Development of an Artificial Intelligence Model for Intraoperative Parathyroid Gland Identification During Thyroidectomy

Clinical

For patients satisfying the inclusion and exclusion criteria and providing consent to participate in the study, baseline characteristics would be obtained and stored in a de-identified manner for linking with intraoperative image data. The following characteristics would be captured:

i.               Age

ii.              Sex

iii.            Treating Centre

iv.            Thyroid pathology: FNAC/ final histopathology

v.              Size and number of thyroid nodules

vi.            Ultrasound TIRADS score

vii.           Computed tomography or magnetic resonance imaging (if indicated and available)

viii.         Surgery undertaken and approach

Once the lateral thyroid region has been exposed depending upon the approach chosen for the thyroid gland, the following structures will be identified and exposed on the side being operated and photographed.

a.     Internal jugular vein and common carotid artery laterally

b.     Recurrent laryngeal nerve

c.     Tubercle of Zuckerkandl

d.     Superior and inferior parathyroid glands

A high-definition camera will be placed 15 cm from the operative field such that each frame captures all the structures listed above. Multiple images will be obtained with the camera placed lateral and anterior to the lateral thyroid region (Figure). In case of endoscopic thyroidectomies, the robotic camera will be used, zoomed out to incorporate the superior and inferior parathyroid glands, recurrent laryngeal nerve and the tubercle of Zuckerkandl.

To obtain images for training the model on the parathyroid gland vascularity, images would be repeated after the specimen is excised.

Intraoperative confirmation of PG identification will be undertaken by a combination of the following techniques:

1.     After obtaining images with white light, the same field will be observed with NIRAF / ICG angiography (subject to the availability and practice at the treating center).

NIRAF images will be obtained after turning off the lights in the operating room and holding the NIR filter 15 cm from the operating field. The autofluorescence from the parathyroid glands will be visualized on the monitor of the NIRAF system.

ICG angiography is performed after injecting ICG into the peripheral vein at a dose of 5mg and waiting for 1 to 2 minutes. The field will be visualized with infrared filter and images will be obtained.

2.     Frozen section analysis of devascularized PG

3.     In-vivo confirmation of parathyroid gland (Wei et al.) (4): Suspected parathyroid glands will be  subjected to insitu needle aspiration with 24G needle. Multiple passes will be obtained to increase yield. Syringes will be rapidly emptied and thin cytological smears will be prepared. Fixation will be done using a rapid Diff-quik staining or toluidine blue for intraoperative onsite evaluation followed by permanent fixation. Parathyroid gland confirmation will be performed by the pathologist.

 Patients undergoing total thyroidectomy will be followed up by serum iPTH assay on postoperative day 1 and 1-year follow-up to determine the incidence of temporary and permanent hypoparathyroidism, respectively.

Image Annotation

Selection of representative frames and subsequent image annotation will be performed by a designated team of high-volume thyroid surgeons. High-volume thyroid surgeon will be defined as a dedicated head and neck surgeon with greater than or equal to 10 years of experience and performing or supervising a minimum of 30 thyroidectomies or parathyroidectomies per year. The anatomical structures annotated will be:

1.     Common carotid artery

2.     Internal jugular vein

3.     Recurrent laryngeal nerve

4.     Tubercle of Zuckerkandl

5.     Superior parathyroid gland: normal, ischemic

6.     Inferior parathyroid gland: normal, ischemic

 Workflow for Image Analysis

The annotations will be processed to classify each pixel falling within an annotation as a specific class, denoting the distinct structures mentioned above. The annotated images and the raw images will be further processed to transform pixel sizes into uniform sizes of 1 mm × 1mm, and pixel intensities across all images will be normalized with the z-score normalization method. The raw images will be resized into 256×256 pixels to enhance the computational efficiency.   Data augmentation techniques, 1) Rotational shift, 2) Brightness, 3) Contrast, and 4) Random zoom, will be performed on the whole dataset to train the deep learning architectures efficiently and render them robust.

State-of-the-art medical image segmentation algorithms, 1) CNN-based U-NET, Res-UNet, Attention-UNet, 2) Vision transformer-based Swin-UNet, and 3) Hybrid of CNN and Transformer based  Swin-UNETR and Trans-UNet will be evaluated on their performance to segment desired structures by classifying each pixel into specific categories post-training on the dataset [17]. During training, the weights of the architectures will be optimized by the categorical cross-entropy loss function. The entire dataset will be divided as 70% for training, 15% for validation during training, and 15% for testing. The performance of the models will be evaluated with the following metrics: 1) Dice score, 2) Accuracy, 3) Sensitivity, 4) Specificity, and 5) Intersection Over Union [18].

The best-performing model will be further tested on data from distinct centers to assess its generalizability. Following this, the model will be packaged into an API (Application Package Interface) format to enable its real-time execution

The computation required to train the algorithms will be conducted on Nvidia RTX A4000 GPU with 24 GB VRAM.

Outcomes:

The AI model will be compared in terms of parathyroid gland identification and vascular integrity of the parathyroid gland with blinded high-volume surgeon. The following outcomes will be calculated:

Accuracy

Sensitivity

Specificity

Dice Score

d. Sample size: The AI model needs to be trained on at least 1000 images and 300 patients.

e. Statistical analysis: The entire dataset will be divided as 70% for training, 15% for validation during training, and 15% for testing. The performance of the models will be evaluated with the following metrics: 1) Dice score, 2) Accuracy, 3) Sensitivity, 4) Specificity, and 5) Intersection Over Union. The computation required to train the algorithms will be conducted on Nvidia RTX A4000 GPU with 24 GB VRAM.