Baris Karakullukcu

NKI

In short, Baris Karakullukcu's research entails the following:

Utilizing artificial intelligence to segment echo images of tongue cancer intraoperatively, to facilitate radical resectionudy

Squamous cell carcinoma (SCC) of the tongue is a rare tumor of the oral cavity. The standard treatment involves surgery aimed at removing the tumor with a minimum margin of 5 mm of healthy tissue. However, due to insufficient intraoperative feedback on tumor borders, this margin is not achieved in up to 85% of patients. Tongue SCC can be effectively differentiated from normal tissues in ultrasound (US) images, allowing for predictions regarding the depth of invasion and resection margins. Nonetheless, the interpretation of these US images is highly subjective and dependent on the assessing clinician’s expertise.

To address this variability, we aim to standardize the US evaluation technique through artificial intelligence (AI) models and advanced image processing techniques. This approach can automate tumor segmentation in US images, yielding more accurate and consistent intraoperative information.

The primary objective is to train and validate AI models while developing a novel Simultaneous Localization and Mapping (SLAM)-based tumor reconstruction algorithm for automatic segmentation of SCC and normal tissues. We have created an ultrasound system capable of imaging the entire resection specimen and constructing a 3D volume from 2D stacked images. The specimens are sectioned in the pathology department along the same axis as the ultrasound images and stained. Histopathology slides are then evaluated, with tumors delineated and registered to the ultrasound images to generate ground truth data. Existing data has already been collected in this standardized manner, and new data will be gathered over the next three years. Seventy percent of the collected data will be used to train the AI models, while 30% will serve as a validation set. We hypothesize that the AI model can segment tumor and normal tissue with a Dice Similarity Score (DSS) of 90 when compared to histopathology specimens. 

In the future, the developed model could provide immediate feedback to surgeons in the operating room, enhancing the adequacy of tumor resection.