AI-powered medical imaging for ventral hernia repair

Understanding pathophysiology

Incisional hernias (IH) are a common, morbid, long-term complication after abdominal surgery. They occur in approximately 1%-10% of all abdominal surgeries, and they are increasingly common in the United States and globally^[1,2]. Advancements in artificial intelligence (AI) present an opportunity to deepen the understanding of hernia pathophysiology, an area that remains incompletely understood^[3].

McAuliffe et al. utilized preoperative computed tomography (CT) imaging to train a deep learning (DL) model to identify linear and volumetric measurements of viscera, adipose, and muscle in the abdominopelvic cavity. Their work showed that widening of the rectus complex, increased visceral volume, and abdominopelvic skeletal muscle atrophy were associated with the eventual development of IH^[4]. These findings suggest a common pathophysiologic mechanism for developing ventral hernia, rooted in increased intra-abdominal volume or pressure against a weakened abdominal wall.

Surgeons can use this work to preoperatively assess radiological images to evaluate the risk of developing hernia. In their follow-up work, Talwar et al. used an extended dataset and employed three machine learning (ML) techniques - Ensemble Boosting, Random Forest, and support vector machine - to identify optimal CT-based biomarkers (OBMs) that could predict IH formation^[3]. The study found a parsimonious model of three OBMs achieved a predictive accuracy of 0.83, underscoring the potential of AI to synthesize imaging data and identify patterns linked to hernia formation^[3]. These findings lay the groundwork for future studies using preoperative CT scans for risk stratification in VHR.

Preoperative surgical planning for VHR

Advancements in AI-based imaging have demonstrated significant potential in enhancing preoperative planning for VHR.

Love et al. introduced a method to predict the need for additional myofascial release (AMR) during open Rives-Stoppa hernia repair based on the rectus width to hernia width ratio (RDR) measured from CT scans^[5]. Their approach relied on manual measurements of key anatomical features and was shown to be a practical tool, showing a linear decrease in likelihood of requiring AMR as RDR increased^[5]. Though useful for planning, 10% of patients with the highest RDR still required AMR. Additionally, the model failed to account for key biomechanical features, like the elasticity of the abdominal wall, underscoring the need for tools that leverage AI and CV to provide advanced insights into surgical planning.

Building on this foundation, Elhage et al. applied DL models to enhance preoperative planning for patients undergoing abdominal wall reconstruction (AWR), a procedure to restore complex abdominal hernias. They developed three convolutional neural network (CNN) models to predict the need for component separation, surgical site infections (SSIs), and pulmonary complications^[6,7]. CNNs consist of multiple layers that process images; the initial layers detect simple patterns, such as contrast differences, while deeper layers identify more complex structures, like shapes^[8]. This hierarchical feature extraction enables CNNs to recognize anatomical structures, making them well-suited for image analysis.

Using a dataset of 369 patients and 9,303 preoperative CT images, they trained the models with an 80/20 split, classifying each image as “yes” or “no” for each prediction task^[6]. The models then generated a probability score (0-1), reflecting the likelihood of each outcome. Elhage et al. found that the CNN models outperformed surgeons in predicting the need for component separation and accurately identified SSIs^[6]. However, the models struggled with detecting pulmonary complications, likely due to the small sample size of rare outcomes. To address these limitations, the authors proposed architectural refinements to improve model efficacy.

In a follow-up study, the same group used a generative adversarial network anomaly (GANomaly) model to improve predictions for rare complications following AWR. It was trained on an expanded dataset of 510 patients with 10,004 CT images. The GANomaly model generated synthetic CT images to better represent low-frequency events such as mesh infection and pulmonary failure^[9]. Compared to the prior model, the GANomaly model improved sensitivity, increasing from 25% to 68% for mesh infection and 27% to 73% for pulmonary failure^[9]. The model’s predictive ability improved from 0.61 to 0.73 for mesh infection and 0.59 to 0.70 for pulmonary failure^[9]. However, precision declined from 93% to 78% for mesh infection and 92% to 67% for pulmonary failure^[9]. This highlights a trade-off: while sensitivity improved for rare complications, the risk of overdiagnosis increased, a useful lesson for future AI development.

