US-based models.

Kainz et al. (2021) conducted a prospective, multicenter study to evaluate a deep learning framework designed to use compression US to detect lower-extremity DVT. The datasets included US video loops collected from two sites using two different devices, ensuring heterogeneity across acquisition sources [29]. The model architecture integrated vessel segmentation, anatomical landmark detection, and compressibility classification into a single pipeline to assess vein patency. Internal and external validation were both performed, with external testing conducted across sites and devices [29]. Their model achieved a sensitivity between 82% and 96%, a specificity between 70% and 82%, and an accuracy between 75% and 83% [29]. This system was designed to guide non-specialists, with radiologists providing the ground truth labels [29]. An additional feature of this work was the evaluation of cost-effectiveness, demonstrating the authors’ emphasis on clinical integration [29].

Nakayama et al. (2023) conducted a prospective study using their ResNet101 on both stationary and portable ultrasound images of the popliteal vein. Instead of identifying the thrombi directly, their deep learning model was developed to determine whether images were of sufficient diagnostic quality to allow reliable assessment of DVT [33]. By automatically categorizing the images into “Satisfactory,” “Moderately Satisfactory,” and “Unsatisfactory” groups, the algorithm assessed the quality of image acquisition [33]. The team also reported comparable accuracies of 76% and 72% for stationary and portable ultrasound, respectively [33].

Oppenheimer et al. (2024) conducted a prospective study across two university hospitals to evaluate an AI-assisted US system integrated within a telemedicine framework for the diagnosis of DVT [34]. It provided novice operators real-time feedback on venous compressibility at the groin and popliteal fossa. The scans were then reviewed by remote specialists [34]. The Auto-DVT model demonstrated 100% sensitivity and 95% specificity [34].

Nothnagel and Aslam (2024) performed a prospective clinical evaluation of the ThinkSono AI-guided US. The algorithm utilized a CNN to analyze US in real time, providing acquisition guidance and classification of venous compressibility [35]. This work demonstrated a point-of-care implementation, in which a physician used the AI tool to acquire and interpret US images [35]. The reported diagnostic performance was 100% sensitivity and 90–92% specificity [35].

Curry et al. (2025) conducted a double-blind, multicenter, prospective trial of AutoDVT, a handheld compression US device used to diagnose DVT at the common femoral and popliteal veins [36]. The model combined CNN optimized for vessel recognition with algorithms that actively assessed probe positioning, vein centering, and compressibility during scanning [36]. The model demonstrated a sensitivity of 68% and a specificity of 80% for the detection of proximal DVT [36].

Avgerinos et al. (2025) conducted a prospective study to evaluate the ThinkSono AI-guided US platform for bedside detection of lower-extremity DVT [37]. Similar to the ThinkSono used in other studies, this model provided real-time prompts to non-expert operators to ensure correct probe positioning and adequate compression. Image adequacy was assessed using the American College of Emergency Physicians (ACEP) quality scoring [37]. This model achieved 100% sensitivity and 95% specificity [37].

Speranza et al. (2025) conducted a retrospective study evaluating the ThinkSono-guided point-of-care US. The study included both radiologists and emergency medicine physicians as operators, who performed compression ultrasound at the common femoral and popliteal veins with real-time AI guidance [38]. The reported diagnostic performance had a sensitivity of 90–98% and a specificity of 74–100% [38]. While ultrasound-based models dominate the literature, MRI studies provided insight into higher-resolution thrombus characterization [30].

MRI-based models.

Sun et al. (2021) retrospectively evaluated a deep learning segmentation model designed for the detection of lower extremity DVT on black-blood MRI. Its architecture of a 3D U-shaped CNN with a generative adversarial network (GAN) framework automatically segmented venous thrombi [30]. It was validated on multicenter datasets using sequences like DANTE-SPACE and DANTE-FLASH [30]. The team noted high performance in their deep learning model on black blood MRI, with a sensitivity ranging from 93% to 95%, a specificity between 92% and 97%, and an accuracy between 94% and 96% [30].

CT-based models.

Seo et al. (2023) conducted a retrospective study and developed a RetinaNet-based object detection framework with different backbones (ResNet50, ResNet152) to detect lower limb DVT on CT venography. The authors generated synthesized image patches with a three-slice input (one image, one image above, and one image below) to provide better anatomical context [15]. This study combined the detection of thrombosis with a focus on data preparation [15]. The reported diagnostic performance showed an average sensitivity of 83.3% [15]. Beyond single-modality approaches, some studies explored hybrid architectures to leverage complementary imaging strengths [39,40].

Studies evaluated across multiple imaging datasets.

