Variational Shape Completion for Virtual Planning of Jaw Reconstructive Surgery

Amir H. Abdi,Mehran Pesteie,Eitan Prisman,Purang Abolmaesumi,Sidney Fels

DOI: https://doi.org/10.1007/978-3-030-32254-0%5C_26

2019-07-16

Abstract:The premorbid geometry of the mandible is of significant relevance in jaw reconstructive surgeries and occasionally unknown to the surgical team. In this paper, an optimization framework is introduced to train deep models for completion (reconstruction) of the missing segments of the bone based on the remaining healthy structure. To leverage the contextual information of the surroundings of the dissected region, the voxel-weighted Dice loss is introduced. To address the non-deterministic nature of the shape completion problem, we leverage a weighted multi-target probabilistic solution which is an extension to the conditional variational autoencoder (CVAE). This approach considers multiple targets as acceptable reconstructions, each weighted according to their conformity with the original shape. We quantify the performance gain of the proposed method against similar algorithms, including CVAE, where we report statistically significant improvements in both deterministic and probabilistic paradigms. The probabilistic model is also evaluated on its ability to generate anatomically relevant variations for the missing bone. As a unique aspect of this work, the model is tested on real surgical cases where the clinical relevancy of its reconstructions and their compliance with surgeon's virtual plan are demonstrated as necessary steps towards clinical adoption.

Machine Learning,Graphics

What problem does this paper attempt to address?

The problem that this paper attempts to solve is: in mandibular reconstruction surgery, how to accurately reconstruct the missing part of the mandible. Specifically, when part of the patient's mandible has been missing or deformed before the surgery, doctors need to infer and reconstruct the missing part based on the remaining healthy structures. Since the shape of the mandible is crucial for functions such as chewing, speaking, and swallowing, its reconstruction requires not only functional accuracy but also aesthetic rationality.

### Main problems
1. **Unknown preoperative morphology**: In some cases, the surgical team cannot obtain the complete preoperative mandibular morphology information of the patient, which complicates the reconstruction work.
2. **Non - deterministic shape completion problem**: For the reconstruction of the missing part, there are multiple possible solutions rather than a single definite answer.

### Solutions
To solve the above problems, the author introduced an optimization framework based on deep learning to train the model to complete (reconstruct) the missing part of the mandible. The main methods include:
- **Voxel - weighted Dice loss function**: By using the weighted Dice loss function, the model can utilize the information of the surrounding structures and give priority to the accuracy of the reconstruction area.
- **Multi - objective probability solution**: Expand the conditional variational auto - encoder (CVAE), consider multiple possible reconstruction results, and weight them according to their consistency with the original morphology.

### Specific contributions
1. **Randomly generate training samples**: Designed a framework for generating training samples for deterministic and probabilistic shape completion models.
2. **Voxel - weighted Dice loss function**: Ensure that the model gives priority to the missing area during reconstruction and maintains consistency with the surrounding structures.
3. **Target - weighted variational objective function**: Solve the non - determinacy of the shape completion problem and allow the generation of multiple reasonable reconstruction results.

### Experimental verification
The author demonstrated the effectiveness of the model through quantitative and qualitative evaluations and tested it in real - life surgical cases, proving the clinical applicability of the model and the rationality of the reconstruction results.

### Formula summary
- **Voxel - weighted Dice loss function**:
\[
L_{\text{VWDice}}(X, Y, \hat{Y}, W)=1-\frac{2 \sum_{ijk}[\hat{Y} \circ(Y + X) \circ W]_{ijk}}{\sum_{ijk}[(\hat{Y}+Y + X) \circ W]_{ijk}}
\]
where \(W\) is a weight matrix, defined as:
\[
W_{ijk}=\begin{cases}
1 & \text{if } B_{ijk}=1 \\
\frac{1}{(2\pi\sigma_w)^{3/2} N_w(i,j,k)} & \text{if } B_{ijk}=0
\end{cases}
\]

- **Target - weighted loss function**:
\[
L_{\text{VWDice - TW}}=\alpha \sum_{i = 0}^{m} \Lambda_{0i}(Y) L_i^{\text{VWDice}}(X, Y_i, \hat{Y}_i, W_i)+\gamma KL(P_{\text{post}}||N(0, I))
\]
where \(\alpha\) is a normalization parameter, \(\gamma\) is a weight constant, and \(\Lambda_{ij}(Y)\) represents the geometric consistency measure between members \(i\) and \(j\) in the set \(Y\).

Through these methods, the author has successfully improved the accuracy and clinical applicability of mandibular reconstruction.