benign from malignant TANU.
MATERIALS AND METHODS
This study enrolled 648 TANUs, which were divided into the training (Train), validation (Val), external test (Test), and surgically confirmed (SC) cohorts. Handcrafted radiomics features were extracted using PyRadiomics, and deep features were obtained using pretrained ResNet models. Robust features were selected via LASSO regression. The deep learning radiomics (DLR) signature integrated the deep transfer learning (DTL) and radiomics signatures. The combined signature integrated the DLR and clinical signatures. The performance of different signatures was evaluated using the Area Under the Curve (AUC). We also compared the diagnostic performance of our best signature with FNAC. Gradient-weighted Class Activation Mapping (Grad-CAM) visualizations were used to interpret the deep learning component.
RESULTS
The DTL signature achieved the highest AUC of 0.959 in the Test cohort (the DLR, combined, radiomics, and clinical signatures were 0.937, 0.930, 0.835, and 0.599, respectively). Furthermore, the DTL signature outperformed FNAC in diagnostic accuracy, with higher specificity (96.2%) and overall accuracy (89.3%). Grad-CAM confirmed that key image regions significantly contributed to the signature's decisions.
CONCLUSION
By eliminating the need to integrate clinical and radiomics features, the DTL signature streamlines the diagnostic process while maintaining robust performance, making it a highly effective tool for clinical decision-making.
Ultrasound (US) is the primary imaging modality for thyroid evaluation, yet its diagnostic accuracy remains highly operator dependent. Automated identification of thyroid anatomy and adjacent vascular structures may improve standardisation, support image-guided procedures, and enhance preoperative decision-making in endocrine surgery. In this study, 300 B-mode ultrasound videos from 280 patients (111,679 frames) were used to develop a deep learning-based segmentation model for simultaneous delineation of the thyroid gland and carotid artery. Data were split at the patient level into training/validation (n = 240) and an independent test set (n = 60). An expert-in-the-loop annotation workflow combining manual annotation, model-assisted pseudo-labelling, and expert refinement was implemented to enable efficient dataset construction. Segmentation performance was evaluated using the Dice similarity coefficient (DSC) and intersection over union (IoU). On the independent test set, the model achieved a DSC of 0.96 ± 0.02 and 0.96 ± 0.03 for the two segmented structures, with corresponding IoU values of 0.93 ± 0.03 and 0.93 ± 0.05. The model maintained consistent performance across varying anatomical presentations, including pathological nodules. Separately, evaluation of the expert-in-the-loop annotation pipeline showed high agreement between model-generated pseudo-labels and final expert-corrected annotations, indicating that the semi-automatic approach provided a strong initial approximation requiring only limited refinement. This multi-structure, video-based segmentation approach provides a clinically relevant foundation for artificial intelligence-assisted thyroid imaging, with potential applications in anatomical localisation, thyroid volume estimation, procedural guidance, three-dimensional reconstruction, and image-guided or robotic ultrasound systems.
M. Bolomiti, Joel Burman, K. Radiya et al.· Updates in Surgery· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
