Research

Overview of Surgical Planning Laboratory

The Surgical Planning Laboratory (SPL) is a clinically oriented computer science laboratory at Houston Methodist Research Institute, with ongoing research projects in medical image analysis, mathematical modeling and simulation, surgical device design, as well as large-scale clinical software systems. At SPL, we utilize various types of medical images to perform segmentation, create 3D models, simulate surgical operation, and predict surgical outcome.

While the lab covers a wide range of topics, the focus of SPL has been craniomaxillofacial surgery—It pioneered computer-aided surgical simulation in this domain and has been innovating for nearly two decades. We embraced the recent rise of artificial intelligence, integrating many deep learning techniques into our clinical practice, and leading the efforts to create a federation of neural networks across multiple clinical sites for craniomaxillofacial image informatics.

AnatomicAligner

To provide a platform for orthognathic virtual surgical planning, we developed AnatomicAligner from the ground up. Today, surgeons can plan the correction of various types of facial deformity on the platform, through image segmentation, cephalometric analysis, virtual osteotomy, skeletal reconstruction, and custom splint and template design. In March 2020, AnatomicAligner acquired FDA 510(k) clearance for commercial use.

In the operating room, the virtual plan created in AnatomicAligner is a roadmap for surgeons: The cutting guides precisely locate the incision; the plates help set each bony segment in place. Using this platform, complex cases that would traditionally take multiple surgeries may be accomplished in just one, with increased accuracy and diminished risks.


A completed surgical plan in AnatomicAligner

This platform is never complete—It integrates our research output constantly and rapidly. During the past decade, the platform has seen many updates. For instance, we added functions for automatic dental articulation, which replaced the decades-old practice of using dental stone models; we also added automatic image segmentation, to replace an eight-hour laborious manual process. Although the software’s framework is set, each function within each module can be updated, so AnatomicAligner will remain state of the art.

This project was supported by the National Institute of Health, R01DE022676.

Image Segmentation

Orthognathic surgical planning requires segmentation of different facial components, e.g., maxilla, mandible, soft tissue, to name a few. This is accomplished by creating masks for each component, selecting the region of interest while filtering out the background. Thresholding based on grayscale values is a common way to initialize such a mask, but it is far from perfect and manual refinement is a daunting task. Additionally, anatomical landmarks form the basis of cephalometric and diagnostic analysis for surgeons, and labeling these landmarks is also a routine task that requires considerable manual effort.


A comparison between initial and manually edited segmentation results

The low image quality of Cone-Beam Computed Tomography (CBCT) is a major factor that complicates the segmentation process. Compared to traditional Computed Tomography (CT), CBCT exposes patients to less harmful radiation, and the imaging device provides much ease of use. But the added safety and convenience for patients come at the cost of image quality: To segment a CBCT scan for surgical planning purpose, a trained specialist must go through a laborious manual editing process, amounting to eight hours per patient.

To speed up this process, we found great success using deep learning technology. Specifically, based on the concept of Convolutional Neural Network (CNN), we built SkullEngine, a framework that handles high-resolution 3-D images in a coarse-to-fine fashion, producing segmentation masks and detecting anatomical landmarks simultaneously. To date, SkullEngine has been integrated into AnatomicAligner platform, and it has reduced manual editing time to under one hour per patient.

This project was supported by the National Institute of Health, R01DE022676, and is pending renewal.

Outcome Prediction

Craniomaxillofacial surgery restores normal facial appearances and functions; the planning of such surgery must consider both aspects. While the bone is seen as rigid and its movement described simply by rotation and translation, soft tissue has complex biomechanical properties and behaves in a highly non-rigid fashion. This makes soft tissue changes difficult to predict following skeletal reconstruction. As a result, surgeons plan and operate on the bone, yet restoring normal facial appearance becomes an afterthought.

Our research proved that such prediction is feasible using the Finite Element Method (FEM). We modeled the soft tissue volume as a hexahedral mesh and simulated skeletal reconstruction by initializing nodal movements on the mesh inner surface (in contact with the bone). The mesh would then respond with changes on the outer mesh surface (skin), bringing us the predicted post-operative face. Through multiple improvements, our method maintained state of the art accuracy, even in challenging regions such as the lips.


The framework for finite-element based outcome prediction

An accurate prediction of soft tissue changes is a valuable tool: It provides useful feedback for surgeons during surgical planning, and it helps reassure patients during consultation. Now that we have achieved good prediction accuracy, we look to improve the speed by finding more efficient ways of mesh modeling, as well as physics-informed deep learning of soft tissue behavior.


The framework for learning based outcome prediction

This project is supported by the National Institute of Health, R01DE021863.

Helical Distractor

When the amount of correction required to remove mandibular or maxillary retrognathia goes beyond the limits of orthognathic surgery, distraction osteogenesis can be used to gradually enlarge the deformed mandible or maxilla. This situation occurs when there is extreme retrusion, or when surrounding soft tissues are unyielding due to scarring. It is most common in previously repaired clefts of the lip and palate.



Unfortunately, the movements needed for successful distraction are complex, involving translations in all cardinal directions (anteroposterior, transverse, and vertical) and rotations about all cardinal axes (pitch, roll, and yaw). Despite this fact, current internal distractors can only execute linear or circular motion. A linear distractor can shift a bone segment into a new location but cannot rotate it; A circular one can blend translation and rotation but only in a single plane.

In theory, it is obvious that a helical distractor would be better and more capable, since helical motion can place a bone segment into any position and orientation in a three-dimensional space. With that thought, we simulated treatment using helical distraction, for mandible and maxilla respectively, to prove that such treatment indeed leads to better outcome. A patent application is pending for the invention of helical distractors, and we believe this brings about major change in the treatment of hypoplastic skeletal deformities.