HomeEducationDoctorate (PhD & EngD)For current candidatesPhD infoUpcoming public defencesPhD Defence Tim Boers | Enter the matrix: on how to improve thyroid nodule management using 3D ultrasound

PhD Defence Tim Boers | Enter the matrix: on how to improve thyroid nodule management using 3D ultrasound

Enter the matrix: on how to improve thyroid nodule management using 3D ultrasound

The PhD defence of Tim Boers will take place in the Waaier building of the University of Twente and can be followed by a live stream.
Live Stream

Tim Boers is a PhD student in the department Multi-Modality Medical Imaging. (Co)Promotors are prof.dr. M. Versluis and prof.dr. S. Manohar from the faculty of Science & Technology and dr. S. Braak from ZGT. 

Roughly two-thirds of the adult population has a thyroid nodule, of which 90% are benign. Of the adults that have a nodule, approximately 5% will experience symptoms that include a feeling of a marble stuck in the throat, difficulty swallowing and breathing, and cosmetic complaints. Via ultrasound imaging, an initial malignancy risk classification is determined, after which a fine needle aspiration can be performed to improve the malignancy risk assessment. Based on the patient's characteristics and the malignancy risk assessment, a treatment plan is suggested. Surgical removal of the afflicted lobe is the current treatment standard for non-hyperfunctioning benign thyroid nodules. A minimally invasive alternative is available; radiofrequency ablation (RFA) is a proven suitable alternative for this type of nodule, facilitating volume reduction by burning a large part of the nodular tissue via an RF electrode. Thyroid nodule management primarily makes use of ultrasound as the imaging modality for diagnosis, image guidance during therapy, and follow-up. Although ultrasound is relatively easy to apply, it is hard to standardize for repeated measurements and across various users. This inter- and intra-observer variation in volume measurements has a range that is too large. This makes it more difficult to perform an accurate diagnosis; moreover, the requirement to do a fine needle aspiration is based on the size of the nodule, amongst other ultrasound features. During follow-up, the nodule volume reduction ratio (VRR) is used as a means to assess treatment efficacy; however, with the aforementioned observer variation, the VRR becomes less reliable. These challenges may be overcome by using 3D ultrasound. Creating a volume instead of slices allows the radiologist to scroll through the volume instead of only having the slices of the required scanning planes; this reduces observer variation, especially when measurements are required. Additionally, the increase in available imaging data can be used during interventions to make better-informed decisions. The development of and training with new technologies and methods should be done in a simulation environment as much as possible to prevent an increased complication risk for patients. In this thesis, a platform is created on which these methods can be developed. Further,  it offers insight into the use of 2D and 3D ultrasound for thyroid nodule management.

The current and future use of ultrasound in thyroid nodule management has been reviewed in Chapter 2. Although ultrasound is the standard for thyroid nodule management, it can be improved. The current inter- and intra-observer variation is too large and should be reduced. This can be done by applying 3D ultrasound and automatic segmentation methods. Additionally, adding additional analysis methods, such as elastography and Doppler or contrast enhanced ultrasound, may improve the accuracy of the malignancy risk scoring system TI-RADS as well as aid follow-up in identifying ablated from non-ablated areas. Computer-aided needle-based interventions may improve needle placement and, thereby, intervention effectiveness.  

To assess the impact of changes to an intervention, a baseline has to be created. The effectiveness of RFA in Dutch hospitals, the effect the learning curve has on treatment effectiveness, and what factors contribute to predicting technical treatment success 1 year after RFA are investigated in Chapter 3. In three centers, 346 patients with 366 nodules were retrospectively included. A learning curve was encountered in the VRR and the amount of energy applied per volume for up to 30 cases per center. These learning curve cases were excluded from further analysis to understand the current effectiveness of RFA in these centers. A median of 70.4% VRR at 12 months after treatment was found. The literature describes that baseline nodule volume and the amount of energy applied per volume are parameters that influence the VRR. Based on our analysis, significant differences were encountered; however, these were not clinically relevant. What can be clinically relevant is that when a nodule has not reached a VRR of 50% at 6 months, it is unlikely to reach treatment success at 12 months. A re-treatment can be considered when the patient still experiences nodule-related complaints.

With the baseline known, the next step was assessing the effect of changing a transducer from 2D to 3D for RFA while keeping all other factors constant. Measuring this in patients would put them, unnecessarily, at higher risk for complications and introduces patient variability to the study, which then requires a large group to be included to counter this. However, when using phantoms, we are able to keep the majority of the factors constant. Phantoms mimic a part of the human body, either in a simplistic way or in a complex fashion, nearing anthropomorphism. What these phantoms have in common is the capacity to offer a reference frame in which singular changes to an intervention can be assessed.