Joseph et al. (2024) conducted a retrospective study and presented a deep learning framework termed “Deep Vein Net,” which integrated a CNN with a Sooty Tern Optimization (STO) algorithm. The excellent exploration ability and convergence speed of this metaheuristic optimization method mimic the foraging behavior of sooty terns [40]. In this application, the STO algorithm was used to adjust the hyperparameters of the CNN, thereby enhancing its ability to identify subtle spatial features across several MRI slices [40]. The reported diagnostic performance included a specificity of 96%, an accuracy of 96%, and a precision of 97% [40].

Arun et al. (2025) created the Deep R-Belief Neural Network, which was built from stacked Boltzmann machines. This approach enabled unsupervised pre-training, which was then followed by supervised fine-tuning, allowing the model to capture complex feature representations [39]. The network was evaluated across heterogeneous imaging datasets, allowing assessment of performance on ultrasound, CT, and MRI datasets independently. The algorithm was also trained on more than 12,000 publicly available datasets [39]. Their model achieved a specificity of 97%, an accuracy of 98.9%, and a precision of 97.5% [39].

US-based models.

AutoDVT from Kainz et al. (2021) demonstrated several strengths. It was clinically relevant for bedside use by non-experts and introduced technological innovation by combining vessel segmentation, landmark detection, and compressibility assessment into a single model [29]. The study also included external validation across two sites and devices, as well as cost-effectiveness analyses [29]. However, the authors also recognized several limitations. The authors acknowledged the risk of domain shift, a phenomenon in which the algorithm’s performance deteriorates when the dataset is applied to unfamiliar US devices [29]. Furthermore, the modest positive predictive values (65–89%) indicated that positive results would require expert review before diagnosis [29].

Nakayama et al. (2023) and their model exhibited comparable performance across both portable and stationary US, indicating adaptability in various healthcare settings. A key strength was its focus on image quality and the use of heatmaps to enhance interpretability [33]. The study had several limitations. The sample size was small (n = 20) and reported accuracies were modest (72–76%) [33]. The model lacked external validation and was limited to the popliteal vein, leaving performance at other sites untested. The intermediate category also showed higher misclassification rates, which could potentially lower its real-world applicability [33]. Overall, this study illustrates both the potential and the limitations of applying AI-assisted compressibility assessment to portable US. Further optimization and multicenter evaluation are needed before clinical deployment [33].

Oppenheimer et al. (2024) and their AI model provided novice operators with real-time feedback on venous compressibility. A strength of the study was its focus on clinical integration, combining AI guidance with telemedicine expertise [34]. The tool demonstrated high sensitivity (100%), suggesting promise as a frontline triage tool. Its specificity (95%) indicated fewer false positives compared to other US-based AI systems [34]. The model also prompted users to apply adequate compressions when scanning. However, the algorithm was trained only on two-point compression, leaving calf and iliac thromboses undetectable [34]. The AI tool was also only tested on a small sample size of 70. The authors emphasized the need for larger, prospective validation across diverse settings to establish applicability [34]. AutoDVT demonstrates promise as a rapid triage adjunct, but broader validation across diverse populations and US devices is necessary before clinical adoption [34].

Nothnagel and Aslam (2024) and their ThinkSono Guidance platform enabled non-experts to obtain adequate compression images [35]. A strength of this study was its focus on real-time image acquisition. The American College of Emergency Physicians (ACEP) quality score added objectivity and transparency to the evaluation process by grading images [35]. The scans could also be uploaded for independent expert review, improving diagnostic safety in the workflow [35]. The model also achieved near-perfect sensitivity, which is valuable for minimizing missed DVT cases, while maintaining specificity around 90%. However, this tool had a few limitations [35]. It was limited to a three-region proximal scan and did not evaluate the calves. The sample size was modest, and external testing was not reported, limiting generalizability [35]. Nevertheless, these findings strengthen the case for AI as a supportive triage tool in bedside US [35].

Curry et al. (2025) assessed AutoDVT, an AI system engineered to deliver real-time guidance and assessment of venous compressibility on handheld US [36]. The tool’s strength lies in its design to actively direct probe positioning and maintain reproducibility across operators. It could finish a targeted scan in as short as four minutes [36]. However, the algorithm demonstrated important limitations. It demonstrated modest diagnostic performance with a sensitivity of only 68% and a specificity of 80%, limiting its utility as a standalone tool [36]. Its diagnostic scope was restricted to the femoral and popliteal veins, and sensitivity was modest at 68%. The model was also only trained and validated on a single handheld device, which raised concerns about its generalizability [36]. Further optimization and broader validation across devices are needed before clinical integration [36].

The ThinkSono from Speranza et al. (2025) also directed novice operators through two-point compression scans, giving feedback on probe position, vein centering, and compression adequacy [38]. It had a decent sample size of 381 patients across 11 hospitals, demonstrating robustness in its results. However, the algorithm remained constrained to a two-point compression protocol, with calf and iliac thromboses unaddressed [38]. Inter-observer agreement for image quality was also low, and the system is not autonomous, instead requiring clinician input to establish a diagnosis [38].