To further refine predictive modeling, Wilson et al. compared three DL models - one trained on structured clinical data alone, one trained on CT imaging, and one combining both datasets to predict complications^[10]. The clinical-only DL model, incorporating variables such as age, BMI, diabetes history, and tobacco use, outperformed both the image-only and mixed models^[10]. However, the researchers noted that the small sample size and unbalanced dataset likely contributed to the underperformance of the image-based models in predicting hernia recurrence^[10]. Future work should focus on identifying image-based features that enhance predictive accuracy and ensuring the use of larger, more diverse datasets for model training. Despite these challenges, DL applications in VHR should be further explored to augment preoperative planning. Table 1 summarizes the studies reviewed in this section, comparing model types, tasks, training data, and key findings.

Intraoperative decision support

Augmented reality (AR) systems that integrate imaging and AI offer an avenue for improving intraoperative decision making in VHR. While the use of AR in VHR is nascent, real-time AI-based imaging has shown success in related surgical applications, highlighting its potential for future integration.

Bamba et al. developed a CNN-based CV system for real-time detection of the GI tract, blood, vessels, gauze, and clips^[11]. The system was trained on annotated surgical videos from various procedures, including colectomy, hernia repair, and sigmoid resection, with experts manually labeling each visible tool and anatomical object^[11]. A total of 1,070 images were utilized for training. The model, developed using IBM Virtual Insights, demonstrated a mean sensitivity and positive predictive value (PPV) of 92% and 80%, respectively^[11]. When identifying blood vessels - anatomy of particular relevance to VHR - the model achieved sensitivity and PPV of 79.3% and 82.1%, respectively^[11]. Notably, however, the group noted that the CNN model frequently misclassified reddish fat as blood vessels, suggesting the need for further refinement.

The study represents a significant step toward real-time AI-based imaging in intraoperative settings. A key limitation of this system is that it was trained on endoscopic images, which differ from the open surgical approach used in VHR. While the model performs well in confined endoscopic settings, its efficacy in open surgeries, where lighting conditions and perspectives are variable, remains unverified. To address these limitations, hybrid imaging approaches that integrate infrared or depth-sensing modalities may enhance anatomic differentiation. By building upon current advancements, AI-powered imaging can significantly enhance intraoperative decision making and improve patient outcomes.

Postoperative monitoring after VHR

AI-based imaging applications have the potential to enhance postoperative monitoring following VHR by detecting wound complications^[12-15]. Rochon et al. developed an AI-powered tool to identify surgical wounds requiring priority review^[15]. The AI utilized was “You Only Look Once’ (YOLO), an object detection algorithm known for real-time processing capabilities^[15]. While YOLO requires significant computational power, it offers faster processing than traditional CNN models. It was trained on 37,974 images and tested on 3,634 images^[15]. The model categorized images based on the presence of four factors: discoloration, unexpected fluids or tissues, retained sutures or clips, and incisional separation.

The model demonstrated an overall sensitivity of 89% and 100% intra-rater reliability in identifying wounds requiring urgent review^[15]. However, accuracy fell for detecting incision separation and skin discoloration in darker skin tones, likely due to dataset imbalance (71% lighter skin, 29% darker skin tones)^[15]. YOLO’s high computational demands may also limit adoption. Despite these limitations, AI-driven wound assessment presents a promising avenue for postoperative VHR monitoring.

Another study explored the use of mobile thermal imaging and deep learning to detect SSIs following cesarean section surgery in Rwanda, aiming to develop an accessible postoperative monitoring tool for resource-limited settings^[16]. Two CNN models, a naïve CNN (baseline) and ResNet50 (pre-trained for feature extraction), were trained on 500 thermal wound images, which were manually labeled by a physician^[16]. These images were captured using a mobile thermal imaging device, which records heat patterns emitted from the wound site. The naïve CNN achieved an accuracy of 0.86, while ResNet50 demonstrated an accuracy of 0.94^[16]. The study was limited by the small number of SSI-positive images, which may have affected the generalizability of the model. Expanding the dataset could improve model performance and advance the development of a cost-effective postoperative monitoring modality.