The literature has described the impact of inter- and intra-observer variation on thyroid and thyroid nodule measurements. Differences of up to 30% can occur, which is reason to investigate the use of more advanced ultrasound transducers, such as 3D ultrasound. The impact of this variation is evident when we look at the TI-RADS scoring systems for malignancy risk assessment. These systems use the longest-axis measurement in order to determine, per risk category, whether to initiate fine needle aspiration or to enlist the patient in a watchful waiting regime. If the measurement is prone to change per operator or per measurement moment, then at one moment the patient needs to undergo a fine needle aspiration, while with another operator the patient would be enlisted into a watchful waiting regime. Reducing the measurement variation can make these TI-RADS scoring systems more robust. Chapter 4 has shown that utilizing a volume-based measurement technique that the matrix transducer offers results in improved measurement accuracy. Despite the phantom nodule being spherical and thus easier to measure with 2D ultrasound and calipers, we were still able to achieve higher measurement accuracy with the matrix transducer.

To improve phantom-based study result extrapolation to the clinic, these phantoms should mimic the human situation as closely as possible. The anthropomorphic phantom shown in Chapter 5 is well characterized and is, in most aspects, similar to the human neck. The results of this study on this phantom indicate that extrapolation to the clinic should be possible. The phantom serves as a platform on which thermal treatments, such as RFA, can be improved. Hopefully speeding up development processes as well as having a more tested and developed system before it hits clinical studies and comes into contact with patients, thus reducing the burden on patients.

The following step was to put the anthropomorphic phantom to use. A comparison was made between RFA guided by 2D ultrasound and by 3D ultrasound. Chapter 6 shows that the impact of 2D and 3D ultrasound on RFA efficacy does not differ from one another; however, the matrix transducer is perceived by the operator as being more user-friendly for needle placement due to the dual-plane imaging. Additionally, we saw a trend for a lower nodule ablation percentage and a lower volume ablated outside the nodule when using 3D ultrasound. Although a lower volume ablated outside the nodule is preferable, the nodule ablation percentage should be kept the same. A proper balance between safety and effectiveness should be found. It remains a question whether these ablation results are due to the operator still gaining experience with the matrix transducer or if the additional scanning plane made the operator ablate more restricted. What is clear is that determining, in the ablation zone, what part of the nodule has been ablated and what has not remains difficult due to gas bubble formation. To solve this, planning and navigational software should be developed. This should lead to higher ablation percentages while remaining as safe or becoming even safer than we currently are by keeping track of what has been ablated and what has not, and where critical structures are situated that need to be avoided. An additional use case for these phantoms is their capacity to compare dominant and non-dominant hand ablations. Our operator showed differences in the ablation percentage, amount of energy introduced, and time taken between hands using 2D ultrasound, while these differences were not significant when using 3D ultrasound. Additional research is required that employs more operators to find stronger evidence supporting a difference between the ablating hands.

In Chapter 7, a deep learning algorithm is employed to automatically segment 3D ultrasound scans. The hypothesis was that 3D data would improve segmentation accuracy. A comparison was made between a 2D ultrasound-tracked sweep data set and a 3D matrix ultrasound data set. The thyroid, carotid artery, and internal jugular vein were segmented using a 2D and a 3D U-net with three methods: 2D, 2.5D, and 3D. The results show that automatic segmentation is possible for 3D ultrasound data. However, the current 3D algorithm does not outperform the 2D approach. The limited field of view and lower resolution can play a role in the accuracy of the segmentations, although newer iterations of the U-net should also be used. An advantage of the 3D data set is that motion artifacts are less prevalent than in the 2D tracked sweep data set.

In conclusion, as also discussed in the general discussion in Chapter 8, the work in this thesis showed that RFA for benign thyroid nodules in the Netherlands is safe and effective and has a relatively short learning curve. Challenges remain for thyroid nodule management, for which 3D ultrasound can be a solution. To make full use of 3D ultrasound, stitching algorithms should be integrated into the ultrasound systems to acquire larger volumes. These volumes can then be used in computer-aided diagnosis and intervention systems. To improve the applicability of 3D ultrasound in the clinic, integrating analysis methods such as 3D elastography and 3D Doppler is suggested. To expedite such developments and testing, an anthropomorphic phantom platform was created for thermal-based therapies. This phantom should also be used as a training platform to gain insight into ablation performance as an operator.