Avgerinos et al. (2025) provided important prospective evidence for the clinical utility of ThinkSono in real-world practice. ThinkSono demonstrated high diagnostic accuracy. All scans achieved an American College of Emergency Physicians (ACEP) quality score of 3 or more, indicating the adequacy of the images collected [37]. The AI tool could complete a scan within 7 minutes on average, increasing efficiency. However, its limitations included limited anatomical scope, as the tool was restricted to the femoral and popliteal vein compression, excluding potential calf or iliac DVT [37]. Furthermore, the system required expert input, limiting its role as a fully autonomous diagnostic tool. Lastly, it was tested on a small sample size of 53 and included few DVT cases [37]. Overall, this study highlighted the feasibility and efficiency of AI-guided US for lower-extremity DVT but reinforced the need for larger multicenter trials to confirm generalizability [37].

While US-based AI models were the most extensively studied and clinically feasible, MRI-based approaches provided an opportunity to push diagnostic performance even higher, albeit with trade-offs in cost and accessibility [30].

MRI-based models.

Sun et al. (2021) used a 3D U-shaped segmentation model within a GAN framework to differentiate thrombus from tissues exhibiting similar intensity and morphology on MRI [30]. One strength of this work was its focus on pixel-level segmentation to differentiate thrombus from adjacent tissue with similar morphology. The model was validated on DANTE-SPACE and DANTE-FLASH sequences from three different centers [30]. Performance was strong and outperformed state-of-the-art models such as 3D U-Net, V-Net, 3D nnU-Net, and Cascade nnU-Net [30]. However, the need for extensive preprocessing steps such as resampling, normalization, and data augmentation made clinical implementation computationally demanding. Additionally, the model was not prospectively tested in real-world clinical workflows, and its performance across diverse MRI vendors or protocols was indeterminate [30]. Furthermore, MRI’s high cost and limited availability reduce immediate applicability in routine DVT evaluation, which more often relies on US. Overall, Sun et al. (2021) provided compelling evidence, but larger multicenter trials are needed before MRI-based AI models can be translated into clinical practice [30].

In contrast to MRI, which offers near-perfect diagnostic metrics but limited clinical scalability, CT-based AI studies remain relatively sparse and exploratory, highlighting both the feasibility and current gaps in this domain [15].

CT-based models.

The RetinaNet model from Seo et al. (2023) synthesized three consecutive CT slices to mimic how radiologists assess DVT across multiple images, improving diagnostic accuracy [15]. RetinaNet’s one-stage architecture performed detection and classification concurrently, making it more computationally efficient. Furthermore, the authors addressed data quality by excluding cases with significant artifacts to ensure the reliability of the training dataset. However, the authors noted a few limitations [15]. The AI model was restricted to iliofemoral veins, and clinically important calf veins were excluded. The sample size was modest, raising concerns about generalizability. The model was also not externally validated [15].

Beyond single-modality applications, frameworks using different imaging modalities may represent the next frontier, aiming to leverage complementary strengths across imaging techniques to achieve even greater diagnostic robustness [39,40].

Studies evaluated across multiple imaging datasets.

Joseph et al. (2024) employed “Deep Vein Net,” a hybrid algorithm that integrated discrete waveform pre-processing, Sooty Tern Optimization for feature selection, and a fuzzy Extreme Learning Machine classifier [40]. The model reported outstanding accuracy across CT and MRI, surpassing deep learning networks including Google Net, Ghost Net, and AlexNet [40]. However, several limitations were present. The study was retrospective and lacked external validation. Furthermore, its restriction to CT and MRI, which are not first-line modalities for DVT diagnosis, limited its immediate applicability to clinical workflows [39]. Overall, Joseph et al. (2024) provided a technically sophisticated model, but its clinical utility remains uncertain until validated prospectively in a diverse population [40].

Arun et al. (2025) and their model combined RegNet for feature extraction with Deep Belief classifier. It had high diagnostic performance, outperforming models such as CNN, SESARF, and XGBoost [39]. Although CT, MRI and ultrasound datasets were evaluated separately for comparative analysis, the proposed framework focused primarily on duplex ultrasound-based DVT classification rather than true multimodal fusion. No prospective testing or external validation was reported, raising concerns about generalizability [40]

Overall, these modality-specific findings suggest that ultrasound-based AI models are closest to clinical translation, MRI-based models highlight the technical frontier of accuracy, CT-based models remain exploratory, and approaches incorporating multiple imaging datasets represent an emerging but unvalidated future direction [15,29,30,33–40].